Repository navigation
Agent Roles
James Nikolas Putradi edited this page Nov 20, 2026
·
1 revision
There is no human law author. This page is the practical view of Appendix J; the appendix is normative.
| Role | Writes | Must not read | Scored on |
|---|---|---|---|
| Conductor | task graph, manifests, locks, merges | — | protocol integrity |
| LAA |
model.bend, LAWS.bend, witnesses, abs.yaml, props_locked.sv, manifest |
oracle, seeded bugs, PROOF_ref.bend, agent traces |
law discrimination (LA3-LA6) |
| RPA |
PROOF_ref.bend + effort log |
oracle, traces | provability within the effort cap |
| OAA | oracle TB, oracle SVA, golden RTL, seeded bugs, checklist | every LAA artifact | oracle recall on the bug corpus |
| PA | props_c2.sv |
the Bend laws, the oracle | parity within 5 pp |
| AA | LA-gate evidence, adversarial review | oracle plaintext | defects found |
| IA | the RTL and PROOF.bend under test |
every protected path | the arm's visible gates (and nothing else) |
- Separate Freebuff runs. Different session, different workspace, no shared conversation. Roles communicate through committed files and issue links only.
- Read-only mounts. The policy proxy mounts protected paths read-only and logs every read attempt.
- Different model families where the compute budget allows, recorded per artifact. A same-family LAA/OAA pair is refused by the provenance check unless declared.
- Different objectives. LAA is rewarded for discrimination, OAA for recall, AA for findings. No role is scored on another role's metric.
-
Hash locks. Once
locked, a law file changes only through an amendment. -
Evidence, not sidecars. The policy-proxy audit log is the proof the barrier held;
provenance.yamlis a claim about it.
A law author and an implementation agent from the same model family can agree on the same blind spot: the laws omit what that family tends to omit, the agent passes easily, and C3 looks strong while measuring little.
Four defences, all pre-registered:
- Mutation pressure (LA3, LA4): a blind-spot law set fails to kill plausible bugs.
- Independent oracle recall: OAA's kill rate is measured against the same corpus.
-
Adversarial auditor: AA proposes implementations that satisfy every law and violate
spec.md. - Ablation A7 + LAW-FAR: same-family versus cross-family FAR becomes a moderator, not noise.
If LAA and OAA read the spec differently:
- both readings are published in
ambiguities.yaml; - the reading stricter on the design becomes the law; the other becomes an oracle-only requirement;
- if the readings are behaviourally incompatible, the task is quarantined — never silently resolved.
Release approval, ethics, external correspondence, and adjudication of an unresolved blocker. A human may spot-check 3% of locked law sets for quality. No human writes law content.
Code Apache-2.0 · documents and task data CC BY 4.0 · this wiki records decisions, not the current state of the code — the design document and the repository are normative.