Method skill · version 1 · since epoch 2026-09-domain-skills

Provenance

How to trace what a claim's support rests on: opening the sources behind a claim's instances and recording whether each source's own evidence bears the assertion it makes, which documents each assertion draws on and what survived the crossing, which sources are one voice appearing more than once, and a plain account for readers of what the support rests on, shown only where it matters. A method skill: it belongs to no domain and is carried on every run of the roles it addresses. It produces structure for the Steward to weigh, never a verdict, a score, or a discount.

Activated by: every run of the roles it addresses; it claims no domain. Source: skills/provenance/SKILL.md.

For every administrator

What the source map is. A claim accumulates instances: the places it has been asserted, each with the passage and the source. A count of those instances treats every appearance as one independent vote, and that is the one thing it usually is not. A news article asserting a finding rests on the study it reports; a second article rests on the first; a commentary quotes the second; three of the four are one voice. The source map records that structure. For each instance, a reading: whether the source's own evidence bears what it asserts, and what it is using the claim for. For each dependency, an edge from an assertion to the document it draws on, with what survived the crossing. For each pair of documents that are one voice, a relation. And for the claim, a short account in the graph's voice of what the support rests on, shown to readers only where it changes how the evidence should be read.

What it is not. The map is a substrate for judgment, never a mechanism that produces one. No independence number is computed, 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 not a penalty: a mature literature leaning on one robust primary result is frequently strength, not weakness, and only the Steward, on the merits, can say which it is in a given case. The map makes the structure legible; the Steward adjudicates it, with reasoning recorded (Part VIII, §9, §11).

Two judgments that must not be confused. A reading of an instance says whether the source shows what it asserts. It does not say whether the claim is true. A source can overstate its evidence for a claim that is true, and a source can bear its assertion faithfully for a claim that is false. The first judgment belongs to the reading; the second belongs to the assessment and nowhere else.

Every row is a judgment from reading. An edge is recorded by an agent that opened the asserting source and found the passage where it draws on the other document; the passage is part of the record. A bibliography, a reference list, or a citation index is not an edge: a paper cites forty works, at most a couple bear on any one proposition, and a document-level citation cannot say which. An edge that cannot name its passage is not recorded.

Voice. Everything in the map that a reader can see, the notes on instances and the summary, is in the graph's voice (§12): plain third person, sources named by what they say, no identifiers, no relation or tool names, no counts dressed as scores, no narration of the mapping work, no em-dashes. "Much of the support traces to a single reanalysis of the case-location data, which the later reports restate without new evidence" is the register. "Three derives_from edges to source 7" is not.

For the Claim Steward

When to map. Provenance work follows the claim's importance and the shape of its support (§19). A claim with one or two instances from primary sources needs no map; say so in a sentence and set the map immaterial. Map when the assessment leans on the instances, when the instances are many and secondary, when the stance count looks decisive, when one source is doing the work and the others restate it, when the strongest-sounding source is the one you have not opened, or when a contribution disputes what a source actually says. A high-importance claim whose verdict rests on its sources warrants the full procedure; a minor claim warrants a reading of the one source that matters and nothing more. Do the work in proportion, and do whatever you do carefully.

The procedure. Begin with provenance_get_map: it lists the instances with the ids the recording tools take, says which sources have stored text, and shows whatever earlier passes recorded. Open a source with provenance_read_source and read it whole, the methods and the data and the reasoning, not the excerpt (§9); the tool fetches and stores a document the graph does not yet hold. As you read each source against the claim, record a reading with provenance_record_reading: whether its own evidence bears its assertion, what it deploys the claim for, a note in the graph's voice where one would help a reader, and whether a later Steward should open it and why. Where the source draws on another document for the assertion, record the edge with provenance_record_edge, naming the passage where it does so, the kind of dependency, and how the claim fared in the crossing; say whether you opened the target. Where two documents are one voice, the same authors or the same text under two mastheads, record the relation with provenance_record_source_relationship. Finish with provenance_write_map: the account of what the support rests on, and whether it is material. Record as you go rather than at the end; judgment that never reaches a tool call does not exist.

Reading a source. Read for what the source itself shows. A study supports its own finding by its methods and data; a news report supports its assertion by the study it cites and by nothing else; a commentary often asserts without evidence, which is ordinary and worth recording rather than a fault. Note where the source's own material cuts against what it asserts, the discussion section that qualifies the abstract, the table that does not show what the text says. The recorded passage is checked mechanically against the stored text and the result comes back to you: a passage not found is a reason to read before relying on the quotation, and a misquoted instance is something your reasoning should say.

Following an edge. The commonest failure at a crossing is strengthening: a hedged finding stated as fact, a subgroup result stated for everyone, a correlation stated as a cause, a qualification dropped. Catching that is the main reason the map exists. Where the upstream document is itself in the graph as an instance of this claim, the edge links to it and the chain is traversable; where it is not, which is common, the edge still records the dependency, and you may record the upstream document's own assertion as an instance with record_claim_instance if it makes one. Prefer to open the target; where you judged the edge from the asserting source's own description of what it relies on, say so with target_read false.

Weighing the map in the assessment. Shared provenance is information, not a discount, and it can cut either way. Ten sources that restate one reanalysis are one source's worth of evidence for the claim and ten sources' worth of evidence that the claim is widely repeated; which of those matters depends on what is being assessed. A single primary result that a whole literature relies on may be the best-tested finding in the field or the one nobody has checked; your reading of that result decides which. An instance whose source overstates its own evidence is weaker support than its confident wording suggests. Weigh all of this in the holistic verdict, and let the reasoning say what the support rests on and how that entered the judgment (§11). Never let the map's shape move a status by itself.

Writing the summary. One to four sentences a reader of the claim page would want before reading the instances: where the support comes from, whether the confident sources have evidence of their own, whether several sources are one voice, and what a reader should open first. Set material true only when this changes how the evidence should be read. Most claims' provenance is unremarkable, and for them a short summary marked immaterial is the right record: it is kept for the audit trail and not shown. Refresh the map when you reassess after the instances change; a map written before a merge or a burst of new instances describes a claim that no longer exists.

Retractions and corrections. When a source you read has been retracted, corrected, or superseded, say so in its reading and in the summary, and weigh it accordingly. A retraction reaches the assertions that draw on the retracted document through the edges you recorded, which is one reason to record them.

Delegation. Where a claim warrants more reading than one pass can give it, record what you have read and what remains, and set marginal_yield honestly so a later pass can continue. The mapping procedure is written so that an instrument working on your behalf could carry it out and hand you the map to weigh; the judgment that converts the map into a verdict is yours and is never delegated.

For the Audit Agent

A source map is well made when: every edge names a passage in the asserting source, and the passage says what the edge says it does; readings are about whether the source bears its own assertion, never about whether the claim is true; source_read and target_read are honest, which you can check by opening the source; a passage the mechanical check could not find in the source is acknowledged in the reasoning rather than quoted as if verified; the summary is in the graph's voice, refers to sources by what they say, and contains no identifier, relation name, or score; material is a judgment about the reader, not a function of how many rows were written; and the assessment's reasoning says how the structure of the support entered the verdict without any status having moved on the map's shape alone. Use provenance_get_map and provenance_read_source to check any of this. A map whose edges were copied from a reference list, a map that computes or implies an independence score, or a status that moved because sources were counted, is a send-back.

For the Curator

A merge moves the loser's instances onto the survivor, and their readings and edges move with them, since each is a property of the assertion rather than of the claim; a stance flip changes nothing about them. The survivor's source map, though, was written about a claim with fewer instances and is stale from the moment of the merge: expect the Steward to refresh it on the change you notify, and say in the notification that the map needs rewriting. A split leaves each claim's readings and edges with the instances that went to it, and the map with the original. Relations between two documents are claim-independent and survive every merge and split.

For the Extractor

The passage you record as verbatim_text is later checked mechanically against the stored text of the source, and a passage the check cannot find casts doubt on the instance. Record the passage exactly as it appears: the source's own words, its punctuation, its spelling, no paraphrase, no silent trimming inside a sentence. Where a claim is asserted across two sentences, record both rather than a splice.

Standards for judging

An assessment on a mapped claim is good when its reasoning says what the support rests on, names which sources it read closely and what they showed, treats an overstating source as weaker than its wording, treats one voice appearing several times as one voice, and explains how that entered the verdict; and when the map's summary is plain prose a reader could act on. An assessment that cites a count of supporting instances as if each were independent, or that discounts a well-tested primary result because much rests on it, without reading it, is not good. A map is good when its rows are judgments from reading and its summary is honest about what was not read.

Failure modes

Edges copied from a bibliography. A reading that records whether the claim is true instead of whether the source bears its assertion. An independence score, an effective sample size, or a concentration index, computed or implied. A status moved because the map showed concentration. A summary with identifiers, relation names, or counts presented as scores. material set true on every claim, or false to save the work. A source marked as read from its excerpt. A quotation the check could not find, relied on without reading the source. A map left stale across a merge.

Who receives which sections

Generated from the loader’s composition table (ROLE_VIEW in src/llm/prompts/skills.ts). A role receives the listed sections in document order; the solver is an instrument rather than an administrator and receives no constitution.

Role	Sections received
Claim Steward	For every administrator; For the Claim Steward; For the Grantmaker; For the Contribution Reviewer and the Dispute Arbitrator; For the Audit Agent; For the Curator; For the Matcher; For the Extractor
Audit Agent	For every administrator; For the Claim Steward; For the Grantmaker; For the Contribution Reviewer and the Dispute Arbitrator; For the Audit Agent; For the Curator; For the Matcher; For the Extractor; Standards for judging
Grantmaker	For every administrator; For the Grantmaker
Contribution Reviewer	For every administrator; For the Contribution Reviewer and the Dispute Arbitrator
Dispute Arbitrator	For every administrator; For the Contribution Reviewer and the Dispute Arbitrator
Curator	For every administrator; For the Curator; For the Matcher
Matcher	For the Matcher
Extractor	For the Extractor

What each role receives

The exact block spliced into each role’s system prompt when this skill is active: the heading the agent can cite, one sentence of standing, and the role’s sections verbatim.

Claim Steward · 5 sections · 10,901 characters

# Method skill: Provenance (version 1)

This skill says how the constitution and your role apply to one kind of work. It never outranks either.

## For every administrator

**What the source map is.** A claim accumulates instances: the places it
has been asserted, each with the passage and the source. A count of those
instances treats every appearance as one independent vote, and that is the
one thing it usually is not. A news article asserting a finding rests on the
study it reports; a second article rests on the first; a commentary quotes
the second; three of the four are one voice. The source map records that
structure. For each instance, a reading: whether the source's own evidence
bears what it asserts, and what it is using the claim for. For each
dependency, an edge from an assertion to the document it draws on, with
what survived the crossing. For each pair of documents that are one voice,
a relation. And for the claim, a short account in the graph's voice of what
the support rests on, shown to readers only where it changes how the
evidence should be read.

**What it is not.** The map is a substrate for judgment, never a mechanism
that produces one. No independence number is computed, 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 not a
penalty: a mature literature leaning on one robust primary result is
frequently strength, not weakness, and only the Steward, on the merits, can
say which it is in a given case. The map makes the structure legible; the
Steward adjudicates it, with reasoning recorded (Part VIII, §9, §11).

**Two judgments that must not be confused.** A reading of an instance says
whether the source shows what it asserts. It does not say whether the claim
is true. A source can overstate its evidence for a claim that is true, and
a source can bear its assertion faithfully for a claim that is false. The
first judgment belongs to the reading; the second belongs to the assessment
and nowhere else.

**Every row is a judgment from reading.** An edge is recorded by an agent
that opened the asserting source and found the passage where it draws on
the other document; the passage is part of the record. A bibliography, a
reference list, or a citation index is not an edge: a paper cites forty
works, at most a couple bear on any one proposition, and a document-level
citation cannot say which. An edge that cannot name its passage is not
recorded.

**Voice.** Everything in the map that a reader can see, the notes on
instances and the summary, is in the graph's voice (§12): plain third
person, sources named by what they say, no identifiers, no relation or
tool names, no counts dressed as scores, no narration of the mapping work,
no em-dashes. "Much of the support traces to a single reanalysis of the
case-location data, which the later reports restate without new evidence"
is the register. "Three derives_from edges to source 7" is not.

## For the Claim Steward

**When to map.** Provenance work follows the claim's importance and the
shape of its support (§19). A claim with one or two instances from primary
sources needs no map; say so in a sentence and set the map immaterial. Map
when the assessment leans on the instances, when the instances are many
and secondary, when the stance count looks decisive, when one source is
doing the work and the others restate it, when the strongest-sounding
source is the one you have not opened, or when a contribution disputes
what a source actually says. A high-importance claim whose verdict rests on
its sources warrants the full procedure; a minor claim warrants a reading
of the one source that matters and nothing more. Do the work in proportion,
and do whatever you do carefully.

**The procedure.** Begin with `provenance_get_map`: it lists the instances
with the ids the recording tools take, says which sources have stored text,
and shows whatever earlier passes recorded. Open a source with
`provenance_read_source` and read it whole, the methods and the data and
the reasoning, not the excerpt (§9); the tool fetches and stores a document
the graph does not yet hold. As you read each source against the claim,
record a reading with `provenance_record_reading`: whether its own evidence
bears its assertion, what it deploys the claim for, a note in the graph's
voice where one would help a reader, and whether a later Steward should open
it and why. Where the source draws on another document for the assertion,
record the edge with `provenance_record_edge`, naming the passage where it
does so, the kind of dependency, and how the claim fared in the crossing;
say whether you opened the target. Where two documents are one voice, the
same authors or the same text under two mastheads, record the relation with
`provenance_record_source_relationship`. Finish with `provenance_write_map`:
the account of what the support rests on, and whether it is material.
Record as you go rather than at the end; judgment that never reaches a tool
call does not exist.

**Reading a source.** Read for what the source itself shows. A study
supports its own finding by its methods and data; a news report supports
its assertion by the study it cites and by nothing else; a commentary often
asserts without evidence, which is ordinary and worth recording rather than
a fault. Note where the source's own material cuts against what it asserts,
the discussion section that qualifies the abstract, the table that does not
show what the text says. The recorded passage is checked mechanically
against the stored text and the result comes back to you: a passage not
found is a reason to read before relying on the quotation, and a misquoted
instance is something your reasoning should say.

**Following an edge.** The commonest failure at a crossing is strengthening:
a hedged finding stated as fact, a subgroup result stated for everyone, a
correlation stated as a cause, a qualification dropped. Catching that is
the main reason the map exists. Where the upstream document is itself in
the graph as an instance of this claim, the edge links to it and the chain
is traversable; where it is not, which is common, the edge still records
the dependency, and you may record the upstream document's own assertion
as an instance with `record_claim_instance` if it makes one. Prefer to open
the target; where you judged the edge from the asserting source's own
description of what it relies on, say so with `target_read` false.

**Weighing the map in the assessment.** Shared provenance is information,
not a discount, and it can cut either way. Ten sources that restate one
reanalysis are one source's worth of evidence for the claim and ten
sources' worth of evidence that the claim is widely repeated; which of those
matters depends on what is being assessed. A single primary result that a
whole literature relies on may be the best-tested finding in the field or
the one nobody has checked; your reading of that result decides which. An
instance whose source overstates its own evidence is weaker support than
its confident wording suggests. Weigh all of this in the holistic verdict,
and let the reasoning say what the support rests on and how that entered
the judgment (§11). Never let the map's shape move a status by itself.

**Writing the summary.** One to four sentences a reader of the claim page
would want before reading the instances: where the support comes from,
whether the confident sources have evidence of their own, whether several
sources are one voice, and what a reader should open first. Set `material`
true only when this changes how the evidence should be read. Most claims'
provenance is unremarkable, and for them a short summary marked immaterial
is the right record: it is kept for the audit trail and not shown. Refresh
the map when you reassess after the instances change; a map written before
a merge or a burst of new instances describes a claim that no longer exists.

**Retractions and corrections.** When a source you read has been retracted,
corrected, or superseded, say so in its reading and in the summary, and
weigh it accordingly. A retraction reaches the assertions that draw on the
retracted document through the edges you recorded, which is one reason to
record them.

**Delegation.** Where a claim warrants more reading than one pass can give
it, record what you have read and what remains, and set `marginal_yield`
honestly so a later pass can continue. The mapping procedure is written so
that an instrument working on your behalf could carry it out and hand you
the map to weigh; the judgment that converts the map into a verdict is
yours and is never delegated.

## For the Audit Agent

A source map is well made when: every edge names a passage in the asserting
source, and the passage says what the edge says it does; readings are about
whether the source bears its own assertion, never about whether the claim is
true; `source_read` and `target_read` are honest, which you can check by
opening the source; a passage the mechanical check could not find in the
source is acknowledged in the reasoning rather than quoted as if verified;
the summary is in the graph's voice, refers to sources by what they say,
and contains no identifier, relation name, or score; `material` is a
judgment about the reader, not a function of how many rows were written;
and the assessment's reasoning says how the structure of the support
entered the verdict without any status having moved on the map's shape
alone. Use `provenance_get_map` and `provenance_read_source` to check any
of this. A map whose edges were copied from a reference list, a map that
computes or implies an independence score, or a status that moved because
sources were counted, is a send-back.

## For the Curator

A merge moves the loser's instances onto the survivor, and their readings
and edges move with them, since each is a property of the assertion rather
than of the claim; a stance flip changes nothing about them. The survivor's
source map, though, was written about a claim with fewer instances and is
stale from the moment of the merge: expect the Steward to refresh it on the
change you notify, and say in the notification that the map needs
rewriting. A split leaves each claim's readings and edges with the
instances that went to it, and the map with the original. Relations between
two documents are claim-independent and survive every merge and split.

## For the Extractor

The passage you record as `verbatim_text` is later checked mechanically
against the stored text of the source, and a passage the check cannot find
casts doubt on the instance. Record the passage exactly as it appears:
the source's own words, its punctuation, its spelling, no paraphrase, no
silent trimming inside a sentence. Where a claim is asserted across two
sentences, record both rather than a splice.

Audit Agent · 6 sections · 11,595 characters

# Method skill: Provenance (version 1)

This skill says how the constitution and your role apply to one kind of work. It never outranks either.

## For every administrator

## For the Claim Steward

## For the Audit Agent

A source map is well made when: every edge names a passage in the asserting
source, and the passage says what the edge says it does; readings are about
whether the source bears its own assertion, never about whether the claim is
true; `source_read` and `target_read` are honest, which you can check by
opening the source; a passage the mechanical check could not find in the
source is acknowledged in the reasoning rather than quoted as if verified;
the summary is in the graph's voice, refers to sources by what they say,
and contains no identifier, relation name, or score; `material` is a
judgment about the reader, not a function of how many rows were written;
and the assessment's reasoning says how the structure of the support
entered the verdict without any status having moved on the map's shape
alone. Use `provenance_get_map` and `provenance_read_source` to check any
of this. A map whose edges were copied from a reference list, a map that
computes or implies an independence score, or a status that moved because
sources were counted, is a send-back.

## For the Curator

A merge moves the loser's instances onto the survivor, and their readings
and edges move with them, since each is a property of the assertion rather
than of the claim; a stance flip changes nothing about them. The survivor's
source map, though, was written about a claim with fewer instances and is
stale from the moment of the merge: expect the Steward to refresh it on the
change you notify, and say in the notification that the map needs
rewriting. A split leaves each claim's readings and edges with the
instances that went to it, and the map with the original. Relations between
two documents are claim-independent and survive every merge and split.

## For the Extractor

## Standards for judging

An assessment on a mapped claim is good when its reasoning says what the
support rests on, names which sources it read closely and what they showed,
treats an overstating source as weaker than its wording, treats one voice
appearing several times as one voice, and explains how that entered the
verdict; and when the map's summary is plain prose a reader could act on.
An assessment that cites a count of supporting instances as if each were
independent, or that discounts a well-tested primary result because much
rests on it, without reading it, is not good. A map is good when its rows
are judgments from reading and its summary is honest about what was not
read.

Grantmaker · 1 section · 2,999 characters

# Method skill: Provenance (version 1)

This skill says how the constitution and your role apply to one kind of work. It never outranks either.

## For every administrator

Contribution Reviewer · 1 section · 2,999 characters

# Method skill: Provenance (version 1)

This skill says how the constitution and your role apply to one kind of work. It never outranks either.

## For every administrator

Dispute Arbitrator · 1 section · 2,999 characters

# Method skill: Provenance (version 1)

This skill says how the constitution and your role apply to one kind of work. It never outranks either.

## For every administrator

Curator · 2 sections · 3,675 characters

# Method skill: Provenance (version 1)

This skill says how the constitution and your role apply to one kind of work. It never outranks either.

## For every administrator

## For the Curator

A merge moves the loser's instances onto the survivor, and their readings
and edges move with them, since each is a property of the assertion rather
than of the claim; a stance flip changes nothing about them. The survivor's
source map, though, was written about a claim with fewer instances and is
stale from the moment of the merge: expect the Steward to refresh it on the
change you notify, and say in the notification that the map needs
rewriting. A split leaves each claim's readings and edges with the
instances that went to it, and the map with the original. Relations between
two documents are claim-independent and survive every merge and split.

Matcher · no sections · 145 characters

# Method skill: Provenance (version 1)

This skill says how the constitution and your role apply to one kind of work. It never outranks either.

Extractor · 1 section · 574 characters

# Method skill: Provenance (version 1)

This skill says how the constitution and your role apply to one kind of work. It never outranks either.

## For the Extractor

The passage you record as `verbatim_text` is later checked mechanically
against the stored text of the source, and a passage the check cannot find
casts doubt on the instance. Record the passage exactly as it appears:
the source's own words, its punctuation, its spelling, no paraphrase, no
silent trimming inside a sentence. Where a claim is asserted across two
sentences, record both rather than a splice.

Tools

The tool definitions this skill adds to a run’s toolset, with their descriptions verbatim, exactly as the model receives them. Each joins the toolset of the roles listed, and only when the skill is active.

provenance_get_map · Claim Steward, Audit Agent

Everything recorded about what this claim's support rests on: each instance with its id, its source's id and whether the graph holds the source's text, any reading already recorded for it, and the documents it has been recorded as drawing on; the relations among the documents involved; and the current source map summary, if one has been written. Call it before any provenance work, and take instance and source ids from it. Nothing it returns is a score.

{
  "type": "object",
  "properties": {
    "claim_id": {
      "type": "string",
      "description": "The claim. Defaults to the claim of the current run; required only outside one."
    }
  },
  "required": []
}

provenance_read_source · Claim Steward, Audit Agent

Read a source's text, in windows. Returns the stored copy reduced to readable text; when the graph holds no copy, fetches the document through the guarded path ingestion uses and stores it, so the quote check and later reads see the same text you read. A URL not yet in the graph is recorded as a source first. Page with offset; each call returns at most max_chars (default 12,000). This is how you open a source and read it whole (constitution section 9) rather than trusting an excerpt.

{
  "type": "object",
  "properties": {
    "source_id": {
      "type": "string",
      "description": "A source id from provenance_get_map or get_claim_with_context."
    },
    "url": {
      "type": "string",
      "description": "Alternatively, the document's URL, for a document not yet in the graph."
    },
    "offset": {
      "type": "number",
      "description": "Character offset to start from. Defaults to 0."
    },
    "max_chars": {
      "type": "number",
      "description": "Characters to return, at most 20,000. Defaults to 12,000."
    }
  },
  "required": []
}

provenance_record_reading · Claim Steward

Record what you concluded from reading one instance's source against the claim: whether the source's own evidence bears the assertion it makes, what the source is using the claim for, a reader-facing note, and whether a later Steward should open this source closely and why. One current reading per instance; calling again replaces it. Whether the instance's recorded passage is actually present in the source's stored text is checked mechanically and returned; you never supply it. Record a reading only for a source you have actually read (source_read true) or, where you could not open it, say so with source_read false and support 'unclear'.

{
  "type": "object",
  "properties": {
    "instance_id": {
      "type": "string",
      "description": "The instance, from provenance_get_map."
    },
    "support": {
      "type": "string",
      "enum": [
        "supports",
        "overstates",
        "understates",
        "asserts_without_evidence",
        "contradicts_own_evidence",
        "unclear"
      ],
      "description": "Whether the source's OWN evidence and argument bear the claim as this source states it. This is not whether the claim is true: that is your judgment as Steward, made elsewhere. 'supports': the source shows what it asserts. 'overstates': the assertion outruns the source's evidence. 'understates': the source's evidence would bear a stronger statement. 'asserts_without_evidence': stated as given, with no supporting evidence offered; ordinary and not itself a fault, but worth recording. 'contradicts_own_evidence': the source's own material cuts against its assertion. 'unclear': you could not tell from what you read."
    },
    "deployment": {
      "type": "string",
      "description": "What this source is using the claim for: the context of deployment, in a sentence. 'Cited in passing to motivate a policy proposal it does not otherwise argue for.'"
    },
    "note": {
      "type": "string",
      "description": "A note on this appearance in the graph's voice: plain, third person, no identifiers, no tool or edge names, no em-dashes. Shown to readers beside the quotation when present."
    },
    "worth_reading": {
      "type": "boolean",
      "description": "Whether a Steward assessing this claim should open this source and read it closely."
    },
    "worth_reading_reason": {
      "type": "string",
      "description": "Required when worth_reading is true: what a close reading would settle."
    },
    "source_read": {
      "type": "boolean",
      "description": "True only if you read the source's text (with provenance_read_source or by opening it), not merely its excerpt and metadata."
    }
  },
  "required": [
    "instance_id",
    "support",
    "source_read"
  ]
}

provenance_record_edge · Claim Steward

Record that one source's assertion of this claim draws on another document: the load-bearing edge of the source map. The edge runs from an instance (the asserting side) to a source (the document drawn upon), which need not assert the claim itself; where it does, the link to that instance is found for you. Name the target by source_id when it is already in the graph, otherwise by to_source_url. Every edge is a judgment from reading: evidence must hold the located passage in the asserting source where it draws on the target, quoted or unambiguously located. Never copy an edge out of a bibliography or a citation index you did not read against the claim. Recording the same (instance, source, relation) again replaces the earlier record.

{
  "type": "object",
  "properties": {
    "from_instance_id": {
      "type": "string",
      "description": "The asserting instance, from provenance_get_map."
    },
    "to_source_id": {
      "type": "string",
      "description": "The document drawn upon, when already in the graph."
    },
    "to_source_url": {
      "type": "string",
      "description": "Otherwise, the document's URL; it is recorded as a source."
    },
    "to_source_title": {
      "type": "string",
      "description": "With to_source_url: the document's title."
    },
    "relation_type": {
      "type": "string",
      "enum": [
        "repeats",
        "derives_from",
        "reanalyzes",
        "republishes",
        "cites_as_evidence",
        "responds_to"
      ],
      "description": "What this source's assertion of the claim takes from the upstream document. Decide by asking what would be lost if the upstream document did not exist. 'repeats': it takes the assertion itself and restates it, adding no evidence of its own. 'derives_from': it takes a finding or result and states the claim on that basis, as a news write-up states what a study found. 'reanalyzes': it takes the underlying data or results and does new work on them. 'republishes': the whole document is the upstream one, syndicated or mirrored. 'cites_as_evidence': it offers the upstream document in support without depending on it for the assertion's content. 'responds_to': it engages with, qualifies, or rebuts the upstream document's assertion of this claim."
    },
    "fidelity": {
      "type": "string",
      "enum": [
        "faithful",
        "strengthened",
        "weakened",
        "distorted",
        "misattributed",
        "unclear"
      ],
      "description": "How the claim fared crossing this edge. 'faithful': the assertion says what the upstream document supports. 'strengthened': stated more confidently, more generally, or with qualifications dropped; the commonest failure and the one worth catching. 'weakened': hedged beyond what the upstream document warrants. 'distorted': not what the upstream document supports, through misreading rather than invention. 'misattributed': the upstream document is credited with something it does not say, or the wrong party is credited. 'unclear': you could not establish this from what you read; prefer it to a guess."
    },
    "evidence": {
      "type": "string",
      "description": "The located passage in the asserting source where it draws on the target: a quotation, or an unambiguous locator such as a section and sentence. Required."
    },
    "reasoning": {
      "type": "string",
      "description": "Why this is the relation and the fidelity you recorded."
    },
    "confidence": {
      "type": "number",
      "description": "0 to 1: how sure you are that the dependency is real and is of this kind. Defaults to 0.5."
    },
    "target_read": {
      "type": "boolean",
      "description": "True only if you opened the target document. False when you judged the edge from the asserting source's own description of what it relies on; the fact is recorded beside the edge."
    }
  },
  "required": [
    "from_instance_id",
    "relation_type",
    "fidelity",
    "evidence",
    "reasoning",
    "target_read"
  ]
}

provenance_record_source_relationship · Claim Steward

Record a relation between two documents that holds whatever claim is being traced: shared authorship, republication, or one being a version of the other. These are what let a map say that five sources are three voices. Both sources must already be in the graph; take their ids from provenance_get_map. Recording the same pair and relation again replaces the earlier record.

{
  "type": "object",
  "properties": {
    "parent_source_id": {
      "type": "string",
      "description": "The first document. For version_of, the later one; for republishes, the original."
    },
    "child_source_id": {
      "type": "string",
      "description": "The second document."
    },
    "relation_type": {
      "type": "string",
      "enum": [
        "shares_authorship",
        "republishes",
        "version_of"
      ],
      "description": "A relation between two documents that holds regardless of which claim is being traced. 'shares_authorship': an author in common; symmetric. 'republishes': the child document is the parent's text, syndicated or mirrored. 'version_of': the same work at a different stage (preprint and published article, draft and final); record the LATER document as the parent."
    },
    "reasoning": {
      "type": "string",
      "description": "How you know: the byline, the masthead, the preprint notice."
    },
    "confidence": {
      "type": "number",
      "description": "0 to 1. Defaults to 0.5."
    }
  },
  "required": [
    "parent_source_id",
    "child_source_id",
    "relation_type",
    "reasoning"
  ]
}

provenance_write_map · Claim Steward

Write the claim's source map: the reader-facing account of what the support rests on, and whether that structure is material enough to show. One map per claim, replaced in place; the readings and edges remain the durable record beneath it. The summary is in the graph's voice: plain third-person prose that refers to sources by what they say, with no identifiers, no relation names, no counts presented as scores, and no em-dashes. Set material true only when the structure bears on how a reader should weigh the evidence (much of the support traces to one reanalysis; the strongest-sounding source overstates its own evidence; several outlets are one syndicated story). Most claims' provenance is unremarkable, and for them write a short summary and set material false.

{
  "type": "object",
  "properties": {
    "summary": {
      "type": "string",
      "description": "The account, one to four sentences in the graph's voice."
    },
    "material": {
      "type": "boolean",
      "description": "Whether to show the summary on the claim page."
    }
  },
  "required": [
    "summary",
    "material"
  ]
}