ADR-284: Bounded nightly SOTA research agent

Field	Value
Status	Accepted - implementation gated off by default
Date	2026-07-29
Builds on	ADR-283

Context

RuView needs a repeatable way to notice relevant state-of-the-art work and turn it into reviewable repository activity. A nightly model with simultaneous network, repository-write, policy-evolution, and execution authority would create an unacceptable prompt-injection and supply-chain boundary. It could also confuse generated confidence with scientific evidence or silently turn a research suggestion into production code.

Cognitum exposes an OpenAI-compatible completion service and a public application registry. The contributor harness already commits a Darwin genome and a signed Flywheel replay gate. Those components can support nightly research without granting unattended learning promotion.

Decision

Add a scheduled GitHub Actions workflow that runs daily at 03:17 UTC, remains disabled until a maintainer enables a repository variable, and supports a manual evidence-only dry run.

The live flow has seven jobs:

1. Collect bounded public Cognitum-registry and recent arXiv evidence.
2. Ask Cognitum cognitum-mid for one proposal that is locally validated against a strict schema.
3. Score proposal completeness using the frozen Darwin policy and verify an honest-null Flywheel replay.
4. Deduplicate or create one issue.
5. For a locally classified low-risk proposal only, ask Cognitum for a tiny declarative transform and bounded test vectors. Trusted repository templates turn that data into the prototype module, tests, JSON, and README.
6. In a job with no external secret or GitHub write token, revalidate every artifact, verify the Flywheel replay, and perform static and syntax checks without executing generated code.
7. In a job with no model credential, re-hash the validated artifacts, create a new branch, open one draft PR, link it to the issue, and explicitly dispatch the credential-free contributor-harness verifier.

The jobs exchange bounded JSON artifacts. Cognitum receipts retain the provider, endpoint, exact resolved tier/model, request ID, a recomputable routing attestation, and digest metadata. Credential-free validation rebuilds the deterministic request and verifies its digest. The raw-output digest is audit metadata only because raw model transcripts are not retained.

Security and evidence policy

Retrieved titles, abstracts, descriptions, and links are untrusted CLAIMED evidence. Source hosts, paths, media types, redirects, time, byte counts, records, and citations are validated. The fixed trusted prompt states that evidence has no instruction authority. Model output is parsed as one JSON object and locally reconstructs risk, citations, implementation disposition, and fingerprint.

Risk classification is deliberately conservative. Security, authentication, cryptography, workflow, dependency, release, deployment, production, firmware, hardware, network-server, native/Wasmtime plugin, HomeKit pairing, STT/TTS, and satellite-voice proposals are issue-only.

Autonomous implementation is restricted to new files beneath a fingerprinted examples/research-sota/nightly/ directory. The model cannot supply paths or source text. It selects only a schema-bounded scalar transform and matching test vectors; repository-owned templates deterministically emit exactly five files. It cannot edit existing files or add dependencies. File count, size, line count, paths, symlink ancestry, numeric bounds, operation schema, secret-shaped values, canonical template digests, and accuracy claims are checked. Emitted source receives syntax checking, but is not executed.

The deterministic score is named PROPOSAL_COMPLETENESS. It is explicitly not a novelty, scientific-quality, safety, or performance score.

Darwin and Flywheel boundary

Nightly automation reads the committed Darwin genome as frozen prompt policy. It never calls Darwin evolution or any Cognitum evolve, pod, guidance-mutation, brain-write, or promotion endpoint.

Flywheel evaluates the unchanged policy with the repository's honest-null fixture. The signed replay must verify, report zero verified improvements, and report no promotion. This canary proves only that the committed Flywheel gate stayed root-only, rejected its candidate, and did not promote under the frozen fixture. It does not evaluate the proposal. The no-learning/no-promotion boundary for the nightly run comes from the workflow's static authority split, closed commands, and artifact validation.

Credentials and publication

Scheduled enablement requires:

• repository secret COGNITUM_NIGHTLY_API_KEY, limited to completions:mid; and
• repository variable RUVIEW_NIGHTLY_SOTA_ENABLED=true.

Model jobs receive no write-capable GitHub token. GitHub mutation jobs receive no model key. Validation receives neither. The publish job has the additional actions:write permission solely to dispatch the read-only ruview-harness-flywheel.yml verifier with Darwin disabled, because a PR created by the workflow token may not trigger ordinary pull-request workflows.

The agent creates draft PRs only. It cannot approve, merge, release, promote a Darwin candidate, or update canonical shared-brain records. Before any branch write, it re-fetches the issue and repository rules. Publication requires the issue to remain open, bot-authored, correctly labelled, and fingerprint-bound; main must require at least one approving review and the Verify contributor harness job-name check. The publisher requires that exact check name and GitHub Actions integration ID from GitHub's effective-active-rules endpoint. GitHub hides ruleset bypass actors from read-only tokens, so the workflow is not given an administrative token to inspect them. Its safety does not depend on that metadata: the publisher can create only a non-default branch and draft PR and contains no merge, approval, or main-push path. Branch protection and maintainer review remain the authority boundary.

Consequences

RuView gains a low-volume research flywheel with durable evidence, stable deduplication, and inspectable failure artifacts. A compromised paper, registry record, or model can at worst propose bounded new example files that still require static gates and human review.

The tradeoff is intentionally limited autonomy: production ideas become issues, generated prototypes are not executed, and a missing credential, service outage, schema drift, or validation ambiguity stops the run rather than guessing. Maintainers must explicitly enable the schedule and permit Actions to create pull requests.