Summary

Discharges FALSIFY-SHIP-018 / AC-SHIP2-008 / GATE-ARCH-370M-007 at
PARTIAL_ALGORITHM_LEVEL. Lands a pure two-number threshold function
verdict_from_pass_at_1(correct, total, threshold_pct) -> Ship018Verdict
in crates/aprender-train/src/models/llama_370m.rs plus const
AC_SHIP2_008_MIN_HUMANEVAL_PASS_AT_1_PCT = 30.0 pinning the spec
§5.2 AC-SHIP2-008 floor — proves the HumanEval pass@1 decision rule
at cargo test time, independent of a trained 370M artifact.

This is the 6th PARTIAL for MODEL-2 (after SHIP-012/015/017/019/020).
Spec v2.22.0's "exhausted" verdict is now falsified 4 times.

What ships today

• verdict_from_pass_at_1(correct: usize, total: usize, threshold_pct: f32) -> Ship018Verdict
  — inclusive-floor comparison with conservative-Fail guards.
• Ship018Verdict enum { Pass, Fail }.
• AC_SHIP2_008_MIN_HUMANEVAL_PASS_AT_1_PCT: f32 = 30.0 (contract floor).
• Two unit tests:
  • falsify_ship_018_humaneval_pass_at_1_threshold_logic — boundary
    (f32-exact 50/100 = 50.0% with ±ULP proving inclusive >=),
    below-floor rejection (49/164 ≈ 29.88%, 29/100 = 29.0%), generous
    pass (82/164, 164/164), hard-red (0/164, 1/164), monotonicity
    sweep (correct ∈ 0..=164 at total=164 never flips Pass → Fail),
    div-safety (total=0 fails closed), correct>total sanity, non-finite
    threshold guard (NaN / ±∞ all Fail), provenance pinning (const
    stays = 30.0).
  • falsify_ship_018_gate_arch_370m_007_has_partial_discharge_marker
    — parses SOVEREIGN_CONTRACT_YAML via serde_yaml, asserts the
    gate binds FALSIFY-SHIP-018 / AC-SHIP2-008,
    discharge_status == PARTIAL_ALGORITHM_LEVEL,
    evidence_discharged_by non-empty, full_discharge_blocks_on
    present, ship_blocking: true.
• contracts/model-families/llama-370m-sovereign-v1.yaml v1.5.0 →
  v1.6.0 (stays ACTIVE) with new GATE-ARCH-370M-007 block + changelog
  entry.
• docs/specifications/aprender-train/ship-two-models-spec.md v2.23.0
  → v2.24.0 with amendment block.

Pattern lesson

The "exhausted" verdict has now been falsified 4 times:
SHIP-019 → SHIP-017 → SHIP-020 → SHIP-018. Whenever a SHIP-spec
gate names a threshold/tolerance/ratio/cut-off and the compute-heavy
harness is separable from the decision function, the threshold
function can land today. Remaining 5th-PARTIAL candidate worth a
survey pass: SHIP-016 (apr qa 8-of-8 aggregate — separable
but not a single-threshold proof). SHIP-013/014 genuinely need real
RTX 4090 compute (CE loss + 21-day wall-clock).

Full discharge blocks on

Real 370M .apr checkpoint from AC-SHIP2-003/004 compute-dispatch +
three independent apr eval --benchmark humaneval --json median
pass@1 values at seed=0 on the SHIP-TWO-001 canonical host; feed
each into verdict_from_pass_at_1 and require all three Pass.
Fixture-swap only — no harness rewrite required.

Status after this PR

MODEL-2 ship-gates: 3/12 ACTIVE (001, 011, 012) + 6/12 PARTIAL
(002 via SHIP-012, 005 via SHIP-015, 007 via SHIP-017, 008 via
SHIP-018, 009 via SHIP-019, 010 via SHIP-020) = 9/12 touched (75.0%).
Remaining 3 genuinely compute-blocked: 003 (CE ≤ 2.2 val loss),
004 (≤21-day wall-clock), 006 (apr qa aggregate).

Also bundled

A 6-line pre-existing fmt fix in crates/aprender-train/src/train/device.rs
under Toyota Way "all defects are your defects" — same pattern as
PR #1005. Without it, cargo fmt -p aprender-train --check fails on
main for reasons unrelated to this PR.

Test plan

• cargo test -p aprender-train --lib models::llama_370m →
  11/11 pass (2 new + 9 existing).
• pv validate contracts/model-families/llama-370m-sovereign-v1.yaml
  → Contract is valid.
• cargo fmt -p aprender-train --check → clean.
• cargo clippy -p aprender-train --lib -- -D warnings → clean.
• CI (workspace-test, ci/test, ci/lint, ci/coverage) green.

Refs: SHIP-TWO-001, task #151, FALSIFY-SHIP-018, GATE-ARCH-370M-007.

🤖 Generated with Claude Code