Define ADR-0014 corrected bounded pretrained meta-trainer cycle (reduced steps, no training)

[Rook-CodeVolt]

Outcome

Prepare one corrected bounded pretrained meta-trainer training cycle (ADR-0014) that responds to ADR-0013's completed, evidence-backed negative result with a single, documented hyperparameter correction. This issue authorizes design, dry validation, and independent review only. Training must not start until the exact candidate ADR and live-execution plan are approved and merged.

Context

ADR-0013's first real bounded run (commit 13b316e6bea25d141ecf1c89284e40e68a1b3ff9, run_id adr0013-metatrainer-sft-20260920) completed end-to-end cleanly under real gated authorization but produced an honest negative result: the post-training candidate scored worse than the immutable baseline on all three evaluation suites:

• meta_trainer (target task, 20 items): 0.0% -> 0.0% (no improvement)
• capability_retention (arithmetic, 10 items): 30.0% -> 0.0% (regressed)
• safety (15 items): 46.7% -> 33.3% (regressed)

Raw held-out outputs show classic full-parameter-SFT overfitting/degeneration on the 40-example corpus (repetitive token loops such as "46-46-46-46..." and "PEPFET, a variant of PEPFET..."). This matches ADR-0013's own "Stop and reject this candidate" trigger. ADR-0013's own text is explicit that a continuation "requires a new issue/ADR with a fresh bounded configuration" and "does not authorize an automatic retry, LR sweep, extra epoch, or second seed" — this issue and its companion ADR-0014 are that new proposal.

Scope

• Reuse ADR-0013's immutable model identity, corpus (pilot-metatrainer-v2), evaluator, evaluation suites (meta_trainer/capability_retention/safety), contamination machinery, and resource-budget shape unchanged.
• Change MAX_STEPS from 120 (3 epochs at batch 1 over 40 examples) to 40 (1 epoch), with documented rationale citing ADR-0013's regression numbers and the observed overfitting/degeneration pattern.
• Also lower LEARNING_RATE from 1e-5 to 5e-6 (halved). ADR-0013's 1e-5 was not itself excessive for full-SFT, but with the step count cut to a single epoch, the remaining corrective lever is per-step update size: full-parameter SFT (not LoRA) updates every parameter every step, so the epoch reduction and the LR reduction are complementary defense-in-depth against re-hitting the same degeneration pattern, not redundant. Halving (rather than a steeper cut) keeps the run a live, comparable second measurement point in the same experimental family as ADR-0013, not a barely-perturbing no-op. Documented in full in the ADR-0014 configuration section.
• Do not touch dataset content, security/sandboxing/gate-verification code, or resource budgets.
• Use a fresh run id (adr0014-metatrainer-sft-20260922), a fresh approval-signature namespace (codevolt-adr0014, distinct from ADR-0013's codevolt-adr0013) and fresh reviewed host paths, so no artifact, signed approval, or scratch directory can be confused between the two cycles.
• Run tests, lint, document/link checks, and non-training dry validation only.

Acceptance criteria

• ADR-0014 follows repository style and cites ADR-0013's exact regression numbers as its evidence basis.
• Model repo, immutable revision, license, and content hashes are unchanged from and identical to ADR-0013's.
• Trainer/evaluator paths and exact dependency pins are unchanged from ADR-0013's; no unsupported adapter behavior is newly claimed.
• Dataset, manifest, registry, and supporting-file hashes are unchanged and the 40/20 split is preserved; the corpus is read from its existing location, not duplicated.
• Baseline and post-training use the same held-out data, renderer/template identity, decoding settings, and rubric as ADR-0013.
• Pass/continue/stop thresholds and all mandatory stop conditions are unchanged from ADR-0013.
• Resource budgets and allowed network destinations are unchanged; live training remains offline after prerequisite fetches.
• No secrets, raw weights, publication, deployment, or promotion are authorized.
• Reproducible dry-run and future execution commands are documented; this issue's implementation does not execute training.
• The exact candidate receives independent security review before any live execution.
• CI (lint + full test suite) passes on the PR introducing this ADR and its runner.

Authorization boundary

The owner has conditionally authorized drafting this corrected cycle after ADR-0013's negative result, per comment on kanban task t_b366da7a: "Draft ADR-0014: a new, freshly-reviewed bounded cycle proposing reduced MAX_STEPS/epochs (and LR if justified) as a corrected configuration, citing ADR-0013 stop-and-reject result." That authorization is limited to drafting, dry validation, and independent review for this exact configuration. It does not authorize training before review/merge, additional runs, an unattended loop, automatic promotion, publication, deployment, or any increase in authority.

Edit: the LR-unchanged line in the original scope has been corrected above to match the implemented and documented decision (LR halved to 5e-6), which is within the "and LR if justified" discretion the owner's comment granted. See ADR-0014 for the full rationale.

[Rook-CodeVolt]

PR opened: #86

[Rook-CodeVolt]

Independently re-verified and closing.

PR #86 (merged to main, merge commit 7f74971, head 48db059): 6 files changed (CHANGELOG.md, docs/ROADMAP.md, docs/decisions/0014-corrected-bounded-metatrainer-cycle.md, examples/pilot-metatrainer-v2/run_bounded_cycle_adr0014.py, examples/pilot-metatrainer-v2/approval_allowed_signers_adr0014, tests/test_bounded_metatrainer_cycle_adr0014.py). Pulled the merged diff myself and diffed run_bounded_cycle_adr0014.py line-by-line against ADR-0013's run_bounded_cycle.py: the only substantive changes are MAX_STEPS 120->40, LEARNING_RATE 1e-5->5e-6, SAVE_STEPS 40->20, SEED, RUN_ID (adr0014-metatrainer-sft-20260922), APPROVAL_NAMESPACE (codevolt-adr0014), and re-scoped path/env identifier strings -- no change to host-containment, gate/signature verification, resource-budget, or contamination-check logic. approval_allowed_signers_adr0014 has the same 3 signer keys (rook-owner, maya-security, maya-dataset-rights) as ADR-0013, confirmed via repo content fetch. No dataset files touched anywhere in the diff.

Maya-CodeVolt's review (PRR_kwDOUT4pss8AAAABOndGbw, submitted 2026-09-22T08:43:32Z, state APPROVED, reviewed commit 48db059) matches this diff exactly -- same file list, same hunk-level findings, same signer verification.

--execute path confirmed still fail-closed and never invoked: the runner requires --execute plus --review-gate plus an exact approved scratch root and gate path (run_bounded_cycle_adr0014.py lines ~908-937); no such call exists in CI or history, and no run artifacts/scratch dirs for adr0014-metatrainer-sft-20260922 exist anywhere in the repo tree.

CI: workflow run 35705827258 on head 48db059 -- test(3.9)/test(3.11)/test(3.13) all conclusion=success.

All acceptance criteria satisfied: ADR-0014 present and cites ADR-0013's exact regression numbers; model/dataset hashes and paths unchanged from ADR-0013 (only identifiers re-scoped); fresh run_id and approval namespace; no dataset/security/gate code touched; CI green; independent security review (Maya) completed and matches the actual diff; no training executed.

Live execution of ADR-0014 remains unauthorized. Before any --execute call against run_bounded_cycle_adr0014.py, 3 fresh signed approval documents (security, dataset-rights, owner) bound to head SHA 48db059 / run_id adr0014-metatrainer-sft-20260922 / namespace codevolt-adr0014 are required, per the existing three-gate pattern. That is a separate, future gated step outside this issue's scope (drafting + dry validation + review only).