Governance
How the research program learns to govern itself: incidents, controls, lessons, and what each agent role is shown.
Public repositories
| Repository | Role | Verified | Last checked |
|---|---|---|---|
yoheinakajima/activegraph | The ActiveGraph runtime this Atlas runs on. | verified | 2026-08-03 |
yoheinakajima/activegraph-lab | The always-on lab/assistant runtime; optional future Portfolio Lab. | verified | 2026-08-03 |
yoheinakajima/activegraph-packs | Shared pack library (research, codebase, team-ops, governed evolution). | verified | 2026-08-03 |
yoheinakajima/distributed-discovery | Distributed Discovery program, theory set, lab pilot, and TreasureBench artifacts. | verified | 2026-08-03 |
yoheinakajima/ouroboros | Ouroboros program; public home of the self-modification audit. | verified | 2026-08-03 |
yoheinakajima/synthetic-players | Synthetic Players program and persona-panel paper artifacts. | verified | 2026-08-03 |
Incidents
resolved Vacuous assertion in the site self-containment test (low)
tests/test_atlas.py asserted site self-containment with a trailing `or True`, making the check unfalsifiable while appearing to pass. Found by the owner's external review of commit c23e926 — the Atlas's own CI did not catch it. Kept on the record deliberately: a dogfood cycle that reveals and repairs a real weakness is stronger evidence than a frictionless demo.
tests/test_atlas.py asserted site self-containment with a trailing `or True`, making the check unfalsifiable while appearing to pass. Found by the owner's external review of commit c23e926 — the Atlas's own CI did not catch it. Kept on the record deliberately: a dogfood cycle that reveals and repairs a real weakness is stronger evidence than a frictionless demo.
Controls
Advisory findings stay advisory until validated hard
Structural audits are valuable, but false-positive-prone heuristics must not become mandatory CI. Hard failures protect contracts (references, identifiers, cycles, sync); inference produces reviewable clues. Adopted from the Distributed Discovery ActiveGraph pilot and built into this Atlas's CI design.
Structural audits are valuable, but false-positive-prone heuristics must not become mandatory CI. Hard failures protect contracts (references, identifiers, cycles, sync); inference produces reviewable clues. Adopted from the Distributed Discovery ActiveGraph pilot and built into this Atlas's CI design.
Ruff lint tier in CI (vacuous-assertion class) hard
CI runs ruff with the SIM and B rule families; SIM222 flags `expr or True`-style always-true conditions, the exact class of the vacuous-assertion incident. Adopted alongside replacing the vacuous test with a real link resolver.
CI runs ruff with the SIM and B rule families; SIM222 flags `expr or True`-style always-true conditions, the exact class of the vacuous-assertion incident. Adopted alongside replacing the vacuous test with a real link resolver.
Preregister success criteria before empirical campaigns advisory
Claim-grade empirical work declares its uplift/success criteria before execution. The Ouroboros substrate comparison ran against a preregistered uplift criterion (and reported the null); the TreasureBench campaign is gated on explicit authorization with a frozen instrument for the same reason.
Claim-grade empirical work declares its uplift/success criteria before execution. The Ouroboros substrate comparison ran against a preregistered uplift criterion (and reported the null); the TreasureBench campaign is gated on explicit authorization with a frozen instrument for the same reason.
Method lessons
Keep false-positive-prone audits out of mandatory CI
The Distributed Discovery ActiveGraph pilot found real value in structural auditing while correctly keeping a false-positive-prone heuristic advisory rather than promoting it to a merge blocker. The Atlas encodes this as the hard/advisory audit split.
clue: actionability=high, decay=low, granularity=mechanism, representation=structured-record, scope=cross-project, source_type=direct-observation, visibility=public
The Distributed Discovery ActiveGraph pilot found real value in structural auditing while correctly keeping a false-positive-prone heuristic advisory rather than promoting it to a merge blocker. The Atlas encodes this as the hard/advisory audit split.
clue: actionability=high, decay=low, granularity=mechanism, representation=structured-record, scope=cross-project, source_type=direct-observation, visibility=public
Retaining a change is not expressing it
Retention, expression, behavioral mediation, task improvement, and recursive improvement are five different claims. A persisted self-modification establishes only the first; each further claim needs its own evidence. Any self-improvement story in this portfolio is audited against this separation.
clue: actionability=high, decay=low, granularity=mechanism, reliability=preregistered-null, representation=structured-record, scope=cross-project, source_type=direct-observation, visibility=public
Retention, expression, behavioral mediation, task improvement, and recursive improvement are five different claims. A persisted self-modification establishes only the first; each further claim needs its own evidence. Any self-improvement story in this portfolio is audited against this separation.
clue: actionability=high, decay=low, granularity=mechanism, reliability=preregistered-null, representation=structured-record, scope=cross-project, source_type=direct-observation, visibility=public
Declared role exposure
This records the slice each role's generated context DECLARES — not what an agent actually loaded, read, or used. Behavioral exposure (invocation receipts, read tracing) is a planned, separate measurement; see plans/adr-backlog.md.
| Role | Sees | Events | Dominant source | Share |
|---|---|---|---|---|
idea-curator | ideas, their proposed relations, and the inbox | 0 | — | 0% |
paper-editor | active works' axes and their status/identifier events | 25 | owner-statement | 61% |
portfolio-lab | the full public projection (pinned read, propose-only) | 37 | owner-statement | 72% |
portfolio-manager | all works (summary), decisions, notes, forecasts | 4 | owner-statement | 100% |
scientific-auditor | every status assertion with its full source list | 31 | owner-statement | 68% |