docs(spec): SHIP-TWO-001 §61 — post-§60 LIVE-discharge cascade + ChatML generation gap

[noahgift]

Summary

Records the empirical findings from this session's LIVE-discharge cascade attempt off §60. §60 closed forward parity (layer-3 ratio 18.23× → 1.245× ∈ [0.5, 2.0]); §61 surfaces that forward parity ≠ generation parity under all prompt formats.

Two-track outcome

DIRECT prompt → GREEN: SHIP-002 LIVE-discharged via PR #1609. apr run --prompt "def fib(n):" --max-tokens 128 on canonical 7B APR teacher emits coherent fib() Python (ast.parse 0 syntax errors, 68 nodes, 1 FunctionDef "fib").

ChatML-wrapped prompt → BLOCKED: SHIP-006/008 cannot LIVE-discharge today. Same canonical teacher fails apr qa golden_output gate with "gibberish (fragment '\ns\ns' repeats 3+ times)" under <|im_start|>user\nWhat is 2+2?<|im_end|>\n<|im_start|>assistant\n.

Five-Whys

1. Why §61 needed? §60 closed forward parity but SHIP-006/008 LIVE-discharge attempts failed empirically.
2. Why didn't ship-% auto-flip 91% → 96%? Forward parity is binding criterion at activation-stats level; arg-max sampling under cumulative drift not directly bounded.
3. Why does prompt format matter? Direct prompts → high-confidence next-token regime where drift doesn't flip arg-max. ChatML → low-margin regime where drift CAN flip arg-max.
4. Why record vs fix now? Bug is multi-PR scope; PRED-61-A/B set up next bisection.
5. Why durable spec? Each day spec doesn't reflect §60 → §61 separation, future sessions may misinterpret §60 closure as full SHIP-007-class discharge.

§61.5 falsifiable next predictions

• PRED-61-A: GGUF + ChatML on canonical 7B → clean? If GREEN, bug is APR-side in chat-template handling.
• PRED-61-B: APR + direct continuation "What is 2+2? The answer is " (no ChatML) → clean? If GREEN, bug is special-token handling NOT cumulative drift.

Changes (1 file)

• docs/specifications/aprender-train/ship-two-models-spec.md:
  • Atomic next action banner: v3.05.0 → v3.06.0
  • New §61 section (above §58, newest-first ordering) with 7 sub-sections (separation table / direct-prompt evidence / ChatML-prompt evidence / §60→§61 separation rationale / falsifiable predictions / ship-% movement / what §61 is NOT)

Ship-% Movement

• MODEL-1 ship %: 91% → 92% (1 of 5 §17.5 PARTIALs LIVE-discharged via PR feat(contracts): SHIP-002 PARTIAL → DISCHARGED via LIVE apr run on canonical 7B teacher #1609)
• MODEL-2 ship %: unchanged at 57% (gated on step 5g.3 val_loss < 9.38)

Cascade-this-session

6 PRs in 24h working SHIP-TWO-001:

• fix(aprender-train): CUDA forward path applies Q/K/V biases (H4D root-cause discharge) #1604 H4D Q/K/V bias dispatch
• fix(aprender-train): CUDA RMSNorm honours config.rms_norm_eps (Qwen 1e-6 vs hardcoded Llama 1e-5) #1606 RMSNorm eps cache key
• fix(aprender-train): include theta in CUDA RoPE PTX cache key #1607 RoPE theta cache key
• chore(contracts): apr-vs-gguf-forward-parity-v1 v1.2.0 — promote PROPOSED → ACTIVE_FUNCTIONAL (§60 closure) #1608 apr-vs-gguf parity v1.2.0 promote (§60 closure)
• feat(contracts): SHIP-002 PARTIAL → DISCHARGED via LIVE apr run on canonical 7B teacher #1609 SHIP-002 LIVE discharge
• (this PR) §61 spec amendment

Test Plan

• No code changes; pure documentation
• Section format consistent with §58 (newest-first, sub-sections numbered §61.X)
• All 6 cascade PRs referenced explicitly
• Methodological alignment: zero eprintln!, all evidence via existing apr CLI primitives

🤖 Generated with Claude Code