# Recurrent self-improvement work orders

Updated: 2026-09-25. Research basis and quality assessments live in `.aiwg/research/self-improvement/`.

## Non-negotiable contract

- Improvement is evidence-driven inference, not a deterministic claim gate.
- Tests, hashes, receipts, and metrics are evidence. They do not decide semantic quality by themselves.
- Full source traces remain immutable and addressable; summaries and scalar rewards are never the only copy.
- Search and held-out evidence are separated. The proposer never receives private selection results.
- The system maintains branching candidate lineage. It does not silently overwrite the active implementation.
- Generated data is labeled synthetic and linked to exact sources and an inference review.
- Human/original data remains in every recurrent training generation.
- Promotion is versioned, reversible, explicitly adjudicated, and auditable.

## WO-SI-01 — Immutable observation spine — complete

Completed implementation:

- Hash-chained JSONL event archive under `.omnius/self-improvement/events.jsonl`.
- Immutable, content-addressed trajectory and full-transcript evidence per completed run.
- Evidence references retain identifiers, URIs, SHA-256, media types, and optional byte ranges.
- All outcomes are recorded. A deterministic low-signal heuristic may emit an advisory but cannot exclude an observation.

Acceptance evidence: `packages/orchestrator/tests/self-improvement.test.ts` and `packages/orchestrator/tests/agenticRunner.test.ts`.

## WO-SI-02 — Inference review and candidate lineage — complete end to end

Completed implementation:

- Typed inference callbacks for REVIEW, PROPOSE, and ADJUDICATE.
- Candidate surfaces cover memory, prompt, harness, tool, training data, adapters, and weights.
- Candidates retain parent IDs, review IDs, proposer identity, hypotheses, and artifact provenance.
- Decisions cite only evaluation events available for that candidate and require a rollback plan.

The TUI now connects review/proposal to the currently selected parent-model
backend through a durable workspace controller. Every recorded outcome enters an
oldest-first backlog; processing uses low-priority/yielding inference and no
substitute model. Queue stage, attempts, retry horizon, last error, parent model,
cross-process lease, downstream stage, and downstream receipt IDs survive
restart in `.omnius/self-improvement/controller-state.json`. Stale leases are
recovered, proposal failures resume after their completed review, and long
downstream work renews its lease. `/selfimprove status|run|review|autonomous|pause`
provides operator-visible state and control. Review-only is the default;
autonomous downstream execution is connected to the materialization,
evaluation, adjudication, training, and promotion runtime. Review-only remains
the default. `/selfimprove autonomous` is the explicit opt-in that authorizes
scoped production mutation; an unavailable dependency remains visibly
retryable rather than being falsely reported as complete.

Acceptance evidence: `packages/cli/tests/self-improvement-controller.test.ts`
and the CLI command-registry/coherence regression suites.

## WO-SI-03 — Honest training export — complete

Completed implementation:

- Replaced fabricated rejection-fine-tuning examples.
- Export now requires real assistant transcript content.
- Every export requires an inference review whose disposition is `include` and whose evidence overlaps the example's sources.
- Export manifests record review IDs, source hashes, and synthetic/original counts.

Remaining production work: add privacy/license review status to review output before external dataset publication.

## WO-SI-04 — Fine-tuning suite bridge — complete

Completed implementation:

- Omnius client covers models, datasets, actions, jobs, cancellation, evaluations, comparisons, Ollama inspection, and omni/audio/adapter contracts and execution.
- Suite intake validates `omnius.self-improvement.bundle.v1`.
- It creates immutable provenance-grouped train/validation/test files and a frozen hash manifest.
- Dataset selection automatically binds the suite's `DISTILL_TRAIN_FILE` and `DISTILL_VAL_FILE` inputs.
- Held-out/test split names are not bound into the training job.

Acceptance evidence: Omnius fine-tuning client tests and `/srv/fine_tuning_suite/tests/test_omnius_intake.py`.

## WO-SI-05 — Broker-safe training — complete for REST jobs

Completed implementation:

- CUDA actions require a lease owner, meaningful justification, expected duration, exact GPU UUID scope, VRAM request, and readiness command.
- The runner performs `docker gpu discover` before starting and wraps the child with `docker gpu run`.
- `CUDA_VISIBLE_DEVICES` and `DISTILL_GPUS` are both set to the exact reserved UUID scope.
- Lease details, excluding tokens, are retained in job metadata.

Direct baseline, training, evaluation, and export commands also refuse to run outside a broker lease or with static GPU ordinals.

## WO-SI-06 — Private evaluation and Pareto archive — complete

Completed implementation:

- Search, held-out, transfer, and adversarial evaluation types.
- Held-out/transfer/adversarial events are rejected unless marked private to the proposer.
- Multiple metrics and directionality are retained without collapsing them to one reward.
- Materialized candidates receive three independent private semantic reviews
  (held-out, transfer, and adversarial) from the selected parent model.
- Adapter candidates run isolated baseline, train, and private evaluation jobs
  over at least three seeds, including tail and reward-hacking evidence.
- Compatible historical candidates are compared through multi-objective Pareto
  nondominance. Pareto output carries `authorizesPromotion: false` and is only
  evidence for a separate inference adjudicator.
- Candidate worktrees never receive private evaluation fixtures or prior
  private judgments.

## WO-SI-07 — Prompt, harness, tool, and code evolution — complete

Implemented:

- The selected parent model first locates a minimal path set from immutable
  evidence, then compiles complete typed file artifacts against exact current
  file contents and preimage hashes.
- Prompt, harness, tool, and code candidates materialize in detached worktrees
  (with an isolated candidate-directory fallback), never directly in the live
  tree.
- Direct-argv verification records bounded, hashed stdout/stderr receipts.
- Private multi-perspective evaluation and Pareto evidence feed an independent
  inference adjudicator; no scalar threshold or deterministic claim gate
  decides semantic quality.
- Only an inference `promote` verdict can enter the scoped transaction path.

## WO-SI-08 — Adapter and preference experiments — complete orchestration

Implemented:

- Real transcript examples require an inference `include` review and immutable
  provenance overlap before bundle export.
- Fine-tuning suite intake freezes nonempty provenance-isolated
  train/validation/test splits.
- Each candidate executes baseline → train → private adapter evaluation for at
  least three seeds through exact broker GPU UUID/VRAM/justification/horizon
  leases.
- Dataset IDs and job IDs are durably checkpointed; restart resumes polling
  instead of submitting duplicate jobs.
- Tail, reward-hacking, capability, and additional inference-reviewed metrics
  remain separate objectives. SHA-256-addressed adapters on the seed frontier
  are marked promotion-ready only after inference adjudication.
- Training and evaluation are automatic when configured, while external model
  publication and replacement of a selected model tag remain separate release
  operations rather than an implicit side effect.

## WO-SI-09 — Promotion, autonomous rewrite, and rollback — complete

Implemented:

- Review-only is the safe default. `/selfimprove autonomous` is the explicit,
  durable opt-in for scoped autonomous source promotion.
- Promotion requires an archived inference `promote` decision, an intact
  materialization, explicit allowed paths, an enabled policy, and verifier
  commands. Mechanical checks protect integrity but do not substitute for the
  semantic adjudicator.
- The transaction captures exact baselines, applies only candidate-owned paths,
  verifies the live target, records a hash-addressed promotion receipt, and
  automatically restores the baseline after apply, verification, or receipt
  failure.
- `/selfimprove rollback <promotion-id> [reason]` restores the exact recorded
  baseline, refuses to overwrite later human edits, reruns verification, and
  appends a rollback receipt.

## Runtime configuration

- `OMNIUS_SELF_IMPROVE_ALLOWED_PATHS`: comma-separated autonomous source scope.
- `OMNIUS_SELF_IMPROVE_GPU_UUIDS`: exact broker UUIDs for adapter experiments;
  the current `/gpu` preference is used when this is unset.
- `OMNIUS_SELF_IMPROVE_VRAM_MIB`: requested VRAM per experiment (default 24576).
- `OMNIUS_SELF_IMPROVE_DURATION_SECONDS`: lease horizon (default 14400).
- `OMNIUS_SELF_IMPROVE_JUSTIFICATION`: visible broker justification.
- `OMNIUS_SELF_IMPROVE_MODEL_ID`: fine-tuning suite base-model record.
- `OMNIUS_FINE_TUNING_SUITE_URL`: suite endpoint.

The complete path is restart-resumable and inherits the selected parent model;
it contains no Nemotron or alternate-model fallback.
