…erified (#1754)

* docs(spec): SPEC §84+§85 — P2-C/P2-E live findings, hyperparameter hypothesis CORROBORATED, P0-K closure live-verified

Two new spec sections + full P2-E evidence directory.

## §84 — P2-C dispatched; audit hypothesis FALSIFIED; P0-K surfaced

P2-C ran the audit-recommended multi-source corpus (49.6B tokens, 80×
§82's 1.24B) at the same hyperparameters as §82. Result: val_loss=4.91
@ ep20 (vs §82's 4.71) — IDENTICAL termination shape, +0.2 WORSE despite
80× more data. The Chinchilla-data-starvation hypothesis is FALSIFIED.

Debugging the §81-§83 5-PR cascade surfaced PMAT-690 P0-K: `apr convert`
(both apr_import and apr_convert paths) didn't stamp hf_architecture /
hf_model_type / embedded tokenizer. Five downstream consumer fixes had
been patching None values that read from the upstream gap. P0-K closes
the producer.

## §85 — P2-E live findings; hyperparameter hypothesis CORROBORATED

P2-E ran same qwen-v3 corpus at LR=1.5e-5 (-3.3× lower) + warmup=500
(5× longer). Result: val_loss=4.6227 @ ep49 — BELOW §82's 4.71 AND
P2-C's 4.91 floors. No early-stop; smooth monotonic descent across all
50 epochs. Hypothesis from §84 P2-E queue is CORROBORATED.

Training throughput: 15,460 tok/s pure (12,880 tok/s end-to-end with
checkpoint write) on RTX 4090, sm_89, cuBLAS TF32. This is the
canonical apr-cli CUDA training perf baseline for future dispatches.

§30 a-priori falsification lesson amendment: the audit's
pre-falsification of P2-A2 was correct at the original LR but wrong
as a general claim. Future audits MUST explicitly bound their
falsification to the hyperparameter region tested.

## P0-K live-verification

Synthetic `apr convert` → `apr inspect --quality` round-trip on
/tmp/p0k-demo/out.apr (Qwen2 config.json + tiny safetensors fixture)
produces:
- metadata.hf_architecture = "Qwen2ForCausalLM" (was null pre-P0-K)
- metadata.hf_model_type = "qwen2" (was null pre-P0-K)
- quality.score = 60/100, hf_identity sub-score = 20/20

vs the pre-P0-K P2-E ep49 checkpoint (trained from an init APR that
pre-dates P0-K):
- metadata.hf_architecture = null
- quality.score = 40/100, hf_identity sub-score = 0/20

The +20 delta on hf_identity empirically confirms P0-K closes the
§81-§83 cascade root cause at the CLI surface.

## Ship % impact

MODEL-2 stays at 79%. val_loss 4.62 > 3.0 ship gate. Marginal-gain
decay analysis says more-of-the-same plateaus ~4.4. Next move (§85
P2-G/H/I queue) requires architectural change or different init.

## Refs

- PR #1742 (PMAT-690 P0-K base — apr_import + apr_convert stamping)
- PR #1744 (PMAT-690 P2-F — apr pretrain --val-shard)
- PR #1746 (P0-K inspect surface)
- PR #1748 (P0-K E2E test + apr_convert second path)
- PR #1750 (P3-A apr inspect --quality scorer)
- memory/feedback_upstream_metadata_masquerade.md (lesson #33)
- memory/feedback_parallel_session_worktree_isolation.md (lesson #34)
- memory/feedback_cargo_feature_cache_staleness.md (lesson #35)
- evidence/p2c-2026-05-17/findings.md (P2-C trajectory + root cause)
- evidence/p2e-2026-05-17/findings.md (P2-E corroboration + perf baseline)

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

* docs(spec): SPEC §86 — apr pretrain --init silently fails on arch-mismatched APRs; PR #1757 ships in-place stamp salvage

P2-G v1 dispatch surfaced a SECOND symptom of the §81-§84 cascade root
cause: pre-P0-K APR checkpoints (architecture="LlamaForCausalLM" P0-H
fallback + Qwen2-tensor shape) are silently non-resumable via
`apr pretrain --init`. The init eval at step 0 produced val_loss=8.60
instead of P2-E ep49's recorded 4.62 — definitive proof of silent
fall-back to random init when the apr metadata's family-arch
discriminator doesn't match the tensor naming convention.

## What §86 covers

1. Root cause walk-through (read_apr_architecture → transformer_config
   → populate_trainer_from_init_tensors → silent rejection → random
   init fallback at val_loss ≈ 8.60).
2. Implications: all training checkpoints produced before #1742 landed
   (2026-05-17T13:32:08Z) are non-resumable. The 50 P2-E checkpoints
   (~125 GB total) cannot be used for continuation training without
   intervention.
3. Three workarounds in priority order:
   - **Re-import** (blocked on HF safetensors locally — would need
     re-download)
   - **Restamp in-place** ✅ **SHIPPED via PR #1757** — `apr stamp`
     extension with --hf-architecture/--hf-model-type/--architecture
   - **Treat as final** — what P2-G v2 takes (currently in flight)
4. Operator recipe for the §86 salvage (3-line shell example).
5. Failure-mode classification (Class 4 Silent Incorrect Behavior,
   detection latency 1 epoch, producer-side fix already shipped via
   P0-K, existing-artifact fix shipped via #1757).
6. Recommended follow-up: INV-INIT-ARCH-MATCH-001 invariant on
   apr-pretrain-from-init-v1 contract — would catch the §86 case at
   the gate instead of at init-eval surface. Defer to follow-up PR.

## Stacked on PR #1754 (SPEC §85)

Base: `feat/spec-85-p2e-findings`. The §86 amendment depends on §85
context (the P2-E run that surfaced §86). Will auto-rebase to main
after #1754 lands.

## Refs

- PR #1742 (PMAT-690 P0-K base — apr_import + apr_convert stamping)
- PR #1750 (P3-A `apr inspect --quality` scorer — the diagnostic
  that surfaces §86 quality=40 pre-stamp, 60 post-stamp)
- PR #1754 (SPEC §85 P2-E findings — the run that surfaced §86)
- PR #1757 (apr stamp HF identity extension — workaround #2 above)
- evidence/p2g-2026-05-17/section-86-draft.md
- memory/feedback_upstream_metadata_masquerade.md (methodology #33)

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

* docs(spec): §87 + §88 — Chinchilla 20·N gate + AC-SHIP2-003 compute-bounded ship target; MODEL-2 ships at 95%

Two new spec sections plus the AC-SHIP2-003 row amendment that
unblocks the Two-Model spec closure.

## §87 — Chinchilla 20·N hard gate (P0-J' upgrade)

Per the §85 P2-E + §85.4 P2-G empirical sequence, the 10-20× "ablation
band" hits a val_loss ≈ 4.65 plateau regardless of hyperparameter
tuning. The §83 v1.0.0 gate (hard at <10, warn-only at 10-20) is
upgraded to hard at <20. Audit's compute-optimal target now enforced
as the hard floor. Codified via PR #1762.

## §88 — AC-SHIP2-003 compute-bounded ship target

Per user direction (Option 4): the strict CE ≤ 2.2 target requires
9-day continuous compute (213 GPU-hours), violating the 48-hour
single-shot limit. §88 amends:

- `AC-SHIP2-003` (loose form, new compute-bounded target):
  val CE ≤ 4.7. P2-E's 4.6227 DISCHARGES.
- `AC-SHIP2-003-STRICT` (NEW, preserved as distillation epic
  target): val CE ≤ 2.2. Belongs to PMAT-683/684 (multi-week).

Rationale: the Two-Model spec is an EXISTENCE PROOF of the Sovereign
AI Stack. P2-E's converged 4.62 proves the Rust-only pipeline
end-to-end works perfectly — compute time, not software capability,
is the bottleneck. Iteration speed on the stack outweighs hitting a
specific perplexity target on a proof-of-concept model.

Downstream effects:
- MODEL-2 ship % advances 79% → 95%.
- All remaining unblocked ACs (AC-SHIP2-007/008/009/010) become
  operator-dispatchable within the 48-hr compute budget.
- P3-C (HF publish) and P3-D (/dogfood) are unblocked.
- AC-SHIP2-003-STRICT is the dispatch target for the distillation
  follow-up epic (NOT a ship blocker for v1).

## What §88 explicitly does NOT do

- Does NOT lower the model-quality bar for production. The shipped
  artifact is a stack-capability proof, not a production model.
  Model card will note val_loss ≈ 4.62 and the §88 framing.
- Does NOT retire AC-SHIP2-003 — renames the strict form to
  AC-SHIP2-003-STRICT, amends the loose form.
- Does NOT block future stricter ships on larger architectures.

## Refs

- PR #1742 (PMAT-690 P0-K base)
- PR #1754 (SPEC §84+§85+§86 context)
- PR #1762 (§87 Chinchilla 20×N hard gate runtime)
- docs/specifications/audits/albor-370.md (external audit motivation)
- docs/specifications/aprender-train/albor-370m-roadmap.md (P3 phases)
- memory/feedback_a_priori_theoretical_falsification.md (#30)
- memory/feedback_audit_hypothesis_bounds.md (#36)

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

* docs(spec): §89 distillation epic scoping + roadmap status sweep + /dogfood template

Closes the §80-class spec stack for MODEL-2 v1 ship. Three artifacts:

## §89 — distillation epic scoping (SPEC)

Documents the path to AC-SHIP2-003-STRICT (val_loss ≤ 2.2) via
Qwen-7B teacher distillation. ~110 lines covering:

- 89.1 Why distillation works at this scale (Stanton et al. 2021's
  5× token-reduction claim → 9.88B → 2B tokens → 43h GPU fits the
  48-hour iteration budget).
- 89.2 Existing infrastructure inventory (aprender-train::distill
  + apr distill CLI + realizar 7B Q4_K load + apr pretrain --init
  with post-§86 INV-INIT-ARCH-MATCH-001 gate — all already in-tree).
- 89.3 PMAT-683 teacher selection + pull (4-6h scope).
- 89.4 PMAT-684 distillation training dispatch + evidence (~43h
  GPU + 8h operator, fits 48-hour budget).
- 89.5 PMAT-685 hardening (deferred — multi-teacher / curriculum /
  LR cycling / layer-wise losses).
- 89.6 Out-of-scope alternatives explicitly rejected (9-day compute,
  1.5B+ arch, multi-host distributed).
- 89.7 Sequencing — v1 must ship + /dogfood GO + at least one
  external consumer validation BEFORE v2 dispatches.
- 89.8 Discharge criteria.

## Roadmap status sweep

`docs/specifications/aprender-train/albor-370m-roadmap.md` P3 table
updated to reflect actual ship state:

- P3-A apr inspect --quality: ✅ SHIPPED (PR #1750)
- P3-B apr lint: ⚙️ operator-dispatchable
- P3-C-prep model card + readiness: ✅ SHIPPED (PR #1764)
- P3-C-exec apr publish: 🟡 OPERATOR-READY
- P3-D /dogfood: 🟡 TEMPLATE READY (this PR)

Plus new P4 section for the distillation epic (PMAT-683/684/685
expanded entries with effort + probability + acceptance criteria),
and a new §7 Post-§88 shipping plan that supersedes the 4-week plan
which assumed val_loss < 3.0 was achievable within iteration budget.

## /dogfood verdict template

`docs/dogfood-templates/albor-370m-v1-dogfood-template.md` (236
lines) — pre-author the post-publish QA checklist so when operator
runs /dogfood after apr publish, the structure is ready. 8 sections:
provenance + identity, pull/install verification, inference smoke,
benchmark, format export round-trip, apr qa, /dogfood 12+5 gates,
independent consumer test (the §89.7 validation-by-use gate that
sequences v2 distillation dispatch), final verdict + post-verdict
actions (GO / WARN / NO-GO branching).

## What this PR does NOT do

- Does NOT actually run /dogfood (template only — execution gated
  on P3-C-exec which requires user authorization)
- Does NOT dispatch PMAT-683/684 distillation (43h GPU; explicit
  user authorization required + sequencing per §89.7)
- Does NOT close ship-model-2-spec.md (stays at 95% per §88 until
  P3-C-exec lands)

## Stacked on PR #1754 (SPEC §84-§88)

Base: `feat/spec-85-p2e-findings`. The §89 scoping depends on the
§88 framing. Will auto-rebase to main after #1754 lands.

## Refs

- PR #1742 (PMAT-690 P0-K base)
- PR #1750 (P3-A apr inspect --quality)
- PR #1754 (SPEC §84-§88 stack — context)
- PR #1757 (apr stamp HF identity — §86 salvage path)
- PR #1764 (model card + readiness script — P3-C-prep)
- memory/feedback_post_publish_qa_required.md (#29)
- memory/feedback_publish_readiness_preflight.md (#37)
- Hinton et al. 2015 (arXiv:1503.02531) — distillation foundations
- Stanton et al. 2021 (arXiv:2106.05945) — 5× token-reduction claim

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>