Personas: phase1 on blackholes
3 simulated users — readers, contributors, programmatic clients and an adversarial minority — each an LLM agent playing a manifest entry, reading the graph and submitting through the same path a user takes; then the real review, escalation and arbitration. The findings they filed are triaged for a human first.
Fingerprint: what built each arm
phase1 on blackholes
epoch 2026-09-domain-skills commit e9189e8 profile none: development defaults, not production steward z-ai/glm-5.3-flash persona z-ai/glm-5.3-flash contribution_reviewer z-ai/glm-5.3-flash dispute_arbitrator z-ai/glm-5.3-flash models from run caps not recorded scored 2026-09-18 judge cost none
Persona run
Steps · 21, gists; open one for the verbatim record
No change to the graph.
Past events · 0
Legend
Contributor (simulated) Contribution Reviewer ↗ Extractor ↗ Matcher ↗ Claim Steward ↗ System
verified supported inconclusive doubtful contradicted unknown unassessed (hatched: pending, not a verdict) merged away touched by the current event
Edges: requires grey, supports green, contradicts red dashed, assumes amber dotted. Gists in lists are derived by the exporter; anything marked verbatim is the model’s own text. Attribution “run-window” means a change was credited to a run by time and claim, not by a tool call that names it.