Minerval Architecture

This document describes the Minerval system as it is built today: the domain model, the agent pipeline that populates it, the data layer underneath, and the surfaces that serve it. It is a description of the running architecture, not a roadmap. Where a design decision has interesting consequences, the reasoning is given inline.

The companion documents are the constitution, the text every administrator agent is bound by; the agents pages, which show each agent's actual system prompt verbatim; and the operational policies further down this page, which translate the constitution into concrete rules for each agent.

System Overview

Minerval turns documents into a queryable graph of claims. Ingestion is the expensive, write-side work: an LLM pipeline reads a source, pulls out atomic claims, decides whether each is new, decomposes it into its supporting structure, and assesses its validity. Serving is the cheap, read-side work: the graph is queried directly, with no LLM in the read path. The few surfaces that invoke agents on demand (submitting a source, the browser extension's page analysis, the MCP assessment tools) are authenticated, rate-limited, and metered per account.

   SOURCE               INGESTION                    GRAPH
 ┌─────────┐   ┌───────────────────────────┐   ┌───────────┐
 │ URL or  │──▶│ Extractor → Matcher →     │──▶│ Postgres  │
 │ document│   │ onboard → Claim Steward   │   │ + pgvector│
 └─────────┘   └───────────────────────────┘   └─────┬─────┘
                                                     │ read
   GOVERNANCE (ongoing)                              ▼
 ┌──────────────────────────────────────┐      ┌───────────┐     web ·
 │ Claim Steward (decompose + assess) · │◀────▶│    API    │──▶  extension ·
 │ Curator · Contribution Reviewer ·    │      │ (Fastify) │     MCP clients
 │ Dispute Arbitrator · Audit Agent ·   │      └───────────┘
 │ Grantmaker · Lookout ·               │
 │ Consistency Checker                  │
 └──────────────────────────────────────┘

The work is done once, at ingestion, and reused everywhere a claim recurs: the same claim appears across thousands of documents but is decomposed and assessed a single time. Nor does every claim get the full treatment immediately. Each claim carries an importance score that anchors the expected-value estimates the allocation engine funds work by (docs/allocation.md), so the most consequential claims draw assessment first while minor ones remain searchable stubs until an allocation covers them.

The stack is TypeScript end to end: a Fastify API, background workers driven by a job queue (AWS SQS in production, an in-memory runner locally), and PostgreSQL with the pgvector extension as the single store, carrying vector search and full-text search alongside the relational data. Anthropic Claude models sit behind every agent by default; model ids are centralized in src/llm/models.ts, and in production the load-bearing agents run on Claude Opus 5.5. Any agent can be pointed at OpenAI or OpenRouter instead with a single env var — the Matcher runs on GPT-6 Luna this way — see Providers.

The Domain Model

Six entities carry the epistemic content. The rest of the schema records accounts and governance (contributors, contributions, reviews, appeals, arbitration, reputation) and operations (sources, jobs, usage metering). The epistemic core, where ──< reads "has many":

  Source ──< Instance >── Claim ──< Relationship >── Claim
                            │           (decomposition edge)
                            │                │
                            │                └──< ArgumentSubclaim >──┐
                            │                     (membership: which   │
                            │                      arguments group     │
                            │                      the edge)           │
                            ├──< Assessment   (verdict history;        │
                            │                  one is_current)         │
                            │                                          │
                            ├──< Argument      (a named line of ───────┘
                            │                   reasoning)
                            │
                            └──< Link >── Claim  (lateral see-also; symmetric,
                                                  never a dependency)

Claims

A claim is the atomic unit: a proposition that can be true or false. Empirical claims (directly verifiable or derived), definitional, evaluative, causal, normative, and mathematical claims are all represented the same way and all decompose into subclaims. Two formulations are the same claim when they turn on the same considerations: nothing could count as evidence or argument bearing on one without bearing equally on the other (identical decomposition is a useful diagnostic, not the definition). This is the basis for deduplication, and it extends to negation. A claim and its denial are one node, and each recorded appearance carries a stance saying which side it takes.

Each claim carries its canonical text, a claim_type, a lifecycle state (a claim merged into another records the target in merged_into rather than being deleted), counters for how many of its children have been assessed, an embedding (a 1536-dimension vector) and a text_search column for retrieval, and an importance score.

Importance is a 0 to 1 judgment of consequence-if-wrong times liveness, how actively the claim is disputed or consulted. It is explicitly not a count of dependents: a settled fact cited everywhere scores low, a load-bearing contested premise scores high. The Extractor supplies a prior at ingestion; the claim's Steward sets the authoritative value once it has decomposed the claim and seen its neighborhood. Importance decides both how soon a claim is stewarded and how much effort its Steward spends.

Two further signals complete the picture (issue #172): contestation on the claim, how live the dispute is stated unfused from the consequence half, and marginal_yield on each assessment, the Steward's exit judgment of how much another, stronger pass would improve it. Both are live inputs to the expected-value estimate the allocation engine funds work by (docs/allocation.md): value = importance × contested-factor × expected quality gain, with cost on the other side of the ratio.

Two of the claim's fields carry two different judgments about its kind. claim_type is the proposition-kind facet that the page eyebrow, the cards, the map's bedrock logic, and the territory listing read; its values include mathematical, a proposition of mathematics, true or false by proof rather than by observation, settled by a proof others can check and most firmly by one a machine has checked. domains is a separate tag list (claims.domains, with domains_source recording whether the tag came from the Extractor, the Matcher, inheritance from a parent subclaim, a Steward's set_claim_domains, or the backfill) that selects which domain skills and tools a run on the claim carries (docs/policies.md, "Skills"). The two are independent on purpose: a claim about the economics of a theorem is causal in type and carries mathematics among its domains; a theorem is mathematical in type and mathematics in domain. The Steward sets both, and neither the funding mandate nor importance gates the tools a domain brings.

Arguments

An argument groups decomposition edges into a coherent, named line of reasoning. A single claim routinely has several:

Philosophy: "God exists" has the cosmological argument, the teleological argument, the argument from evil, and others.
Policy: "We should raise the minimum wage" has a poverty-reduction argument (for) and an unemployment argument (against).
Science: "The universe is ~13.8 billion years old" is supported independently by the CMB, stellar evolution, and nucleosynthesis.

Forcing these into one flat set of edges would lose the structure of which subclaim belongs to which line of reasoning. An argument has a stance (for or against: its conclusion is the claim or the claim's denial; older rows may carry neutral, which Stewards no longer record), an optional name and description (a short label), its content (the written form, below), any evidence_urls, and provenance.

The written form. A name is not an argument: the grouping records which subclaims belong together, but not how they combine to bear on the parent claim. Every named argument therefore carries a written form in content: one to three sentences of logically straightforward prose stating the inference, with every subclaim referenced inline as [[claim:<uuid>]] (or [[claim:<uuid>|inline phrasing]] when grammar demands it). For example: "Because [[claim:a]] and [[claim:b]], and given [[claim:c]], the claim follows." The links make the prose and the grouping mutually checkable: every subclaim edge in the argument should appear in the written form, and every reference must be an edge in the argument (the Steward's write_argument tool enforces the latter and warns on the former). Renderers resolve the ids to the claims' canonical text at display time, following merged_into, so links never dangle after a merge. The connective language that the claim bar expels from claim texts ("therefore", "because", "given that") lives here and only here.

The evaluation. The written form deliberately withholds judgment; the judgment lives in a separate steward-maintained evaluation (argument_evaluations, one is_current row per argument with prior rows kept as history). It rates the inference itself (#540): how far the conclusion follows if the premises hold, as certain (it follows necessarily), likely (much more likely, not guaranteed), slight (the premises bear on it only a little) or none (no real support even if they are true). The rating is worded conditionally on every surface ("certain if the premises hold") because it is not a verdict on the conclusion: whether the premises hold is read from their own assessments and shown beside it. An evaluation may also mark the argument decisive: the Steward's judgment that, taken alone and with its premises as they are now assessed, it settles its conclusion. The mark lapses with its premises, and the claim's status has to agree with it. Beside the rating sit two to four sentences of reader-facing prose saying how far the inference goes and which premises, given their current assessments, the argument lives or dies on, with those load-bearing premises linked inline exactly as in the written form. Evaluations written under the earlier scale (holds, holds_with_caveats, fails, contested) were retired as history rather than relabelled, since none of those values maps cleanly onto the new ones. The evaluation is derived within the claim's assessment process, not as a fire-once verdict: the Steward's evaluate_argument tool stamps each evaluation with the assessment it was derived under, so an evaluation left behind by a later reassessment is detectably stale, and the assessment tool's result nudges the Steward to bring evaluations current whenever it records a new verdict.

Two design decisions follow:

The written form is structural; the evaluation is epistemic. The written form states the inference, never a verdict on whether it holds; the verdict and the load-bearing analysis live in the evaluation, which is maintained as part of the claim's assessment and follows the same transparency rules (reasoning visible, open to challenge). It is reader-facing prose, not a discussion surface: contributor exchanges stay in the contribution record.
Arguments are optional and non-exhaustive. A claim with one natural decomposition needs no explicitly named argument; its edges simply have no membership row. Admins create arguments when a distinct line of reasoning bears on the claim, whether or not a source has yet made it.
Arguments may share subclaims. Membership is a relation between an argument and an edge (argument_subclaims), not a column on the edge, so the same premise can be grouped under several of a claim's arguments, arranged differently in each written form.

When the validity of an argument's framework is itself disputed in the discourse, "this framework is valid" is added as a subclaim within that argument, typically with an assumes relation. The rating then judges the inference within the framework, and the framework's own assessment bears on whether the argument's premises hold, like any other premise. The meta-dispute itself stays inside the claim layer, where decomposition, assessment, and contribution already operate.

Decomposition edges

Decomposition is recorded as claim relationships: directed edges from a parent claim to a child claim. Each edge has a relation_type, a free-text reasoning, and a confidence. Which named arguments an edge belongs to is recorded separately, in argument_subclaims (one row per argument and edge), so a child can appear under multiple arguments (shared subclaims) while the dependency itself is stated once; an edge with no membership row is part of the claim's ungrouped basis. A uniqueness constraint prevents duplicate parent/child/relation triples, and self-edges are rejected outright. Edges are the only relations propagation and assessment read; a relation between claims that is not a dependency (a rival explanation, two halves of one position, two formulations kept apart because identity was unclear) is a claim link in its own table, symmetric and non-evaluative, rendered as a see-also on both claim pages and invisible to the tree. The relation types are:

Relation	Meaning
`requires`	The child is a load-bearing premise: the parent is false without it.
`supports`	The child provides evidence for the parent without being logically required.
`contradicts`	The child weighs against the parent.
`specifies`	The child narrows or makes precise part of the parent.
`defines`	The child fixes the meaning of a term in the parent.
`assumes`	Background the parent's framing takes as given (often a framework or scope claim): if it fails the parent is ill-posed rather than simply false. Renamed from `presupposes` (#205).

Assessments

An assessment is a verdict on a claim at a point in time, and it is written for two audiences at once. The summary is the reader-facing verdict: a short paragraph, shown at the top of a claim page, saying what the evidence establishes and where the weight rests. The reasoning_trace is the audit record: how the evidence and decomposition were weighed, kept so the judgment can be reviewed, not so it can be read. Splitting them means neither has to compromise; a trace written to be skimmable makes a worse audit record, and an audit record shown to readers makes a worse summary.

Alongside these sit the status and confidence, a subclaim_summary snapshot of the children at judgment time, and the trigger (and trigger_context) that prompted the assessment. Assessments are append-only history: exactly one row per claim is flagged is_current, enforced by a partial unique index, so the timeline of why a claim's status changed is fully recoverable. The statuses and how they propagate are described under Assessment below.

Node kinds: quantities and questions

A node in the graph is one of three kinds (constitution §2, §7a), recorded in claims.kind: a claim about a proposition, a quantity whose subject is a value, or a question. Every node written before kinds existed is a claim. A fresh claim can become a quantity or a question until it holds an assessment or an argument; after that, a quantity or question on the same subject is a new node. assessments stays a record of verdicts on propositions, so a quantity is never "verified" and a question never "supported".

A quantity's quantity_definitions row says what it measures, in which population, under which conditions, and in what unit. The graph's estimate of it lives in quantity_estimates, append-only with one is_current row like assessments, as a distribution at the precision the evidence supports (a distribution by quantiles, an interval, a bound, an order of magnitude, a single figure, or no figure with a reason) and the share of probability that lies outside every value. A source's statement of the value is an instance with stance states and its figure in stated_estimate, linked to the argument that rests on it. quantity_readings records readings of instruments taken to indicate the quantity, with what was actually measured and how far it bears. An argument about a quantity records what it shows about the value (a figure, a range, a side of a figure, or words alone) in place of a stance. A claim about which side of a figure a quantity lies is recorded in quantity_cuts, so its credence can be checked against the estimate.

A question's candidate answers (claims or quantities) are answers edges from the question, carrying an answer_basis (quality, prominence, or both); its subquestions are part_of edges; its presuppositions are the ordinary assumes and requires edges. answer_relations records, pair by pair, whether two answers exclude each other or can both hold. How well answered a question is lives in question_assessments as a standing (answered, leaning, open, ill-posed, unknown) with the answers it favors. A source that asks the question is an instance with stance asks; one that answers it has stance answers and names the answer node its answer matches. The vocabularies and the JSON shapes are in src/schemas/kinds.ts, and the agents' tools do not offer them until the tools that write them exist.

Requests

A request (constitution Part VIII) is something someone could do that would improve what the graph knows: what is asked, why (what the judgments it bears on would become under its outcomes), and what would satisfy it. An administrator, a contributor, or a mandate files it; a mandate's request is shown as filed by "a funded mandate", never by name. Its keeper is fixed when it is filed: the Steward of the one node it bears on, or the Curator when it bears on several nodes, a source, or something the graph does not hold yet. request_targets holds what it bears on and request_work the work done on it, each with an outcome (met, partly met, negative, cannot be met, declined, revise); a negative result or a refusal is a complete outcome. The keeper decides when the request is met, and a closed request (met, declined, withdrawn, moot) stays closed.

A mandate's Grantmaker can fund the Researcher to take up an open request (fund_request): the request is the brief, the run writes nothing to the graph, and its outcome is recorded as work on the request, after which the keeper is asked to judge it. Contributors propose requests, answers and estimates, and offer material for a request, through the Contribution Reviewer like any other contribution, and propose a question directly (POST /claims/propose with kind: "question"). Research a Grantmaker launches with contribution_budget_usd submits what it establishes as contributions under the mandate: reviewed like anyone's, with the review metered to the mandate's budget job, shown as "a funded mandate", and earning no one reputation or owls. The budget is a dollar amount of the mandate's budget that the Grantmaker delegates for those reviews (research_runs.contribution_budget_micro_usd). Each submission reserves an estimate of its review's cost (the average of recent reviews, or CONTRIBUTION_REVIEW_ESTIMATE_USD before there are any) and is refused when the run's committed review costs plus that estimate would exceed the budget; once reviewed, the real cost (contributions.review_cost_micro_usd) replaces the estimate. The budget is visible to both agents: the Researcher sees what is committed, what is left and what a review is expected to cost in its task message, after every turn and in each submission's result, and the Grantmaker sees the same in the delegate_research and get_research_run results. A review that costs more than estimated can carry a run somewhat past its budget.

Formal statements and machine-checked arguments

A mathematical claim may carry a formal statement: the graph's own rendering of the proposition as a Lean 4 def Statement : Prop, elaborated against a pinned Mathlib revision and toolchain. It lives in its own table (claim_formalizations) rather than in columns on claims, because it has a lifecycle with more than one live row (a draft beside the published one), prizes and attempts pin it by id and hash, and its provenance needs columns of its own. Each row records the pin (pin_id, lean_toolchain, mathlib_rev, image_digest), the statement source verbatim, two hashes (source_hash, which anyone can recompute from the published text and the pin, and expr_hash, over the elaborated body, which the checker compares), the pretty-printed proposition, the Mathlib constants it references, whether a witness example for satisfiable hypotheses elaborated, a reader-facing correspondence note in the graph's voice saying how the formal and informal statements relate and what the formal one leaves out, and a status of draft, reviewed, published, or retired. A partial unique index allows one published statement per claim. Publication is two-pass: the formalize action runs the Steward once to draft, elaborate, and record reviewed, then a fresh-context Steward pass with trigger formalization_review either publishes or returns it to draft. This is the defense against a statement whose every line is correct and whose theorem is the wrong one. A published statement carries review_period_ends_at, and no bounty may bind to it before then; a canonical-form change on the claim returns the statement to reviewed in the same transaction and moves any open bounty to rebinding.

A machine-checked argument is an ordinary argument with stance for (or against, for a disproof) whose evidence is a lean_checks row the checker accepted. lean_checks is the server-side record of every check the platform runs, in three modes (prize, attempt, steward), with the verdict, the per-gate record, diagnostics, resource use, the pin, the image digest, and the metered cost; repeated checks of the same statement, submission, checker version, and mode return the stored row. The claim page derives the machine-checked badge at read time from the argument, its evaluation, and the check. It is not a seventh status: the six statuses are constitutional and closed in code. The badge says the proof checks against the statement; the verdict beside it is still the Steward's judgment of the claim as worded.

Instances and sources

A source is a retrieved document (URL, title, content hash, raw content, type). An instance links a canonical claim to one place it actually appeared: the exact verbatim_text quote (the passage as that source stated it; every source's excerpt is equally verbatim to itself, and none has precedence), the surrounding context, a brief summary_context describing the circumstances ("said during a Senate hearing on banking regulation, in response to questioning about derivatives oversight"), a stance recording whether the quote affirms or denies the canonical claim, and a confidence that the quote really expresses it. Instances are how a single canonical claim accumulates provenance from many documents, and the stance field is what lets a claim and its negation share one node without losing track of who said which.

Source provenance: what the support rests on

A count of a claim's instances treats every appearance as one independent vote, and that is the one thing it usually is not: a news report rests on the study it cites, a second report rests on the first, a commentary quotes the second, and three of the four are one voice. Source provenance (#286) records that structure so a Steward and a reader can see it. It is a substrate for judgment, never a mechanism that produces one: no independence score, no effective sample size, no concentration index, no automatic discount, and no status ever moves because of the map's shape. Concentration on one root is frequently strength, not weakness, and only the Steward can say which in a given case (Part VIII, §9).

Four tables carry it. claim_instance_readings is what the Steward concluded from opening one instance's source against the claim: support, whether the source's own evidence bears the assertion it makes (supports, overstates, understates, asserts_without_evidence, contradicts_own_evidence, unclear), which is explicitly not whether the claim is true; deployment, what the source is using the claim for; a reader-facing note; worth_reading with its reason; source_read, whether the text was actually opened; and quote_check, the one mechanical field, computed by normalized substring matching of the instance's verbatim_text against the source's stored text (verbatim, normalized_match, not_found, no_stored_content) and never written by a model. One reading per instance, replaced in place. claim_provenance_edges is the load-bearing edge: from one source's assertion of the claim (an instance) to the document that assertion draws on (a source, which need not assert the claim itself; where it does, to_instance_id links the chain). The edge is scoped to the claim through its instance, because a bibliography is a document-level fact and the question is claim-level; it carries a relation_type (repeats, derives_from, reanalyzes, republishes, cites_as_evidence, responds_to), a fidelity saying what survived the crossing (faithful, strengthened, weakened, distorted, misattributed, unclear), a required evidence field holding the located passage in the asserting source where it draws on the target, so that an edge copied from a citation index has nothing to put there, and target_read. source_relationships is the thin claim-independent layer, facts about a pair of documents that hold whatever claim is traced (shares_authorship, republishes, version_of), which is what lets a map say five sources are three voices. claim_source_maps is the per-claim reader-facing account in the graph's voice, one row refreshed in place, with material gating whether it appears on the claim page at all; most claims' provenance is unremarkable and the row says so without cluttering the page. Readings and edges hang off the instance rather than carrying a claim id of their own, so a merge or a split moves them with the instance they describe (§5); the map is the one row that goes stale across a merge, and the Curator's notification says so.

The work is the Claim Steward's, under the Provenance skill (skills/provenance/SKILL.md), the first method skill: a skill that belongs to no domain and is carried on every run of the roles it addresses (see Skills). Its tools are in every Steward run: provenance_get_map (the instances with their ids, readings, and edges), provenance_read_source (a source's text, windowed; a document the graph holds no copy of is fetched through the guarded ingestion path and stored, so the quote check sees what the Steward read), provenance_record_reading, provenance_record_edge, provenance_record_source_relationship, and provenance_write_map. The Audit Agent carries the two reads. The skill text carries the judgment about when the work is worth doing (§19): a claim with two primary instances needs a sentence and an immaterial map; a claim whose verdict leans on many secondary instances warrants the full procedure. No per-run cap sits on these tools beyond the iteration budget; the rows are cheap and the judgment is the Steward's. The provenance surfaces on the claim page (the summary above the instances when material, the reading beside each instance, and the edges behind a disclosure, in words), on GET /claims/:id?information_depth=deep (source_map, provenance_edges, source_relationships, and each instance's reading), on the MCP get_claim tool with include: ["provenance"], and in the Steward's own get_claim_instances.

A claim is a canonical proposition with no home source, so the admin get_claim read returns the canonical record and an instance_count, never the instances; get_claim_instances pages through them. The Steward decides which instances the claim carries and in what order: detach_claim_instance removes one whose source does not assert, deny or pose the proposition (claim_instances.detached_at, with the reason in the audit trail; the row stays for the provenance record, and reattach_claim_instance restores it), and set_instance_order ranks the instances a reader should see first (display_rank). Every read that shows or weighs a claim's instances, from the claim page to the provenance story and the monitors, skips detached rows, and the rest follow the Steward's ranks, then the provenance story's order (origins first), then the date of the statement.

Contributions and governance

Anyone can contribute — but the graph is a governed space: open to suggestions, never to direct writes. A contribution targets a claim with a type (challenge, support, propose_merge, propose_split, propose_edit, add_instance, or propose_argument) plus content and evidence. Contributions flow through review (contribution_reviews), can be appealed (appeals), and escalated to arbitration (arbitration_results).

Two intake types extend the same machinery to brand-new content: propose_claim (a suggested claim plus its supporting argument) and propose_source (a document submitted for extraction). These have no target claim while pending — nothing touches the claims table — and only an accepted review materializes them: a proposed claim is canonicalized through the Matcher (so a duplicate or a negation lands on the existing node) and only then created live and handed to its Steward, with a deliberately conservative importance prior; a proposed source is only then queued for extraction. The review gate judges good faith and claim quality (is this a single, disputable, canonical-formable proposition?), never subject matter. Internal seeding by direct service callers (corpus runs, case studies) is the one path that writes without review.

A third family, the prize claim (claim_prize), is the one contribution type that carries files. It is created only by its own multipart route, never by POST /contributions, and its verification and money state live on a linked prize_claims row; the contribution row gives it the identity gate, the public record, review, appeal, arbitration, and audit that every other contribution has, so prizes need no second governance system. The flow is described under the prize-check worker.

Contributors are the account layer as well as the reputation layer; there is one account table, and everyone on it is a potential contributor. Reputation and kudos are kept as append-only event ledgers (reputation_events, kudos_events) with denormalized totals on the contributor row, so every score change traces back to the decision that caused it. Reviews can flag suspected bad faith, and a contributor's standing feeds back into how much their contributions are trusted. This is the machinery the governance agents operate; the rules they apply live in the operational policies below.

Assessment

The six statuses

Validity is expressed honestly, never as a binary. The system implements all six statuses the constitution defines:

Status	Meaning
`verified`	The evidence, examined directly, establishes the claim.
`supported`	The evidence favors the claim, but the examination is incomplete or the evidence is indirect.
`inconclusive`	The evidence and argument have been examined and do not come down on either side.
`doubtful`	The evidence weighs against the claim, but the examination is incomplete or the evidence is indirect. The mirror of `supported`.
`contradicted`	The evidence, examined directly, weighs against the claim.
`unknown`	Insufficient information to assess.

The statuses form an evidential ladder: they describe the evidence, not the debate. Until #524 the set held contested (credible evidence on multiple sides) and unsupported (no credible evidence found). contested described the disagreement rather than the evidence and let a verdict be avoided; unsupported merged two cases the ladder now separates (doubtful where evidence would exist if the claim were true, unknown where it would not be expected). Migration 0066 relabeled existing rows provisionally (contested to inconclusive, unsupported to unknown) and queued the affected claims for a fresh Steward pass.

How actively a claim is disputed is not a status. It is the claim's contestation (0..1, set by the Steward with importance), served on the claim and shown beside the verdict on the claim page, e.g. "Doubtful · actively disputed" at contestation 0.6 or above, "disputed" from 0.3.

The colour treatment in the UI is deliberately muted and never a traffic light: supported and verified are distinct shades of green, inconclusive is amber, doubtful a muted clay and contradicted a clay red, and the rest are warm neutrals. Meaning never depends on colour alone.

Judgment-based propagation

Assessment is a holistic judgment by the claim's Steward, not a mechanical roll-up of child statuses. An earlier design used hard aggregation rules ("if any required subclaim is unsettled, the parent is unsettled"). At scale that makes uncertainty infectious: almost every claim eventually inherits an unsettled subclaim somewhere deep in its tree, and the status field becomes useless.

Instead, the Steward weighs the status of subclaims across all arguments, the materiality of each subclaim to the parent's truth, and the strength of each argument as a whole, and documents the result in its reasoning. The Steward prompt gives guidance and worked examples rather than rules, and is explicit: do not mechanically propagate status changes; assess materiality first.

Propagation is therefore self-limiting. When a Steward materially changes an assessment, it notifies the Stewards of directly dependent claims, each of which re-judges with the same materiality test. Most changes are absorbed within a level or two, because a superior claim is rarely the right locus for a dispute about one of its subclaims.

The Agent Pipeline

Each agent is a model with a system prompt assembled in layers: the full constitution first, then the agent's specific role (whose prompt carries the role's operating standards, summarized in the policies below), then the domain skills active for the run, then the task. The constitution-plus-role prompt is sent as one cached block, plus one per active domain skill, so the constitution is paid for once per agent rather than once per call. The prompts live in src/llm/prompts/, the skills in skills/, and both are vendored verbatim into this site (see the agents and skills pages).

Processing stage

Ingestion runs three steps before governance takes over:

 Extractor ──▶ Matcher ──▶ onboard ──▶ Claim Steward
  read a       new claim    latch +     decompose + assess
  source for   or existing? enqueue     (a governance agent,
  its claims  (agentic search           below)
              over the graph)

Extractor: reads a source and emits the discrete, reusable claims it asserts, in canonical form, each with a provisional importance prior and a confidence that the proposition is a well-formed claim at all. It is a structured extraction call rather than a tool-use loop, and it is deliberately selective: the claims a reader would want checked, not every sentence. A low confidence floor drops obvious non-claims before they enter the graph — a backstop against garbage, not a quality judgment, which stays with the agents.
Matcher: the single decider of claim identity. For each proposition it searches the graph itself, under multiple framings including the negation, and decides match-or-create, recording the stance of the new appearance. It is also a tool the Steward and Curator call before creating anything. Two claims match when the same considerations bear on both; identical decomposition is a diagnostic, not the test. If the Matcher cannot reach a decision within its iteration budget, it defaults to "novel, low confidence": the failure mode is a duplicate the Curator can merge, never a lost claim.
Onboarding is not an agent. A small dispatcher latches the new claim so redelivered messages cannot double-process it, then enqueues its Steward.

Decomposition and assessment are not separate processing agents: they are the Claim Steward's job, because deciding what a claim depends on and whether those dependencies hold is one open-ended judgment that belongs to the claim's owner.

Governance agents

These act through tools over the life of a claim and the graph:

Claim Steward owns a single claim end to end: it decomposes the claim (calling the Matcher before minting any subclaim, so existing claims are linked rather than duplicated), maintains its canonical form and arguments, sets its authoritative importance, and assesses it, re-judging as evidence and depended-on claims change. Its triggers: first onboarding, a subclaim's assessment changing, an accepted contribution, a Curator change, a staleness check, and, on mathematical claims, a formal statement to publish or review, a completed solver attempt, or a prize claim to judge for fidelity. Effort scales with importance; consequential or disputed claims get deeper search (including bounded web search) and an adversarial second pass, minor settled ones a light touch. Decomposition terminates without a depth cap because shared ancestors get linked, not re-created; recursion is bounded economically by the importance brake, and a per-run cap on newly minted subclaims backstops a single runaway pass. On a quantity the same Steward keeps the definition, records the figures sources state and the readings of instruments, and sets the estimate (update_quantity_estimate) in place of a status. Any Steward may file a request, and keeps the requests that bear on its node alone.
Question Steward owns a question the way a Claim Steward owns a claim, in the same runner (the node's kind selects the role and tools): it keeps the wording and the presuppositions, gathers the candidate answers and subquestions, records where sources ask or answer the question, and gives the best answer it can with the question's standing (update_question_assessment).
Curator is the graph-level counterpart: it owns the connective tissue between claims, merging duplicates and counterparts the Matcher missed, splitting conflated claims (§5), and suggesting cross-claim edges for the owning Stewards to adopt. It runs as a funded curate ledger action (docs/allocation.md, "Maintenance and audit"): a Steward's escalation, the bounded reconcile-candidate scan, or an operator records a curation request on an anchor claim, a mandate values and funds the row, and one run reads every concern waiting on the anchor. Every structural operation lands in an append-only reconciliation log with enough payload to reverse it, and the Curator never overrides a Steward's verdict. It also corrects a node created as the wrong kind (change_node_kind, reversible), merges a claim that only states a value into its quantity, and keeps the requests that bear on several nodes or on none.
Contribution Reviewer evaluates each incoming contribution against policy (accept, reject, or escalate), including propose_argument contributions, and flags suspected bad faith. It is also the graph's admission gate: user-proposed claims and sources arrive as intake contributions, and its accept — judged on good faith and claim quality, never topic — is what admits them (materialization itself is mechanical: Matcher first, then claim creation or extraction).
Dispute Arbitrator resolves escalations and appeals through careful adjudication, the highest-stakes governance call.
Audit Agent is quality control over the governance system itself. Each run is invoked with an audit type (a decision audit of specific review decisions, a pattern analysis across recent ones, a contributor review, or an anomaly investigation) and a free-text context saying what prompted it. Runs are fed from two directions: event triggers at the places suspicion is generated (every arbitration overturn and every bad-faith flag requests a decision audit, at most once per contribution), and a scheduler that requests a periodic sweep over recent decisions plus a re-examination of any suspension that has stood unexamined too long — both idempotent through a DB dedupe key, so concurrent processes never double-run an audit. Each request opens an audit ledger action, funded by the platform's Governance mandate (or, for a prize audit, the bounty's review reserve) and never by the mandate the audit examines (docs/allocation.md, "Maintenance and audit"). Findings are persisted rows (audit_findings, attached to their audit_runs row), and every consequence — a re-review, a reputation adjustment through the ledger, a suspension — requires the finding that justifies it. A re-review first neutralizes the superseded decision's consequences (reputation, counters, kudos, a still-active bad-faith flag) so the fresh review's effects don't stack on the old ones. Audit suspensions are severe but not one-way: the suspended contributor can still appeal their own contributions, and the Arbitrator can lift a suspension whose basis an appeal dissolves. The allocation engine's decisions are judging too, so the Audit carries a read-only cross-cut of the action ledger and the mandate plan surface (inspect_ledger, services/ledger-inspection-service.ts): every row in scope in every status with its allocation history, and every plan item targeting those rows with its recorded standing checked against the row it names, so a report about the ledger is verified there instead of taken on the reporter's word.
Grantmaker designs and stewards funded mandates (docs/allocation.md): in conversation with a funder it surveys the territory, quotes honest costs, and drafts or refuses a mandate; once the mandate is live it takes autonomous review passes that value the open action ledger, grow the plan, pace the spend, and move budget between peer mandates.
Lookout is the cheapest administrator with the narrowest question: a standing watch a mandate's Grantmaker posts with a brief (scope in words, where to look, what warrants work), woken by a heartbeat or a trigger (the daily Crossref retraction poll, a poke), that reads the graph, the retraction record, and the open web and raises candidates: a claim to reassess, valued on the mandate's behalf up to a ceiling the Grantmaker delegated; a source to ingest, appended to the mandate's plan; a note for the next review pass. It judges relevance, never truth: it writes no assessment, sets no importance, and moves no money, and every flag is recorded with what became of it, so its precision is on the record and its Grantmaker can tighten or retire a watch that raises noise (docs/allocation.md, "Lookouts").
Consistency Checker reads assessments against each other, one region of the graph at a time (a topic tag, or the claims no sizeable tag covers), for what no single Steward sees: reasoning that conflicts with a neighbor's, evidence one claim records that another never weighed, verdicts that are not a defensible function of what they rest on (§21). Each find is a flag on the claim that looks wrong: its Steward is queued with the tension as context, and the checker's estimate that a pass would change something enters the formula's expected-gain term, so the pass is priced and funded like any other. Its sweeps are ledger actions the General mandate funds, capped per day; it writes no verdict and no valuation (docs/allocation.md, "Consistency sweeps").

Every one of these agents, the Matcher and the Extension Agent's chat included, also carries the issue tools: raise_issue, one channel, in the agent's own words, for a system failure, a gap in its own tools, or a concrete improvement idea arrived at from having just done the work; update_issue, the reporter's own edit path (re-rate the severity, add what it found since, or withdraw a report that was its own mistake); search_issues, a search of the reports on record by keyword and by meaning, or with no query a listing of what was seen most recently on a surface; and get_issue, one report in full with its triage note, its sightings, and the reports collapsed onto it. All are fire-and-forget (they always acknowledge and can never fail a run) and the policies say raising is never a substitute for acting. Reports land in agent_reports, not audit_log: they are about the machinery, not the graph, so they carry ids rather than content and are retained and purged separately. The record is navigable rather than gated (#432): a verbatim repeat collapses onto the dedupe key, an agent that has found its predecessor raises with joins (a sighting, recorded in agent_report_sightings with its own account, and a regression if the report was actioned, which reopens it), and a report written anyway comes back with the near reports on record as advice, with their status and triage note. Whether two reports are the same problem is otherwise the Audit Agent's call at triage, which collapses duplicates onto their representative. The search matches two ways and unions them: by meaning (cosine similarity over the stored embedding, at or above the bar) and by wording (every content word of the query, stemmed, appears in a report's title or body), wording hits first, so a report is never invisible to its own title even when it has no embedding or a paraphrase outscores it. A report recorded while the embedder was down is stored without a vector; the report embedding backfill worker (REPORT_EMBEDDING_BACKFILL_PER_TICK per tick) embeds such reports afterwards. Inside untraced work (the extension, the MCP's on-demand analysis) a report keeps its title, surface, and ids but its body is withheld, so the #356 rule holds for this channel too. External agents on the MCP surface get raise_issue only, attributed and rate-limited, matched against external-origin reports only, and their reports triage as testimony rather than findings.

The far end of the channel is GitHub. Every report written on first sighting is filed as an issue in GITHUB_ISSUES_REPO, labelled agent-generated with its kind and severity (github-issue-service.ts), written as the minerval-agents GitHub App: github-app-auth.ts signs a JWT with the App's private key and exchanges it for an hourly installation token, so the service never holds a long-lived credential (a plain GITHUB_TOKEN serves local runs); a joins with an account, an update_issue note, a regression, and every milestone count follow it there as comments, and a triage decision or a withdrawal closes it with the note. The filing is asynchronous and the report never waits on it; the github-issue-sync worker files the backlog — every open report with no issue yet, the ones raised before the sync existed included — a bounded batch per tick, and retries any filing that failed. The audit scheduler still requests a report_triage audit for each period that saw new reports; the Audit Agent clusters them by underlying gap, ranks by frequency and severity, and records a reading through triage_report that the service-scoped /reports API exposes to maintainers and that closes or annotates the issue. The triage note is stored whole up to TRIAGE_NOTE_MAX_LENGTH (4,000 characters); a longer one is refused with an error naming the limit, never cut (#439).

The administrators (Steward, Curator, Grantmaker, Contribution Reviewer, Dispute Arbitrator, Audit Agent) also carry a note_finding tool, the sibling channel for the other thing an agent notices in the course of its work: something people who hold a question would be better for knowing, because what most of them believe is wrong, or missing, or true for reasons the record now supplies. A finding is written in the graph's voice, rests on typed refs into the graph (claim, assessment, argument, contribution, check, attempt, formalization) that are checked to exist on write, carries an importance of its own from 1 to 10, and is published as written on the public findings page (/findings), where the platform's later writing draws from it. There is no cap and no triage queue; the restraint is the bar in the prompt block, and a run that notes nothing is the norm. What the tool does that the prompt cannot is check the record before it writes: the finding is embedded and searched against every finding on record, and on a near match nothing is written until the agent answers with joins (a sighting, counted and kept in its own words) or distinct_from (a new finding, saying what the earlier one lacks). Findings live in agent_findings with the same no-FK attribution snapshot as reports, so a note outlives the trace it came from; the read side computes a stale flag when a cited assessment is no longer the claim's current one. The extension chat, the MCP surface, the Extractor, the Matcher, the Lookout, and the solver do not carry the tool: an outside agent's discovery is a contribution, and a Lookout's is a flag for a Steward to judge.

One agent lives outside governance entirely. The Extension Agent is the read-only companion behind the browser extension: it judges the phrasings on a live web page against graph state (verdicts range from "egregious" to "fine") and powers the extension's chat, grounded in the same graph tools. It never writes to the graph.

The Lean checker

The checker is a Minerval-owned HTTP service, lean-checker, one container image per pin (lean-checker/; the image, its contract, and its first measurements are in lean-checker/README.md). It is an instrument, not an admin: it answers one mechanical question about one formal statement, holds no standing, and writes nothing to the graph; the API writes the lean_checks row from what it returns.

It runs in two lanes, and the distinction is the security model. The warm lane is a long-lived process with Mathlib imported, serving POST /v1/elaborate (statement publication and vacuity signals), POST /v1/scratch (the Steward's and the solver's iterative work), and POST /v1/search (a proxy to a Loogle mirror pinned to the same Mathlib); it takes semi-trusted input only, Steward- and solver-generated code, never a claimant's file, and its results are never a verdict. The cold lane is one fresh container per check from the pinned image: no network, read-only root, Mathlib read-only, a temporary work directory, memory and heartbeat limits inside Lean and a kill timeout outside, output capped and flagged truncated, no secrets in the image or the environment. POST /v1/check in prize mode queues a job and returns 202 {check_id}; GET /v1/checks/:id returns the record.

The verdict rule, in one paragraph. A submission is accepted when, under the statement's pin, it passes the static policy (no sorry, axiom, native_decide, unsafe, partial, extra import, custom metaprogramming, or disallowed set_option) and compiles with no errors; the target constant exists, is a theorem with no universe parameters, and has a type alpha-equivalent to the published Statement constant (or its negation, for a disproof), compared against the constant rather than an unfolded body so that no reduction is involved and nothing is arguable; its axiom closure lies within propext, Classical.choice, and Quot.sound, and the submission adds no axiom or opaque constant; no new constant is unsafe, partial, or externally implemented; and the new declarations replay through the kernel. It is rejected when any gate fails on the merits, and the failed gate is stated on the contribution page in plain words. It is error when the checker could not decide (timeout, memory, infrastructure), which is never evidence.

Pins. Each statement records its pin, and a submission is checked under the statement's pin, never a newer one. The platform keeps at most three live pins: the current monthly Mathlib tag, the previous one, and any pin still referenced by an open bounty. When the pin advances, a migration job re-elaborates every open statement and carries it forward without a new version only if the elaborated body and the closure of the constants it references hash the same; a statement with a live bounty never changes pin without a new version and the notice the rules require, whatever the hash says. Retired images stay in the registry so any historical verdict can be re-run.

No callback. The checker never calls the API. The API polls GET /v1/checks/:id from the prize-check worker and a recovery sweep, and the checker's security group allows it no egress to the API's or the load balancer's. A design in which the checker reported results back would have to survive the checker being compromised by a submission; polling survives it by construction. The API reaches the service over the VPC's private addressing with a bearer token from Secrets Manager. A second opinion from an outside hosted checker may be requested for a prize verdict and is recorded beside it, never decisive; disagreement between the two is an automatic human sign-off condition.

The Steward reaches the checker through four skill tools (lean_search, lean_elaborate, lean_check, publish_formalization), present exactly when the run carries the Mathematics skill and LEAN_CHECKER_URL is set, with per-run caps as backstops and no importance gate; a checker unreachable at run start yields no Lean tools and a note in the task, and a failed call mid-run returns a structured error the Steward routes around. Every checker call is metered into llm_usage as external usage (provider lean, the pin as the model) so it lands in the escrow accounting of whichever action funded it.

The solver

math_solver (src/llm/agents/math-solver.ts) is the platform's own prover: an agent over the same LLM seam, run as a ledger action of kind attempt_proof, metered through the same chokepoint, traced into agent_runs and agent_steps with tracing forced on, and published on the agents pages like every other prompt. It is an instrument, not an admin: it owns no claim, holds no standing, receives no constitution (only the skill's For the solver section and a harness block), and writes nothing to the graph. Its tools are fixed at run start: the three Lean tools, with lean_check bound to the attempt's statement so it cannot check against a different one; the provider's code-execution container as the computer-algebra toolkit (no network, and it cannot hold Mathlib, which is why Lean is a client tool); a notebook (notebook_write, notebook_read) backed by proof_attempts.notebook; and a terminal report with a strict schema. No web search: a proof found on the web is not the platform solving the problem.

What it may write: proof_attempts.notebook, lean_checks rows through lean_check, and its final report. It may not write to claims, assessments, arguments, argument_evaluations, claim_relationships, claim_instances, contributions, or any money table, and a unit test asserts that no such write occurs under withAgent("math_solver"). The harness validates the report: a proof outcome without an accepted check from this attempt is downgraded to partial, and a computational counterexample without a Lean disproof is a lead, not a result. The solver's narrative is untrusted; the tool log is the record.

Budgets and kill switches. Every attempt has a dollar ceiling (the cost estimate plus ATTEMPT_OVERAGE_FRACTION) read from the usage meter each turn, Lean and container time included; a wall cap and an iteration cap; a per-claim lifetime cap on attempt spend; and the mandate's day room like any allocation. Independent of any mandate, a durable breaker (checkSolverBudget, src/llm/solver-budget.ts) compares the day's solver spend across processes against SOLVER_DAILY_CAP_OWLS and stops new attempts when it is reached, because the in-memory tracker is per process and exempts attributed calls. Three kill switches: SOLVER_ENABLED (the worker exits its loop when false), a solver_paused row in platform_flags polled every turn so an operator halts a run without a deploy, and POST /admin/attempts/:id/cancel, also polled per turn. A halted attempt completes its action with the metered amount and keeps its notebook and transcript. The solver runs with server-side fallbacks off: a refusal records refused and stops, since a maximum-effort attempt silently continued on a different model would be a different product than the mandate funded.

Execution. A dedicated worker (src/workers/solver-executor.ts, its own process, npm run worker:solver) claims attempt_proof actions, runs the attempt under a usage context whose job is the funding job, completes the action with the metered amount, and then invokes the Steward directly with trigger attempt_completed on the strong model, under the same job, rather than enqueueing it: the steward queue coalesces triggers into one pending slot per claim and would run the result as an ordinary reassessment on the standard tier. The Steward reads the check rows the server wrote, re-checks a prize-bearing result with a fresh replay, judges fidelity, records the argument and the assessment, and, where a bounty is bound to the statement, either closes it as solved by the platform or retires a defective statement. Each turn updates the attempt's heartbeat and the action's updated_at, so the reopen sweep reclaims only a dead worker's row and marks its attempt orphaned. Every attempt is disclosed on the claim page, and its report and notebook are published before any bounty opens on the statement.

The researcher

researcher (src/llm/agents/researcher.ts, #298) is the general instrument, of which the solver is the special case. An administrator (the Claim Steward on its claim, a Grantmaker on a funded mandate) launches it with delegate_research: a brief written for a capable assistant that knows nothing of the claim and cannot ask a question back, a model tier, a dollar ceiling, an effort level, and whether the constitution is prepended to its prompt, which it is by default. The run is synchronous: the launcher's tool call blocks until the researcher reports, and the report is the tool result. It answers only to its launcher, owns nothing, and cannot launch instruments of its own. The report opens with an outcome (answered, partly answered, not answered, ill posed) and gives each finding with its locator and the words or numbers it rests on, so the launcher can check the load-bearing ones cheaply.

Tiers. The launcher chooses from three, named for the role rather than the vendor (RESEARCHER_STRONG_MODEL, RESEARCHER_STANDARD_MODEL, RESEARCHER_CHEAP_MODEL). Strong runs the long-run loop with effort and the task budget, for work where the best model class pays, such as a replication. Standard runs the ordinary loop on a Claude model. Both carry the code-execution sandbox (Python, no network), an Anthropic server tool. Cheap runs the ordinary loop on the cheap tier without the sandbox, and is the right choice for reading and mapping a large literature economically. Every tier has web_search (the server tool on Claude, the client-side OpenRouter search elsewhere) and read_page, the graph's read tools (search_claims, get_claim, get_decomposition, get_dependents), the Elicit search tools and the Pangram tools (pangram_check_ai, pangram_find_sources) where configured, a notebook, and, on a claim-scoped task, the Provenance skill's tools except provenance_write_map; on a mathematical claim with a checker configured it also has lean_search and lean_elaborate. The task message names the toolset, so the researcher knows what it can and cannot do, and the tool description names the tiers' affordances, so the launcher can brief it accordingly.

What it may write. Its notebook, the provenance tables through the skill's tools (a reading only for a source it opened, an edge only with the located passage), and the stored copy of a source it fetched. It has no tool that writes a claim, an assessment, an argument, an edge between claims, or an instance, and the executor refuses any tool name that was not offered and any provenance call that names a claim other than the run's, so a name read off a page or invented by the model runs nothing. A unit test holds the line. Its report is the instrument's narrative and is data: the launcher reads it against the sources it names and records what it concludes in its own reasoning (§9, §11). Provenance rows it recorded are reviewed by the Steward with provenance_get_map, and the map's reader- facing summary stays the Steward's to write. A run a Grantmaker launched writes only its notebook: a Grantmaker serves its mandate as well as the graph, and funding is kept apart from what the graph says (§19), so delegate_research from a mandate takes no claim_id and no examine_source_id, and the researcher withholds the provenance and examination tools from any run whose launcher is not a Claim Steward (launcherWritesGraph, src/services/research-run-service.ts).

Budgets and backstops. Every run has a dollar ceiling the launcher set from its own budget, read from the usage meter before every turn on both loops, a wrap-up notice at 85 percent of it, a wall cap (RESEARCHER_MAX_WALL_MINUTES) and a turn cap (RESEARCHER_MAX_TURNS); on the ordinary loop, a run past its ceiling or paused gets one last turn, in which every tool call is refused with the instruction to report, and is then stopped. What a run may be given is the launcher's own money, and the launcher is shown it in the tool description, in a refusal and in each result. A Steward pays from its pass's budget: every lane that runs a Steward sets one on the usage context (withPassBudget in src/llm/usage-context.ts), namely the allocations covering a ledger action, what is left of the background daily budget on the fallback lane, the cap a buyer was quoted for a paid order, the remaining budget of the job funding a money trigger or a deep decomposition, and a mandate's uncommitted budget. What is left is that budget less the pass's metered spend, the research it already launched included; a pass with no budget is not offered delegate_research. A Grantmaker pays from its mandate's uncommitted budget (budget less committed money, grantCommittedMicroUsd), which covers a run's budget_usd and any contribution_budget_usd; there is no deployment cap on either, and no count of runs per pass; a run's wall-clock time is capped (RESEARCHER_MAX_WALL_MINUTES) because its launcher waits on it, and the launcher is told so. A durable daily cap across processes (RESEARCHER_DAILY_CAP_OWLS) refuses launches past it for research no funder pays for (a pass with neither a user nor a job on its usage context), as the LLM budget tracker's limits apply only to unattributed calls; RESEARCHER_ENABLED turns the researcher on (it is off by default, as the solver is, and while off no launcher is offered delegate_research; production sets it in infra/lib/api-stack.ts), and a researcher_paused row in platform_flags halts a run in flight. The spend is metered under the launcher's usage context, so it lands on the claim and the job that funded the launching run, appears in llm_usage under agent researcher, and is added to the launcher's own cost meter, so the launching action is charged for it. The shared pieces of this harness, the stop reasons, the reminder fraction, the task budget, and the container metering, live in src/llm/instrument-harness.ts and serve the solver too.

Budget legibility. A model told "$2.00" does not know what that buys: it does not know its own price, and the dominant cost of a tool loop, every turn re-reading a growing history, is not one it can intuit. So the harness does the arithmetic (src/llm/instrument-harness.ts). The task message translates the ceiling, for a model with list rates in the table, into roughly how many turns of typical size it buys and what a page read early costs by the end of the run, and lists what each tool costs beyond its tokens; a model priced by its provider per call (the cheap tier) gets no up-front figure. After every turn, on both loops, the model sees a spend line read from the meter: what it has spent of what, and about how many turns remain at its last turn's cost. The line replaces the turn counter on the ordinary loop (toolUseLoop's turnNote) and rides beside the provider's task-budget countdown on the long-run loop.

The record. Each launch is a research_runs row: the brief, the model and tier, the ceiling, the spend, the turns, how the run ended (a run that ends on its own without calling report is no_report), the report, and the notebook. A launcher rereads its own runs with get_research_run: a Steward those on its claim, a Grantmaker those its mandate launched. On the claim page the runs are disclosed like solver attempts, with the model and the cost, and with the brief when the Steward launched the run (a Grantmaker's brief comes out of a private funder conversation and is withheld); the report and notebook are not served, and are not presented to readers as the graph's voice. The researcher's prompt is published on the agents pages with the constitution in its default position.

The prize-check worker and the money triggers

A prize claim is filed through POST /claims/:id/prize-claims (multipart, free of any owl charge), whose gate checks that the bounty is open, the statement version is current, the claimant is eligible, no live claim by the same claimant exists on this statement version, the rate limits and cooldowns hold, the attachment policy is met, and the static Lean policy passes as a word-boundary scan that turns away spam before anything runs. One transaction inserts the contribution with review_status = 'checking', a status the ordinary review pipeline and its recovery sweep ignore, the attachments, the prize_claims row (queued), and a prize_review action funded from the bounty's reserve (docs/allocation.md, "Prizes and attempts"); submitted_at is the priority timestamp.

The check runs first, before any agent, as a DB-backed job rather than inside a tool loop, because a prize check may run fifteen minutes and a strong-model run held idle for it would be lost with the process. The worker (src/workers/prize-check-pipeline.ts) claims the prize_review action, selects with FOR UPDATE SKIP LOCKED the oldest queued claim per statement version whose statement has no other claim in flight (strict per-statement serialization, so priority is never lost to a race), posts to the cold lane in prize mode, and polls the record, with a recovery sweep for rows checking past the reclaim window. accepted moves the claim to checked, the bounty to claim_pending, and the contribution to pending for the Reviewer; rejected closes the claim at stage check with the gate summary public, no review row, and no reputation event; error requeues up to a retry cap and then parks the claim in check_error, which holds the statement's queue until an operator resolves it, so an infrastructure failure never costs a claimant their priority. The worker then runs the Reviewer itself, under the bounty's reserve job, with the review claimed in the same statement that marks it pending (the ordinary review pipeline and its recovery sweep skip claim_prize, so no second Reviewer attributed to the claimant can race it), and on admission invokes the Steward directly on prize_claim.

The worker is two-step so a twenty-minute check never holds the runner's other lanes: a tick submits one claim to the checker and returns, and later ticks poll each check in flight once, landing the first finished one. The checker's check id has no column; it lives in an in-process map keyed by prize claim id, and after a restart the worker re-submits each checking row with force: false, which the checker dedupes against the record it already holds, recovering the id without a second run (a retry after a checker error forces a fresh run instead). Each poll heartbeats the row, so the reclaim sweep trips only for a dead worker. The same tick carries three bounded sweeps: a checked claim whose Reviewer run was lost is reviewed again under the reserve; a claim the Audit agent sent back (returned to in_review with its window cleared) is put in front of the Steward again for a fresh decision, a new decision id, and a new audit; and an in_review claim with no decision after 24 hours is re-invoked, at most once a day, and listed on the operator page under in_review_over_24h. Before any money trigger runs, the direct invocation resolves the claim's skills and refuses, loudly, when the trigger's tools (decide_prize_claim, publish_formalization, get_proof_attempt) are not among them, since a run without its own tool would decide nothing.

Six money triggers are invoked directly, never queued. formalize, formalization_review, prize_claim, prize_claim_voided, prize_window_closed, and attempt_completed each call runClaimSteward on STEWARD_STRONG_MODEL from the worker that owns the event (the engine executor for the first two, the prize-check worker, the window closer, and the solver worker), inside a usage context whose job is the funding job. The steward queue is bypassed on purpose: it coalesces into a claim's single pending slot and keeps an existing trigger over a new one, so a prize_claim arriving on a claim already pending reassessment would run as a reassessment on the standard tier, and a fidelity judgment on a prize must not depend on which variant won an auction. Production refuses to run any of them without STEWARD_STRONG_MODEL, the trigger is recorded on the run, and direct invocation means a prize review never waits behind the steward drain. The served model and whether a fallback ran are recorded on the decision; the Audit agent treats a fallback-served acceptance as a send-back.

After the Steward accepts, requestAudit is called with a dedupe key that carries the decision id (a re-acceptance after a send-back must be audited again), the challenge window opens, and promotePayable requires the window to have closed, an audit outcome without a send-back, and, where the amount, the importance, or any anomaly requires it, a human sign-off. The payout itself is mechanical (docs/accounts.md, "Prizes paid in owls").

Models

Model choice follows the value of the judgment, not a single default:

Agent	Production model
Tagger · Matcher · Lookout	GPT-6 Luna (via OpenRouter; `LOOKOUT_MODEL`, with a per-lookout override)
Extractor · Contribution Reviewer · Extension Agent	Claude Sonnet 5.5
Claim Steward · Curator · Dispute Arbitrator · Audit Agent · Grantmaker	Claude Opus 5.5
Solver (`math_solver`)	Claude Opus 5.5 at effort `max` (`SOLVER_MODEL`), fallbacks off
Researcher (`researcher`)	Chosen per launch by the administrator: Claude Opus 5.5 (`RESEARCHER_STRONG_MODEL`), Claude Sonnet 5.5 (`RESEARCHER_STANDARD_MODEL`), or GPT-6 Luna (`RESEARCHER_CHEAP_MODEL`)

The Matcher's judgment is narrow ("same proposition?") over candidates it retrieves itself, so a small model suffices; it is the first agent routed to a non-Anthropic model. The tagger's is narrower still ("which of these existing tags, at what grain?"), makes no epistemic call, and runs over every claim, so it shares that tier: the cheapest capable model, in production as in dev. The Lookout's question ("did something happen that warrants work in my scope?") is relevance, not truth, and it runs often, so it shares the tier too, and a Grantmaker can pin a stronger model on one lookout whose brief warrants it. The load-bearing epistemic work (stewardship, structural adjudication, arbitration, audit) runs on Opus 5.5, with a server-side fallback to Opus 4.8 so a safety-classifier refusal degrades gracefully instead of failing the job. Background assessments carry a standard-model and a strong-model variant on the action ledger, and the upgrade is bought exactly when its marginal gain justifies its marginal cost (docs/allocation.md) — so the most capable model is spent where it buys the most; what goes unfunded is the tail. The exception is the Steward's six money triggers (formalize, formalization_review, prize_claim, prize_claim_voided, prize_window_closed, attempt_completed), which always run on the strong model whatever variant the ledger funded, because a fidelity judgment with money behind it must not depend on an auction.

Providers

Every agent talks to a model through five functions in src/llm/client.ts — complete, completeWithTools, completeStructured, completeStructuredList, toolUseLoop. Those signatures speak the Anthropic dialect and never change, so switching an agent to another vendor is one env var and no code change: MATCHER_MODEL=gpt-5-nano, CURATOR_MODEL=qwen/qwen3-235b-a22b.

Which backend serves a call is decided by the shape of the model id (src/llm/providers/routing.ts, the single source of truth):

ID shape	Provider	Example
`claude-…`	Anthropic direct (`@anthropic-ai/sdk`)	`claude-opus-5-5`
`gpt-…` or `o<digit>`	OpenAI direct (Responses API)	`gpt-5.6-luna`, `gpt-5-nano`
contains `/`	OpenRouter (`vendor/model`)	`qwen/qwen3-235b-a22b`
anything else	rejected, at config load AND at call time	`us.anthropic.claude-…`

Bedrock/Vertex-prefixed ids resolve to nothing and are rejected with a specific message — they 404 against the Anthropic API (issue #11).

Each adapter in src/llm/providers/ speaks its own platform's native dialect rather than a lowest-common-denominator abstraction. Structured output, for instance, is native output_config.format on Anthropic, a strict json_schema text.format on OpenAI, and a forced respond function call on OpenRouter (the most portable mechanism across its model zoo). Each provider also carries its own temperature allowlist, since reasoning models reject sampling params.

OpenAI direct speaks the Responses API, not Chat Completions. Chat Completions was the earlier choice because it maps 1:1 onto the seam's Anthropic-shaped assistant turn; Responses wins anyway, because it is where OpenAI's hosted server-side tools (web search, code interpreter) live and where a reasoning model's chain of thought can be carried across turns of a tool loop. Every GPT-5 model is a reasoning model, so that second point is not optional. The adapter is stateless — store: false, full history resent every call, previous_response_id never used — which means reasoning only survives a tool loop if it is round-tripped explicitly: requests ask for include: ["reasoning.encrypted_content"], and the turn's whole output array (reasoning items with their encrypted_content, message items, function_call items) rides back through the seam's provider-opaque rawContent and is replayed verbatim into the next request's input, in position, with tool results appended as function_call_output items keyed by call_id. OpenRouter has no equivalent surface for our purposes and keeps the Chat Completions translation in providers/openai-dialect.ts; the dialect-independent helpers stay shared between the two.

Anthropic-only, by design: server tools, container-backed execution, ephemeral prompt-cache breakpoints, and the server-side Opus refusal fallback. No agent hard-codes a server tool: the agents that want web search (the Claim Steward, the Grantmaker in every mode — conversation, planning, review — and lookouts) ask tools/web-search-tool.ts for one web_search, which is the Anthropic server tool on a Claude model that runs it and, everywhere else, a client-side tool of the same name and shape that the loop executes through OpenRouter's web plugin (openrouterWebSearch in providers/openrouter.ts: one metered completion on the cheap tier whose citation annotations are the hits). A request that still carries a server tool to a non-Anthropic model fails immediately with a message naming the capability, rather than silently dropping it. OpenAI's own hosted tools are not wired up yet, but they are ordinary entries in the Responses tools array, so the slot for them is the one toResponsesTools already builds. OpenAI gets automatic prefix caching with a stable prompt_cache_key per agent instead of explicit breakpoints.

Metering stays at one chokepoint regardless of provider. Anthropic and OpenAI calls are priced from the rate table in src/llm/pricing.ts; OpenRouter reports its own computed cost per call, which overrides the table (no rate table can cover its zoo). Unknown model ids fall back to the top-tier rate so nothing ever meters as free. Every usage row records which provider served it.

Missing credentials fail as a clear configuration error at call time, not as an opaque 401: ANTHROPIC_API_KEY, OPENAI_API_KEY (shared with embeddings), and OPENROUTER_API_KEY.

The long-run loop

The five functions above are the provider-neutral contract every agent depends on, and they do not change for the solver. The seam has a sixth, longRunToolLoop, which is Anthropic-only by design: streaming (required at the 128,000-token output ceiling), effort up to max, a moving cache breakpoint on the last user message so a hundred-turn loop pays cache-read rates for its history rather than the full input price every turn, per-turn hooks (beforeTurn, afterTurn, reminder) through which the solver's dollar ceiling, wrap-up notice, and kill switches act, and fallbacks: "none". The history is append-only, which the preserved-thinking rule requires and a test pins. A second client with an hour-long timeout serves it, because the default 180-second timeout with retries would abort and re-issue a fifteen-minute turn several times, each billable. The adapter reports the served model, whether a fallback ran, and cache-read and cache-creation tokens on every result; the money decisions record the first two.

Workers and failure handling

Ingestion and the governance pipelines ride SQS queues in production and an in-memory runner locally, with identical handlers. Stewardship is the exception: it has no message queue at all. A claim's steward-state column marks it a CANDIDATE (idempotently — a claim re-triggered while already pending coalesces into one run), the action ledger holds its priced assess/reassess actions, and the executor runs whatever the allocations on that ledger cover (docs/allocation.md), claiming work with FOR UPDATE SKIP LOCKED so concurrent workers never collide. Prize checks and solver attempts are DB-backed jobs with workers of their own, never SQS messages: a check can run fifteen minutes and an attempt six hours, far past the queues' visibility timeout. Tagging is a third DB-backed lane: an untagged claim row is the work item, a scheduler tick leases and tags a small batch most-important-first, and the same drain is the backfill.

Failures are classified before they are counted. Transient API errors (rate limits, server errors, network, exhausted budget) requeue the claim untouched and do not count as attempts, and a run of consecutive transient failures trips a circuit breaker that stops the drain rather than poisoning healthy claims. Only genuine logic errors count toward the retry cap, after which the claim parks in an error state for inspection. The distinction exists because an earlier incident parked dozens of production claims over what turned out to be a billing hiccup.

Persistence

PostgreSQL, not a graph database

The graph is stored relationally in PostgreSQL, accessed through Drizzle ORM, not in a dedicated graph database. Claims are rows; decomposition is an adjacency table (claim_relationships) that holds only what propagation walks; a membership table (argument_subclaims) attaches each edge to the line(s) of reasoning it serves; arguments, assessments, instances and sources are their own tables. A relational store keyed by foreign keys is more than adequate for the tree-shaped reads the product needs, and it lets the same engine carry vector search and full-text search without a second system to operate.

Tree-building (src/services/tree-service.ts) walks the relationship table level by level with a visited set, so each node and edge is fetched exactly once even where shared subclaims give the DAG a diamond shape. The walk is bounded by a cap of 500 nodes per response (MAX_TREE_NODES); children dropped by the cap are flagged on their parent (children_truncated), never silently. Each edge is joined to its argument memberships, and the argument_id, argument_name, argument_stance, and argument_content are carried onto the node, so a client can group a claim's children by argument and render each argument's written form. An edge grouped under two arguments appears once in each group; its subtree renders at the first occurrence and is collapsed at the second, as for any shared subclaim.

Schema at a glance

claims ──< claim_relationships >── claims     (parent / child adjacency)
  │              │
  │              └──< argument_subclaims >── arguments ──▶ claims
  │                   (which arguments group     └──▶ argument_evaluations
  │                    an edge; none = basis)
  │                                         (inference verdicts; one is_current per argument)
  ├──< claim_links >── claims          (lateral see-also: related, rival
  │                                     explanation, counterpart; never walked)
  ├──▶ assessments        (verdict history; one is_current per claim)
  ├──▶ quantity_definitions, quantity_estimates, quantity_readings, quantity_cuts
  │                       (a quantity's definition and estimate history)
  ├──▶ question_assessments, answer_relations
  │                       (a question's standing history; how answers bear on each other)
  ├──< request_targets >── requests ──▶ request_work
  │                       (what someone could do; its keeper; work and outcomes)
  ├──▶ claim_instances ──▶ sources   (provenance: quote + context + stance)
  │        ├──▶ claim_instance_readings         (does the source bear its own assertion)
  │        └──▶ claim_provenance_edges ──▶ sources   (what the assertion draws on)
  │                    sources ──< source_relationships >── sources   (one voice, twice)
  ├──▶ claim_source_maps             (what the support rests on, in prose; shown when material)
  ├──< taggings >── tags             (topic vocabulary; polymorphic on subject_kind)
  └──▶ contributions ──▶ contribution_reviews ──▶ appeals ──▶ arbitration_results
                              contributors ─┘

Around that core sit the account and operations tables: contributors doubles as the account table, api_keys holds hashed keys, llm_usage meters every model call, reputation_events and kudos_events are the append-only score ledgers, reconciliation_events is the Curator's reversible audit log, audit_log is the Steward's append-only decision trail, audit_runs and audit_findings are the Audit Agent's run ledger and durable findings (the run ledger doubles as the dedupe gate for audit triggers), agent_reports (with agent_report_sightings) and agent_findings are the agents' own two channels to the outside (issues with the machinery, mirrored to GitHub; what they found about the world), and jobs tracks queued work. Mathematics adds claim_formalizations and lean_checks (the formal statements and every check), proof_attempts (the solver's runs), research_runs (the researcher's, #298), bounties (owls held against the escrow of the mandate that posted each one; there is no prize fund), prize_claims, prize_payouts, attachments, and platform_flags (operator switches such as solver_paused).

Attachments

Uploaded files (a prize claim's Lean source, documents, and data; a winner's tax form) live in Postgres, in attachments, with bytea bodies keyed by contribution_id: kind (lean_source, document, dataset, code, tax_form), the sanitized filename, a content type determined by magic bytes against an allowlist rather than from the client's header, size_bytes, sha256, visibility, and a storage discriminator, db today with s3 and a storage_key reserved. Bodies are restricted at submission; a Lean source becomes public when the Steward accepts the claim, because the challenge window needs it, and a rejected source stays restricted until the bounty closes so a near-miss cannot be patched by a second account and refiled. Nothing is executed or parsed except the Lean file inside the checker's sandbox; downloads are served as attachments with nosniff and a sandboxing content-security policy. Postgres is the right store at v1 volumes because one transaction landing the contribution, the attachments, and the prize claim is what makes the priority timestamp meaningful; the S3 path (a bucket, a gateway endpoint, presigned PUT and GET, a backfill) needs no schema change and is triggered by files over 10 MiB, attachment storage past about 5 GB, or a second region (docs/infrastructure.md).

Search: vectors and full text

Postgres carries both retrieval paths the pipeline needs. Each claim has a 1536-dimension embedding (a pgvector column) for semantic neighbour search and a tsvector column for keyword search. A query runs both recall paths at once: a claim is a candidate if it matches the keyword query or falls within embedding range. The two signals are deliberately not blended into one score; results are ordered by cosine similarity, with keyword rank as a tiebreak, and keyword matching serves to widen recall. If embedding generation fails, search degrades to keyword-only. Every path serves only active, unmerged claims. This hybrid search serves the public search API, the MCP search_claims tool, and the agents' general search tool. The Matcher's candidate retrieval is the exception: it uses embedding similarity alone, with a deliberately low floor, and widens recall by re-searching under multiple framings rather than by keyword rank. The tag vocabulary carries its own embeddings and is searched the same way (semantic with a lexical widening on the name), by the tagger before it mints and by /tags?mode=search; every claim search accepts a tag as a filter.

Serving Surfaces

The API

A Fastify service at api.claimgraph.io. Reads are public and unauthenticated: claim lookup and search, decomposition trees, dependents, assessment history, contributor profiles, the topic vocabulary, the findings feed (GET /findings, filterable by claim, tag, and importance), and requests (GET /requests, filterable by node and state). A claim's detail carries its kind, and a quantity's or question's detail its quantity or question. Anything that writes or spends model tokens (POST /sources, POST /claims/propose, contributions, appeals, the extension and MCP endpoints) requires a key. No user surface writes to the graph directly: proposed claims and submitted sources become pending intake contributions (HTTP 202) for the Contribution Reviewer, and only direct service callers — internal seeding — keep the immediate path. Interactive OpenAPI documentation is served at /docs on the API host.

The web app

This site, a Next.js app at minerval.ai. It talks to the API server-side with a service key, forwarding the signed-in user's identity through an acting-user header, so browser traffic never carries API credentials.

The browser extension

The extension analyzes the page you are reading: captured text flows through the Extractor and Matcher, then the Extension Agent judges each on-page claim against graph state and the verdicts are anchored as a non-destructive overlay, with a chat popup grounded in the same graph. Analysis answers immediately when the page was analyzed before; otherwise it returns a content hash the extension polls until the pipeline finishes. The work is metered to the user's account, and the key lives in the extension's background worker, never in the page.

Nothing the extension sends is persisted (#356). The page is whatever the reader has open, so the whole pipeline runs untraced (no agent_runs or agent_steps, whatever TRACE_LEVEL says), writes no sources or instances, and keeps its result only in an in-memory cache; chat is the same. The MCP's on-demand analysis tools follow the same rule. Only source submission, where the fetched document is public by construction, produces provenance and a transcript. The claim layer carries its own guard: the constitution (§2) holds that a claim is about the world, not a private person, and the Extractor is where that binds.

MCP

A remote MCP server, speaking streamable HTTP at POST /mcp on the API host, exposes the graph to agentic clients under the same accounts and quotas: tools for searching and reading claims (search_claims, get_claim, get_decomposition), for the pipeline's judgments (match_claim, extract_claims, assess_text), and for contributing (submit_contribution, get_contribution_status), for the machine-readable terms of a prize (get_bounty_terms), plus claim resources and fact-checking prompts. Clients authenticate with an API key or via the OAuth 2.1 authorization flow — the API acts as an authorization server for the /mcp resource, handing sign-in and consent to the web app — which is what lets hosted clients such as Claude.ai connect. Every call is attributed to an account either way.

Accounts, Keys, and Metering

Users and contributors are the same thing: one account table, one identity. Sign-in on the web app goes through Auth.js (GitHub or Google); the API never sees OAuth, only a stable external id of the form provider:subject, against which the account is provisioned on first sign-in.

API keys are prefixed epk_, stored only as hashes, shown once at creation, and scoped user or service. Every model call in the system is metered at the LLM client chokepoint: tokens are priced into micro-USD and recorded per agent, user, and key. Spending runs on the owl economy (docs/accounts.md, docs/allocation.md): one owl of spend covers a dollar of metered cost, agentic operations charge a cap that settles to the metered actual, free signup and monthly grants keep the entry free, and Stripe Checkout sells owl packs. Rate limits at the API boundary are a runaway backstop; the real spend guardrail is the owl balance and the escrowed budgets behind mandates.

Three credentials, not two, once money can move. The service key (MINERVAL_API_KEY) is deployed to the web tier and acts for any user through the acting-user header, so it cannot be the credential that moves money. Seven routes require an operator key (MINERVAL_OPERATOR_KEY), held outside the web deployment and used only from the operator's own session: the bounty confirmation (POST /bounties/:id/confirm), the prize-claim sign-off (POST /prize-claims/:id/sign-off), the void (POST /prize-claims/:id/void), the sanctions screening (POST /prize-claims/:id/screening), the owl grant (POST /prize-claims/:id/pay), the release of a check_error hold (POST /prize-claims/:id/retry-check), and the operator page (GET /operator/prizes). Three routes act for a winner and require both the dashboard session and a one-time code bound to the claim, the account, and the purpose (issued to the owner's session by POST /prize-claims/:id/code; email delivery is a transport change): the payee step (POST /prize-claims/:id/payee), the tax form (POST /prize-claims/:id/attachments), and the withdrawal of a claim (POST /prize-claims/:id/withdraw), so a leaked consumer or service key alone can neither redirect a prize nor abandon a winning claim. Every call to a writing route among these is written to audit_log on the claim with the credential kind and the acting person; the deposit, which has no claim, is appended to the prize_fund_deposits platform flag instead. The service key alone moves no money.

The same operator key funds platform mandates (#529), which are designed in a granting conversation like anyone's. The platform account holds that conversation through a key scoped to /grant-conversations alone, and nobody can act as the platform through x-acting-user. Its money moves through one route, POST /operator/platform-mandates/fund, called by the Platform mandates workflow after a merged pull request pins the Grantmaker's draft by hash. The route records that approval on the grant. See docs/allocation.md, "Platform mandates".

Deployment

The API runs as a container on ECS Fargate behind an application load balancer at api.claimgraph.io, with RDS PostgreSQL (pgvector) and SQS, all provisioned by CDK; a push to main deploys after typecheck and tests, and migrations run at container start. The web app deploys separately to Vercel at minerval.ai. Local development uses docker-compose Postgres and the in-memory queue runner, so the whole pipeline runs on a laptop with no AWS dependencies. The Lean checker is a separate stack with a security group that reaches nothing, and the solver worker is a second ECS service; both are described in docs/infrastructure.md.

Evaluation

The pipeline is graded against fixed corpora, not eyeballed. A harness runs the real application (real routes, real workers, drained to quiescence) over curated clusters of source documents and records every agent message. An LLM judge, deliberately a different model from the agent under test, then scores the assessed claims against the constitution on axes like readability, reasoning fit, impartiality, and decomposition granularity. Runs are compared release over release, so prompt and model changes are judged by what they do to the graph, not by how they read.

The role standards that follow turn the constitution's principles into concrete, per-agent rules: the acceptance criteria the Contribution Reviewer applies, the appeal and audit standards behind it, and the reasoning obligations every agent carries.

Operational policies

Role Prompts and Operating Standards

The Admin Constitution carries the doctrine every agent works under. Each agent's role prompt carries the operating standards for its own job: what a valid contribution of each type must contain, when to escalate, what an appeal has to bring, how audit remedies are chosen. There is no separate policy layer between the two. Where a role prompt and the constitution appear to diverge, the constitution wins, and the prompt needs fixing.

Decisions cite the constitution directly, by section (§2, §13, Part VIII). The policy_citations field on every review and arbitration holds those references, so the audit's consistency check and any later precedent search have a stable key that points at the grounding text rather than a paraphrase of it.

The authoritative text of every role prompt lives in src/llm/prompts/ and the agents pages show each one assembled exactly as its agent receives it. This document summarizes the standards for readers.

Prompt Architecture

Every admin agent's prompt follows this structure:

┌─────────────────────────────────────────────┐
│ LAYER 1: Admin Constitution (cached)        │
│ - Full text of admin_constitution.md        │
│ - Identical across all admin agents         │
│ - Establishes epistemic principles          │
└─────────────────────────────────────────────┘
                    │
                    ▼
┌─────────────────────────────────────────────┐
│ LAYER 2: Role-Specific System Prompt        │
│ - Defines the agent's job and its tools     │
│ - Carries the role's operating standards,   │
│   citing the constitution by section        │
│ - Ends with the shared raise_issue guidance │
│   and the catalog of domain skills          │
└─────────────────────────────────────────────┘
                    │
                    ▼
┌─────────────────────────────────────────────┐
│ LAYER 3: Domain Skills (cached, per skill)  │
│ - One block per skill active for the run    │
│ - The role's sections of skills/<name>/     │
│   SKILL.md, spliced by the loader           │
│ - Selected by the claim's recorded domains  │
│   (the mandate's, for the Grantmaker)       │
│ - Brings the domain's tools                 │
└─────────────────────────────────────────────┘
                    │
                    ▼
┌─────────────────────────────────────────────┐
│ LAYER 4: Task Context                       │
│ - The specific claim/contribution/dispute   │
│ - Relevant graph context                    │
│ - Conversation history (if applicable)      │
└─────────────────────────────────────────────┘

The constitution is read from admin_constitution.md at load time and the process fails loudly if the file is missing; there is no fallback summary, because a prompt silently missing its first layer would be worse than a crash. The constitution-plus-role prompt is sent as one cached block, plus one cached block per active domain skill, so the constitution is paid for once per cache window rather than once per call and a skilled run leaves the shared block's cache entry untouched.

Authority runs in layer order: the constitution, then the role and the operating standards its prompt carries, then the skill. A skill may sharpen how a role's obligations apply and add procedures and tools; it never loosens an obligation.

This architecture ensures:

Consistent application of epistemic principles across all agents
Clear separation between "how to think" (constitution), "what to do" (role), and "how that applies here" (skill)
Efficient caching of the constitution text across agent invocations

Role Standards

What follows summarizes the operating standards each governance role's prompt carries. The text the agent sees is on its page.

Claim Steward

The Steward owns a single claim's page end to end: canonical form, decomposition, arguments, and assessment (Part VIII). Its commitments:

Keep the canonical form the shortest neutral statement of the proposition as actually debated (§3), improving wording on its merits rather than preferring whichever formulation arrived first (§2).
Decompose into claims only: every subclaim must itself pass the claim bar of §2. Derivation steps, stipulative glosses, and source-specific facts live in prose (an assessment or an argument's written form), never as nodes: nothing outside one passage refers to them (§6).
Never mint a subclaim without asking the Matcher whether it already exists, under any wording or as its negation (Part VIII).
Structure runs upward as well as down. A claim that is an argument for, a meta-claim about, or a special case of a proposition the graph lacks gets that proposition minted above it (add_parent_claim), on the same claim bar and after the same Matcher call; where the proposition exists, the edge is proposed to its Steward (propose_parent_edge), since edges into another claim's decomposition are its owner's to write. A source about a controversy states the dispute and only presupposes what is disputed, so the Extractor mints the meta-claim and the Steward supplies the node it is about (#428).
Scale effort with importance (§19): a live crux earns deep structure and broad evidence search; a settled minor claim gets a light, careful pass.
Assess directly on the merits (§9), reaching a holistic verdict across all arguments rather than mechanically aggregating subclaim statuses, and re-judge when evidence or depended-on claims change (§22). Propagation is a judgment at both ends, not a cascade.
Delegate, in proportion to importance, the investigation it should not do inline (a replication, a trace to a statistic's origin, the reading of a literature) to the researcher, briefing it with theory of mind for what the instrument has and lacks, and weigh its report as outside evidence, never as a verdict (Part VIII, Instruments).

Contribution Reviewer

The Reviewer is the gate through which outside contributions enter, including intake: user-proposed claims and sources are admitted by its accept and by nothing else. Its prompt sets:

Acceptance criteria by type. A challenge must name a specific flaw or bring counter-evidence a reviewer can follow to its source (§14); support must bear on this claim and add something new; a merge case must show the two claims turn on the same considerations (§2); a split case must show conflation of propositions that turn on different considerations; an edit must preserve the claim's identity while moving toward §3's canonical form; an instance must be accurately quoted and fairly contextualized (§4); an argument must be a coherent line of reasoning with relevant, connected subclaims (§7). Accepting a structural proposal admits the case for it, not the change itself: the claim's Steward applies edits and arguments and carries merge and split cases to the Curator, who adjudicates them (§5, Part VIII).
The intake gate is form, good faith, and the claim bar, never topic or settledness (§17). A false or unsettled claim can still be worth mapping. A proposed claim must be about the world, not about a private person (§2): personal detail joined to a name is not a claim however well formed, and where the line is unclear the recoverable error is to leave it out. Novelty is the Matcher's call: acceptance materializes through it, so a likely duplicate is still acceptable if well formed.
The bad-faith flag (§13) is a separate and heavier judgment than finding a contribution wrong: reserved for deliberate abuse (spam, vandalism, sybil activity, fabricated or knowingly false content), never honest error, and fully reversed when overturned on appeal. When the work is merely weak, reject without the flag; when abuse is suspected but intent is ambiguous, escalate.
Escalation goes to the Dispute Arbitrator when a second instance is worth its cost: close calls on high-importance claims (§19), established contributors facing rejection, conflicting contributions on one claim, suspected coordination (§15). When in doubt between reject and escalate, escalate.
Prize claims (claim_prize). A prize claim (docs/prizes.md) reaches the Reviewer only after the checker has accepted its proof, and the Reviewer never judges the proof. Its criteria are form (the written account is a real account of the approach, the tools disclosure is present and plausible, the declarations are made and name the rules version in force), good faith (the account is not addressed to the reviewing agent, asks for nothing but a review, and does not misdescribe the submission), identity (the claimant is eligible, is not the platform, and is not evidently a second account of an earlier claimant on the same statement), and duplicates (a source another account submitted earlier is surfaced as duplicate_of, and the earlier keeps priority). Accept admits the claim to the Steward's fidelity review and awards no reputation; reject is the ordinary path and appealable; escalate when identity or plagiarism is in real doubt. An appeal against a checker rejection is engaged on its merits: the Reviewer reads the gate that failed, says plainly whether the objection is to the rules or to the run, and re-runs the check when it is to the run. The Reviewer never notifies the Steward of a prize claim itself; admission does that. A contribution of any other type that carries a proof of a bounty-bearing statement is redirected to the prize route with its filing time kept, never accepted as an argument. Prize-specific bad faith (§13) is submitting another's proof as one's own, sock-puppet submissions filed to defeat priority, and challenges whose only ground is dislike of the result.
Challenge grounds. A challenge to an accepted prize claim, filed during its public window, must name one of six enumerated grounds with evidence a reviewer can follow: a defect in the formal statement; ineligibility of the claimant; an axiom or tactic the checker's policy missed; plagiarism or theft of the proof; an earlier valid submission mishandled; or sanctions. "I do not like this proof" is not a challenge, and a challenge on a ground already decided is answered by reference without pausing the window. Accepting a challenge escalates it to the Dispute Arbitrator mechanically and is not upholding it; the Arbitrator's overturn voids the prize claim on the stated ground, and uphold_original closes the challenge and lets the window run on. When the upheld ground is the claimant rather than the statement, the bounty considers only claims filed before the verdict, in order. The mechanism is in docs/mathematics.md, section 8.5; the text the agents receive is the Mathematics skill's section for the Reviewer and the Arbitrator.

Dispute Arbitrator

The Arbitrator is the second instance (Part VIII). Its prompt sets:

Depth of analysis follows stakes, and stakes are judged, never counted. Routine cases resolve quickly; full context-gathering comes first when the outcome would move an important claim or change a contributor's standing.
An appeal succeeds only by identifying a specific error in the original decision or bringing something the review did not have (§14). Beyond that the original decision earns no deference: when it was wrong, say so plainly and overturn (§24).
Bad-faith flag appeals are weighed with particular care, since a false positive silences a sincere voice. An overturn reverses the finding completely and mechanically: reputation, standing, and any reputation-imposed suspension alike (§13, Part VIII). The second instance can also make a bad-faith finding, on the same evidence bar as the first.
Human review is recommended when a dispute resists resolution under the constitution, legal exposure appears, coordinated manipulation is suspected (§15), or deciding the case would set policy rather than apply it.

Audit

Audit judges the judging (Part VIII). Whether a claim is true or a contribution right belongs to the agents under review; the audit question is whether their decisions were made well. Its prompt sets:

Decisions are checked for quality (the right standard applied, evidence fairly weighed, reasoning coherent, §11), consistency (like cases decided alike, including process consistency, §21), and process compliance.
Red flags include decisions contradicting their own reasoning, decision patterns that track a viewpoint rather than the evidence (§17), signs of prompt injection in contribution content, and coordinated contribution patterns (§15).
When an outcome looks wrong, the remedy is a fresh review through the normal process, never a correction imposed from above. Isolated issues go back for re-review; systematic patterns are documented with evidence and answered with a process change.
Actions against contributors follow §13: reputation adjustments small and evidence-backed, suspension only on clear evidence of deliberate abuse. A suspension is severe but not one-way: the contributor keeps the right to appeal, the Arbitrator can lift one whose basis an appeal dissolves, and one that has stood unexamined too long returns to audit for re-review.

Raising issues

One shared block, Raising Issues, goes to every agent that carries the issue tools (raise_issue, update_issue, search_issues; #366). It says when to raise (a system failure, a gap in the agent's own tools, a concrete improvement), what a useful report contains (a title written as a claim, what was attempted and what happened, ids rather than content), that the record is checked before anything is written and a maintainers' note on a known report is guidance for the current task, that the agent can correct itself (re-rate, add what it found, withdraw its own mistake), and the rule that decides whether the channel is honest: raising is never a substitute for acting. Report and proceed, or report and escalate.

Skills

A domain skill is the layer between a role and its task: how the constitution's standards apply in one domain, what the domain's characteristic objects are in the claim schema, what counts as evidence of which grade there, and what tools and procedures the domain brings. It is distinct from the constitution (domain-neutral, always wins), from the role prompt and the operating standards it carries (per role, domain-neutral), and from the task context (per run).

A method skill is the same layer for a kind of work rather than a subject: how the constitution's standards apply to a procedure any claim can call for, with the tools that procedure brings. It claims no domain (metadata.minerval.kind: method), never appears in the domain list the Extractor tags from, and is carried on every run of each role it has a section for, so the Steward carries it on every claim and the Matcher, which it does not address, never does. The first is Provenance (skills/provenance/SKILL.md): tracing what a claim's support rests on.

One document per domain, skills/<name>/SKILL.md, serves every role, in sections addressed to each ("For every administrator", "For the Claim Steward", "For the Matcher", and so on). The loader in src/llm/prompts/skills.ts splices each role's sections into a block headed # Domain skill: <Name> (version N) that follows the role in the system prompt. The skills pages show every skill verbatim, the composition table, and the exact block each role receives. skills/<name>/tools.json declares the tools a skill brings (the Mathematics skill's Lean tools); they join a run's toolset exactly when the skill is active, and their executors live in code with a startup check that every declared tool has one.

Selection is a recorded judgment, not a filter:

Claim-scoped runs (Steward, Curator, Reviewer, Arbitrator, Audit) read the claim's domains: the Extractor emits a prior, a new subclaim inherits its parent's tags, and the Steward's set_claim_domains is the authoritative call. The Reviewer and Arbitrator reach the claim through the contribution; Audit takes the union over the claims in the decisions under review. The Matcher receives the domains its caller knows, and the Extractor always carries every skill's Extractor section so it can tag.
Mandate-scoped runs (the Grantmaker's conversation and review passes) read the mandate's skills. A mandate's skills select the Grantmaker's view and nothing else: funding never selects a prompt for any agent that writes to the graph (§19).
Method skills are not selected at all: every claim-scoped run carries each one whose composition table gives the role a section.

Precedence follows the rule stated above for role prompts: where a skill appears to diverge from the constitution or the role, the constitution wins, the role wins, and the skill is defective. A skill may refine how a role's obligations apply and may add procedures and tools; it may never remove an obligation. Two skills active at once are both spliced, in alphabetical order, and a conflict between them is a defect to fix in the texts. Assessments and agent runs record the skills they were made under, so the standards applied to any verdict are a query.

Reasoning and Its Audiences

Every admin judgment is accompanied by its reasoning: what evidence was considered, how competing evidence was weighed, what assumptions were made, what uncertainties remain, and what new evidence would change the conclusion (§11). No agent says "this claim is verified" without showing why. There is no fixed template; the obligation is to the content, not a format.

Assessments address two audiences, and the system stores both:

A reader-facing summary, written in the voice of the graph (§12): plain encyclopedic English that walks through the evidence and states the verdict, with the machinery invisible.
The full reasoning behind the verdict (the reasoning_trace field), preserved as the audit record: it may discuss tools used, subclaims consulted, and the weighing itself, and it is what the Audit agent checks reasoning against.

Implementation Notes

Constitution loading

The constitution is loaded once from admin_constitution.md (src/llm/prompts/constitution.ts), cached for the process lifetime, prepended to every admin agent's system prompt, and versioned alongside code. Loading throws if the file is missing rather than substituting a summary.

Role prompts

Each role's prompt is one file in src/llm/prompts/ (contribution-reviewer.ts, dispute-arbitrator.ts, audit-agent.ts, and so on), carrying the role's job, tools, and standards together. Two small blocks are shared: raising-issues.ts is the raise_issue guidance, and bad-faith.ts renders the bad-faith categories from the same enum the tools validate against, so the prompt and the schema cannot disagree.

Versioning

Constitution and role prompts are versioned together. When the constitution changes, every prompt surface is reviewed for compatibility, and the corpus evaluation (see the architecture document) is the check that a prompt change improved the graph rather than just the prose.

Vendoring

scripts/sync-frontend-content.ts copies the constitution, the architecture document, and this document verbatim into the web frontend and regenerates the agent prompt pages and the skill pages from the real prompt code and the skill files. It is re-run whenever any of them changes, so what the site shows is what the agents run; a unit test regenerates the content and fails when the vendored copy has drifted.

When a Decision Falls Short

When an agent's decision fails the standards:

Audit detection: the Audit agent flags it, citing the constitution section it fails
Re-review: the decision is sent back for a fresh review through the normal process
Learning: if the failure is systematic, the remedy is a documented process change, not a quiet correction

These are not failures of the agent but signals that the system needs attention. The goal is improvement, not punishment.