13. Improvement loop

L4 gets better from what it sent and from what it held back. Humans accept every production pin. Nothing promotes itself.

Month-12 quality is a versioned playbook and prompt set — not silent weight post-training on raw closures. Improvement runs offline. Opus 5.5 + GPT-5.6 Sol never sit on the plant request path.

1. Decision summary#

2. Why#

A system that only learns from cards it sent drifts cautious: gates hide their own mistakes. The opportunity ledgerStore of every blocked candidate with gate id and later outcome if known is the honest half of improvement (ADR-025).

ACE-style bullets stay inspectable. A plant owner can read “helpful 12 / harmful 3” on a playbook line. Opaque weight updates on messy closures teach false lessons (ineligible closes, selection bias, detector epoch shifts).

HITL is the product promise: recommend and record; humans decide what enters production registries.

3. Inputs#

Ineligible closes set “action worked” to null. Closure rate alone never ranks patterns.

4. Offline council#

flowchart TB
    %% house-style: l4-improvement-loop
    subgraph in["Inputs"]
        direction LR
        case["Case library outcomes"]
        ol["Opportunity ledger"]
        exp["Exploration outcomes"]
        seam["Seam decision records"]
    end
    off["Offline council<br/>Opus 5.5 + GPT-5.6 Sol"]
    prop["Proposals: bullets,<br/>prompts, soft thresholds"]
    replay{{"Replay holdouts + pass^k"}}
    shadow["Shadow on plant"]
    own(["↩ Named owner accept"])
    lock["Release lockfile pin"]
    case --> off
    ol --> off
    exp --> off
    seam --> off
    off --> prop --> replay --> shadow --> own --> lock

    classDef govc fill:#fff4d6,stroke:#c99a2e,color:#000
    classDef agentc fill:#e8f0ff,stroke:#5b7bd5,color:#000
    classDef loopc fill:#eef7ee,stroke:#4f9a4f,color:#000
    class replay govc
    class own agentc
    class off agentc

Where the two models agree on a process label, it stands as a candidate. Where they disagree, a human decides. Council judgement on blocked items may prioritise backlog review; it is not itself a training label for “should have emitted.”

5. ACE-style bullet playbooks#

Procedural memory holds versioned playbook bullets per domain / family / pattern scope.

Generator → reflector → curator emits add / update / remove. Deterministic code merges — no monolithic rewrite (brevity bias / context collapse). Counters never auto-delete a bullet; low helpful / high harmful triggers a proposal to retire, accepted by the owner.

Directives and hard stops are not editable by this loop.

6. Generator → reflector → curator (one cycle)#

1. Label — council labels a sample of traces and blocked ledger items (12-trace-and-eval.md).
2. Generate — ACE bullets · GEPA-style prompt candidates · registry edits (thresholds, soft gates, patterns, workflows) · stage-graph proposals · seam threshold changes · Jev swap proposals when records are strong enough.
3. Reflect — other model attacks nuisance rate, hard-gate adjacency, and “this only looks good on sent cards.”
4. Curate — drop illegal deltas (hard-gate loosens, write tools, money invention).
5. Replay — holdouts proposers never saw; pass^k and suite gates.
6. Owner pack — dissent visible; tech lead for global; plant owner for plant-scoped.
7. Shadow → canary → pin — unpin rolls back. In-flight runs finish under the lockfile they started; new runs take the new pin.

Proposer and red-team swap every cycle.

7. GEPA-style prompt evolution#

Prompts are registry entries under the lockfile.

If a candidate improves pass^k but raises nuisance or grounding violations, it fails curation.

8. Opportunity ledger → improvement (ADR-025)#

Soft-gate scorecards (miss rate, block precision, foregone Modeled effect, nuisance cost of loosening) are first-class generator fuel.

Hard gates are outside this loop. A pattern of hard-gate blocks that looks wrong is a data problem for the constraint owner — not a threshold knob.

Exploration closures (exploration=true) are the unbiased sample. Prefer them when proposing soft-gate moves. Detail: 22-missed-opportunities.md.

9. Seams → Jev#

Seam decision records are the training and comparison set. A classifier / Jev candidate must beat the logged LLM seam under replay before replacing it (11-models-and-seams.md).

10. Forbidden#

11. Named ownership#

12. Rejected alternatives#

13. Evidence that would change this#

• ACE counters too sparse after N weeks → add exploration sampling or tighten label cadence; do not jump to silent fine-tunes.
• Small verifier on closed seams beats LLM under replay → allow weight/Jev track for that seam only.
• Owner pack fatigue → raise auto-filter bar; do not remove the human pin.

14. v1 slice vs later#