lattice(phase2): H2/H5/H4 verdicts + close CD-H1-01/CD-A2-01 + docs-truth sync

[mattmre]

Summary

Continues the Liquified Lattice Phase II program: runs two GPU campaigns whose apparatus was already
merged (H2 swap re-run, H5 post-bank head-to-head + H4 ablation), closes the two open Carried-Debt
rows, and syncs the roadmap/CHANGELOG to git-verified reality. Docs + evidence only — zero
production .py changes.

Branch: lattice/phase2-continue-20260713 (4 commits). Block flag: CLEAR (0 carried-debt rows).

Commits

Commit	What
results(h2)	Re-run SciFact + NFCorpus query-encoder-swap campaigns on the H1-fixed line; regenerate both result docs. Closes CD-H1-01 + CD-A2-01.
results(h5)	Run the merged H5 head-to-head driver (#290) + H4 compound_cycles (#291). Living-bank verdict: hard negative.
docs(lattice)	Phase II docs-truth: ROADMAP gains a Status column (rungs 9–12 DONE / 13 PARTIAL / 15–17 OPEN); CHANGELOG apparatus + open-steps bullets.
chore(hygiene)	Untrack the two residual #275 NFCorpus campaign logs (completes the L1 tracking-consistency claim; Tier-B-caught).

Key results (every number recomputed from committed manifests by the Tier B reviewer)

H1-fix regression gate PASSED — C3a baseline now equals C0/C2/C2O/C4a per seed on both datasets
(SciFact seed-42 all 0.793459; NFCorpus seed-42 all 0.555004). The contamination was on NFCorpus
(old C3a baselines differed); SciFact was already clean. Closes CD-H1-01. The run exercised the real
all-mpnet-base-v2 swap path through QueryEncoderDrift._backend — closes CD-A2-01.

H5 living-bank verdict = LIVING BANK WINS: False on both datasets (gate: C5 living must beat BOTH
C5s frozen-static AND C5r one-shot router):

Dataset	C5 living	C5s static	C5r one-shot	C5>C5s	C5>C5r	WINS
SciFact	0.131135	0.131135	0.180862	False	False	False
NFCorpus	0.046389	0.046389	0.045817	False	True	False

The living/annealed lifecycle is bit-identical to a frozen static bank; a one-shot router is
competitive-or-better. Living bank NON-PROMOTED. H4 (single-seed): compound_cycles=True 0.005258
vs False 0.236297 (~45× collapse) — compounding is a rejected design.

Evidence / smoke (human-readable; machine fields below)

Result docs: the five docs/drift-recovery-*.md files listed above; summary manifests + H4 run JSONs
under experiment_runs/drift-recovery/. Smoke (floor-tier): H1 regression parity PASS;
scripts/check_block_flag.py → CLEAR; test_post_bank_head_to_head OK;
test_query_encoder_drift.TestQueryEncoderDriftRealModelSmoke (CHELATED_RUN_REAL_MODEL_TESTS=1) →
Ran 1 test, OK. The campaigns themselves are the end-to-end evidence.

Brutal Honesty

• No production code changed (verified: git diff --name-only origin/main..HEAD has zero root .py).
• Tracking hygiene: untracked 22 NFCorpus per-run JSONs (25 MB) + 2 residual results(a5): NFCorpus swap campaign — partial recovery generalizes to second dataset #275 logs so NFCorpus
  matches SciFact's manifest-only policy; the small summary manifests carry the doc-source rows.
• L5 (disclosed): the real-model swap smoke is opt-in (not default CI, avoids a 17-min HF hang);
  H4 is a single-seed (42) ablation — direction unambiguous, magnitude one seed.
• Step 13 is PARTIAL, not DONE — delivered post-bank prune/re-anneal is not the
  isomer/convergence → Evidence-DAG prune the original exit criteria named; flagged for re-scope.
• Older [Unreleased] Findings (Phase I / SHIM-on-hold) are now stale vs lattice delivery; left in
  place and deferred to a separate hygiene pass (DEFERRED_SCOPE).

The initial combined Tier B scored 97 with one LOW hygiene defect (two residual #275 NFCorpus logs
still tracked); commit eb750958 untracks them, and a fresh Tier B pass on the amended tip re-scored
the slice to a genuine 100 (severity none, no new defects). Every NDCG and gate was independently
recomputed from the committed manifests by the reviewer.

§4 required fields

EVIDENCE: five docs/drift-recovery-*.md result docs + summary manifests + H4 run JSONs under experiment_runs/drift-recovery/ (per-number recompute done by Tier B)
SMOKE: floor-tier — H1 regression parity PASS; scripts/check_block_flag.py CLEAR; test_post_bank_head_to_head OK; test_query_encoder_drift.TestQueryEncoderDriftRealModelSmoke (CHELATED_RUN_REAL_MODEL_TESTS=1) Ran 1 test OK; the campaigns are the end-to-end evidence
BHS_SELF_DRAFT: 100
BHS_SELF_DRAFT_AGENT: Fable-5 (chair / implementer)
BHS_TIER_B: 100
BHS_TIER_B_AGENT: grok-4.5 (fresh adversarial reviewer, independent of the implementer)
BHS_TIER_B_SEVERITY: none
BHS_OFFICIAL: 100
CARRY_FORWARD: none
DEFERRED_SCOPE: refresh of the older stale [Unreleased] Findings bullets (Phase I / turn-3171 / SHIM-on-hold) — separate hygiene pass; and step-13 exit-criteria re-scope (post-bank lifecycle vs isomer/convergence→DAG prune)
LOOP_ITERATIONS: 1
OPERATOR_OVERRIDE: none

🤖 Generated with Claude Code

[gemini-code-assist]

Code Review

This pull request updates the project's changelog, roadmap, next-session plans, and experimental manifests to document the outcomes of Phase II drift-recovery experiments. Key updates include recording the negative verdict of the H5 living-bank head-to-head evaluation (resulting in its non-promotion), documenting the H4 compound-cycles ablation study, and updating the query-encoder-swap results following the H2 re-runs, which successfully closes carried debts CD-A2-01 and CD-H1-01. There are no review comments, and therefore no feedback is provided.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.