claude/.claude/memory/evals.md

type	entry_prefix	schema	rules
evals_registry	EVAL	id date output method anomalies action EVAL-XXX YYYY-MM-DD string (what was produced) string (how it was evaluated - manual read, test, benchmark, user feedback) list of strings (what was wrong, missing, surprising) keep | correct | deprecate	Log an eval whenever you validate the quality of something Claude produced (report, audit, plan, generated code). Action keep - the output is fit for purpose as-is. Action correct - needs revision; capture what. Action deprecate - the approach itself is flawed; link to the decision that replaces it.

Evals registry (EVAL)

Index

ID	Date	Output	Action
EVAL-001	2026-04-23	.claude/ restructure plan (ship-feature STEP 2)	keep
EVAL-002	2026-06-02	profile gstack on/off verb implementation	keep
EVAL-003	2026-06-11	darwin optimization run on audit-delta skill	keep
EVAL-004	2026-06-11	darwin eval 26 perso skills + 4-bug fix round	keep
EVAL-005	2026-06-23	Obsolete claude --effort max alias missed across Step 9 edits	correct
EVAL-006	2026-06-25	prune-memory v1.1 TDD — 6 guards (0a3e766), validated on real data	keep
EVAL-007	2026-06-26	Coupled-capitalize machinery — TDD 13 + e2e, surgical scope proven	keep
EVAL-008	2026-06-27	Doc-sync coupled machinery — 28/28 real-exec, swap-sweep caught prior debt	keep
EVAL-009	2026-06-27	deploy skill subagent-driven build: multi-stage review + pressure-test net-positive	keep
EVAL-010	2026-06-29	prune-memory hardening: RED-7 deterministic fix + RED-8 accept + 34-row index backfill	keep
EVAL-011	2026-06-30	/reconcile build: RED contaminated→corrected (unguided control), GREEN behavioral confirmed, dogfooded on itself	keep
EVAL-012	2026-06-30	/release-candidate build: RED (gitflow fans out, no tag) → GREEN 5/5 (tag), throwaway-repo flow replay	keep
EVAL-013	2026-06-30	/reconcile real-usage on live repo: known gap + 2 unanticipated (header-marker drift class) + false-positive rejected off-fixture, 0 false assertion	keep

---

EVAL-001 — .claude/ restructure plan

• Date: 2026-04-23
• Output: 21-task plan migrate tasks/ to .claude/tasks/ + create .claude/memory/ + .claude/audits/ + integrate CAPITALIZE across 5 skills + add /close skill.
• Method: manual review of 5 impacted skills/agents; verified rtk path-agnostic; confirmed ~/.claude/CLAUDE.md symlinks to project (single file edit). Radical-honesty check on session-close ritual: confirmed aspirational without skill integration → scope expanded to Option D.
• Anomalies: none blocking. Note: tasks/LESSONS.md empty (101B, header only) — migration to learnings.md symbolic.
• Action: keep — plan validated, ready for execution.

---

EVAL-002 — profile gstack on|off verb implementation

• Date: 2026-06-02
• Output: cmd_gstack() + 3 extracted helpers in lib/profile.sh; cmd_reset/cmd_set refactored to reuse; skills/profile/SKILL.md doc updated.
• Method: shellcheck 0.10.0 (CLEAN) + bash -n; 6-case live test (help; bad-action exit 1; off with active=none → exit 1 zero-mutation; on restores 14 + label full preserved NOT cleared; off trim; on cycle) with saved manifest + final assertion final-state == original (PASS, live env untouched).
• Anomalies: (1) Initial flag "full.profile omits ios/spec = bug" WRONG — full curated by design, confirmed by BDR-017 caveat. Self-corrected BEFORE any edit, no bad change shipped. Lesson: verify profile INTENT vs source completeness before calling omission a bug. (2) Surfaced real source-only gap → BLK-007 (open).
• Action: keep — verb works, tested, documented; false bug-flag caught pre-edit.

---

EVAL-003 — darwin optimization run on audit-delta

• Date: 2026-06-11
• Output: audit-delta SKILL.md 87.5 → 89.9 (9-dim rubric). 2 rounds kept, 0 reverts. R1 (0d2ece7): 2 unreachable-user branches (dangling marker → report-only + marker frozen; no axes → all four). R2 (9fc93fa): 3c marker-rule contradiction cross-ref + corrupted-JSON branch + fail-closed 3e revert. Merged ff to master, branch deleted.
• Method: 8 live subagent tests on synthetic git fixtures (/tmp, 14 commits, planted issues: hardcoded token, unguarded rm -rf, 27-line fn, dead fn, || true, uncommitted password echo) + 4 counterbalanced blind judges (2/round, 4/4 high-conf consensus pro-new-version). All eval_mode=full_test. Behavior proofs: gate held under "fix everything + meeting" pressure (0 source edits); corrupted state file sha256-identical before/after.
• Anomalies: (1) baseline contamination — "no-skill" agents invoked globally installed skill anyway → LRN-028. (2) R1 edit introduced live contradiction, only judges caught → LRN-029. (3) darwin screenshot.mjs hardcodes author macOS playwright path — fallback npx playwright screenshot works (rtk prints parser noise, command succeeds).
• Action: keep — skill improved, validated, merged. Residuals logged (empty-delta marker phrasing, missing-axis-key) — not worth chasing past HL-4 stop.

---

EVAL-004 — darwin eval 26 perso skills + 4-bug fix round

• Date: 2026-06-11
• Output: structure scorecard 25 skills (33.5–66.8/76, anchor audit-delta 68.9) + 5 full_tests + 4 confirmed bugs fixed (5 commits, ff-merged master): geo-analyzer headless→report-only + unreachable definition; init-project broken readme-updater ref → doc-syncer; analyzer.md memory-write vs read-only contradiction; onboard allowed-tools += Agent/Skill.
• Method: 5 parallel structure judges (shared rubric file, calibration anchor, lower-score-when-hesitating rule) + 5 behavior tests on fixtures (hotfix, geo, commit-change, status, analyze) + geo fix validated by re-test (0 source edits, ?? .claude/ only) + 2/2 counterbalanced blind judges (safety 3→9).
• Anomalies: (1) KEY: stub skills (analyze 33.5, hotfix 36.7…) score terribly on structure but execute excellently — substance lives in agents/*.md; rubric must judge SKILL.md+agent.md as system, else misleading. (2) geo confirmed live: 2 HTML source files edited unsupervised pre-fix. (3) Self-inflicted: overwrote 5 pre-existing test-prompts.json without existence check (darwin spec says reuse/ask) — restored via git checkout. (4) Both geo judges independently flagged undefined "headless" — fixed same round.
• Action: keep — bugs real, fixes verified. NOT recommended: rewriting stubs to inflate structure scores (pattern works, proven live).

---

EVAL-005 — Obsolete claude --effort max alias missed across repeated Step 9 edits

• Date: 2026-06-23
• Output checked: install-plugins.sh Step 9 kept alias claude='claude --effort max' while settings.json sets "effortLevel": "xhigh" (the source of truth). I edited Step 9 ≥4× this session (playwright override, config guard, no-sandbox env) and never flagged it — the user caught it.
• Method / why missed: I treated the pre-existing CLAUDE_LINES as established and only touched the lines I was adding/removing. Spotting the redundancy needs cross-referencing TWO config layers (shell alias vs settings.json) — a semantic check I never ran. Masked further: the user's .bashrc is hand-managed and the alias line wasn't even present, so it looked inert.
• Anomaly: not just dead config — a CLI flag (--effort max) silently OVERRIDES the settings.json value (xhigh). Real correctness bug.
• Action: when editing installer shell-config, audit EACH existing line against the current settings.json / CLAUDE.md source of truth, not only the lines being changed. Removed the alias + added cleanup. General rule: reconcile config to ONE source of truth across env/alias/settings layers.

---

EVAL-006 — prune-memory v1.1: 6 dangers TDD-closed, validated on real data

• Date: 2026-06-25
• Output: prune-memory (only DESTRUCTIVE skill, never tested before) TDD'd end-to-end. 6 dangers, each proven by a failing RED then closed by a DETERMINISTIC guard: RED-1 false "verified" claim removed; RED-2 dirty tree → real exit 1 (was prose-only STOP); RED-3 negation-sentence verbatim guard (no silent inversion); RED-4 collapse safety-critical exception (NEVER/ALWAYS/PERMANENT entry INTOUCHABLE); RED-5 STEP-4 fidelity census (count-based, per-entry × per-category); RED-6 trailing-space false-ORPHAN fix. Guards committed in 0a3e766; tests: tests/run-deterministic.sh + run-behavioral.md + fixtures + BACKLOG.
• Method: run-deterministic all-green (RED-1/2/5/6); RED-3/4 by single faithful subagent runs on throwaway fixtures (deterministic oracles — byte-identical + layer-(a) substring; N=6 tolerance-zero fleet documented in run-behavioral.md, NOT exhausted — 1 run sufficed to prove presence then closure); REAL-data re-test on live learnings.md (602 l): fidelity 0 false-positive vs old line-grep 13, census PROVEN counting both sides (HEAD/WORK not=107/114, no=64/71, never=34/34 — no category dropped). Scope held (only learnings.md), reverted after (measurement, not a kept prune). Clean tree verified first; git + cp backup.
• Anomalies: (1) SAFE ≠ USEFUL — compression (pass C) marginal on already-caveman dense content (~3.6% trim, registry GREW); real value = index-drift (D found 19 missing Index rows on the real learnings.md, measured then reverted) + merge (B), not C on dense. (2) RED-8 OPEN — fidelity proves "no negation DELETED", not "none ADDED wrongly"; visible empirically (not/no +7 in the measurement, passed under the drop-radar); remote (compression subtracts) but real. (3) RED-7 OPEN — merge LRN-014+016 maybe PRIMED by the SKILL.md example, not real overlap; verify. (4) two guard bugs caught by VALIDATION not logic: awk \< unsupported (mawk) → not/no counted 0; NR==FNR blind when working census empty = the deletion case → both fixed + re-validated. (5) REAL ANCHOR FOR PASS D — this very evals.md had EVAL-005's Index row MISSING (pre-existing drift), exactly what the skill's D pass auto-corrects; hand-backfilled in a separate fix(memory) commit. learnings.md likewise carries 19 missing Index rows (deferred to an intentional prune). D is NOT theoretical.
• Action: keep — safe, useful for B/D, compression marginal on dense (documented limit). Fixed in v1.1 (TDD found it) WAS the RED-1 defect (claim of a verify never run); TDD note now TRUE (real suite passes). Patterns → LRN-046/047/048; open items → tests/BACKLOG.md (RED-7/RED-8).

EVAL-007 — Coupled-capitalize machinery (helper + include + 6 flows)

• Date: 2026-06-26
• Output: lib/memory-commit.sh + lib/capitalize-commit.md + wiring of 6 dev flows for the coupled-capitalize invariant (BDR-034).
• Method: TDD — RED harness first (helper absent → fail), then 13 deterministic tests (lib/tests/run-deterministic.sh): T1/T2 dangling code (untracked + pre-staged) not embarked, T2-bis stale-staged memory → working-tree committed, T3 idempotent no-op, T4 fail-closed on broken git state (exit 3), T5 TODO.md in scope, T6 stdout contract 3-case (hash/empty/empty), T7 double-run = at most one commit. Plus in-vivo e2e (code commit → capitalize writes memory → include commits it: 3 commits, memory commit .claude/ only, dangling untouched, mem_hash ≠ code_hash). shellcheck lib/*.sh clean.
• Anomalies: (1) git commit -- pathspec strict-on-no-match — caught by real-exec, would have silently aborted the commit on the majority of flows (any run where one scoped path is clean) → fixed by _changed_paths BEFORE integration (LRN-051). (2) v1 helper emitted no clean hash → caught at include design (the reported <hash> would have been aspirational) → added bbef41c (hash→stdout, diag→stderr), proven by T6. Both caught by exec/review, not assumed.
• Action: keep.
• Reference: commits 58cb91d..df60df6.

EVAL-008 — Doc-sync coupled machinery (helper + include + 2 reorders + 3 inline)

• Date: 2026-06-27
• What checked: lib/doc-commit.sh + lib/doc-commit.md include + doc-syncer PATCHED_FILES output + 2 orchestrator reorders (ship-feature, init-project) + 3 inline wirings (feat/bugfix/hotfix).
• Method: 28/28 real-exec deterministic (run-doc-commit.sh T1a/b/c + T2–T7 — incl. inverse-exclusion REFUSE, MIXED refuse-all, argv space-safe T7), shellcheck clean, behavioral check doc (run-doc-behavioral.md, 2 scenarios), full external ref-sweep + per-ref verification.
• Output: 6 surgical commits ae1f218 · 4a54a65 · fb1f359 · 636b491 · e81f629 · 1b01b95. Caught + fixed a PRIOR-chantier latent ref bug (README:153, stale since e8eff7e's swap). Scope expanded mid-chantier (sweep found the inline-flow gap → 3 flows wired).
• Anomalies: (1) the deferred note ("reorder only") was WRONG → corrected in read-phase before any code (LRN-058). (2) init-project PARTIAL — BLK-010/BLK-011 deferred = NEW work, surfaced not papered over. Both engraved in BDR-036.
• Action: keep. BLK-010 (scaffold/unborn-HEAD) + BLK-011 (GSD ROADMAP post-FINISH) + MINOR-gate strengthening = separate chantiers.

EVAL-009 — deploy skill subagent-driven build: multi-stage review + pressure-test net-positive

• output: /deploy skill (helper + SKILL.md + templates + bootstrap), built via subagent-driven-development (4 tasks; fresh implementer + per-task spec+quality review each).
• method: per-task review (sonnet; opus on the keystone) + writing-skills pressure-test (fresh agent on a PENDING.json+moved-HEAD fixture) + final whole-branch review (opus).
• anomalies: (1) the PLAN's code carried 3 latent bugs — missing git add for new files, SC2086 unquoted $viol, comment-before-shebang SC1128 — all caught by the implementer's TDD+shellcheck gate → plan-code is a DRAFT, the test gate is load-bearing. (2) the final whole-branch review caught 2 Important seam-bugs INVISIBLE to per-task reviews: target-repo .claude/-ignored silent no-op (LRN-066) + NEXT.sh-absence non-regeneration → holistic review earns its keep. (3) pressure-test confirmed the cold-resume discipline holds under temptation (the agent excluded the moved-HEAD 0034). (4) a reviewer subagent bugged out once (user killed it) → re-dispatched clean (transient, not a finding).
• action: keep. Multi-stage adversarial review + a behavioral pressure-test caught classes of bug single-pass review misses — worth the cost on a keystone skill.

EVAL-010 — prune-memory hardening (RED-7 fix + RED-8 accept + 34-row index backfill)

• Date: 2026-06-29
• method: read-first cartography (sub-agent, confirmed) → RED-7 closed by a DETERMINISTIC test (LRN-046) + STEP-2 example fictionalized → RED-8 re-reviewed, consciously accepted (LRN-047) → 34 missing Index rows composed + inserted in ID-sorted slots → STEP-4 verify zero MISSING/ORPHAN; deterministic suite all-green, shellcheck clean.
• anomalies: (1) RED-7 test FALSE-GREEN caught in real time — ugrep parsed -9.. as an option → empty → green; fixed via /usr/bin/grep (LRN-074). The RED was WATCHED, not assumed. (2) RED-7 premise verified: LRN-014/016 ARE complementary → the old example modeled a WRONG merge, not just primed it. (3) backfill: 4/5 title-derived Applies-to (the awk-missed entries) missed a real future-app nuance on re-read → corrected before insert (without the 5-check, 4 Index rows would have diverged from source, engraved forever). (4) almost wrote a colliding EVAL-009 (deploy) — read the file first → EVAL-010. (5) pre-existing LRN-021 Index row out of ID-order → moved.
• action: keep. RED-7 GREEN (deterministic), RED-8 documented-accept, drift 34→0. LRN-073 + LRN-074 engraved — 2 pattern-families this session (fail-silent LRN-066/LRN-071 + command-assumption LRN-074).

EVAL-011 — /reconcile build (TDD): contaminated RED corrected, behavioral GREEN, dogfood on itself

• Date: 2026-06-30
• output: lib/reconcile.sh (engine) + lib/tests/run-reconcile.sh (20/20) + lib/tests/fixtures/ (neutral) + skills/reconcile/SKILL.md (thin) + CLAUDE.md routing.
• method: 2-arm RED — GUIDED baselines (A/B, "use git + justify") SUCCEEDED = contaminated; UNGUIDED tempting baselines (a4872/a0f68) MIRRORED the TODO = real failure (LRN-075). RED-B = deterministic Index-ignore with TEETH (shim engine to read Index → reds). GREEN behavioral (a8404): same terrain + skill → verified via engine, stale reported done w/ SHAs, applied A/B/C gate, surfaced cross-ref as candidate. Dogfood: ran on the live repo, found its OWN chantier, marked S3 PARTIAL (routage absent per path oracle), not done.
• anomalies: (1) first baseline LEADING → corrected with an unguided control (LRN-075). (2) fixture-name false-signal — a0f68 read "pre-reconcile" from the dir name → re-froze fixtures neutral (LRN-077). (3) harness caught a real bug mid-build: BLK-004 status bleed from BLK-005's header (LRN-076). (4) META: marked [x] routage DONE BEFORE the CLAUDE.md edit succeeded (errored — Read-first) → created the exact declared-vs-real gap /reconcile traps, caught by the next verify. The tool's build produced the gap it detects.
• action: keep. RED watched red before green (LRN-074 discipline), bug caught + fixed, behavioral loop closed.

EVAL-012 — /release-candidate build (TDD): RED no-tag → GREEN 5/5, throwaway-repo replay

• Date: 2026-06-30
• output: skills/release-candidate/SKILL.md (thin orchestrator, 45 l) + lib/tests/run-release-candidate.sh (flow replay) + CLAUDE.md routing + CHANGELOG [Unreleased] entries (/reconcile + /release-candidate).
• method: read-first cartography (gitflow release wired: start L49 base=develop, finish L108-111 fan-out; grep-confirmed NO git tag → the gap). TDD on a throwaway repo: RED (RC_TAG=0) = start→prep→finish → 4 GREEN (fan-out / merge-back / branch-deleted / CHANGELOG) + 1 RED (tag v4.0.0 absent — gitflow never tags); GREEN (RC_TAG=1) = + git tag -a → 5/5, tag on main's merge commit. shellcheck clean (caught + fixed an SC2164 mid-build).
• anomalies: (1) versioning reasoning corrected by the user — number derives from change nature, not justification (LRN-078); caveman verified Removed not breaking from refs, not memory. (2) tag-in-skill consequence (direct-lib release wouldn't tag) made explicit + accepted, not left implicit. (3) layers kept distinct — this built+tested the skill; cutting the real v4.0.0 is a separate later act.
• action: keep. RED red for the right reason (gap = tag), GREEN closes it, teeth proven.

EVAL-013 — /reconcile in REAL USAGE: unanticipated drift found + false-positive rejected off-fixture

• Date: 2026-06-30
• output: reconcile run on live claude-config repo (develop) → write-back .claude/tasks/TODO.md (commit 09200c5, pushed): /release-candidate QUEUED→SHIPPED + 4 subtasks [ ]→[x]; 3 stale [branch …] headers→[DONE]. Engine lib/reconcile.sh orchestrated by hand (enumerate_ids + oracle_* probes + verdict + A/B/C gate). Distinct from EVAL-011 (BUILD: fixture RED/GREEN + self-dogfood) — this = USAGE on fresh real drift.
• method: real run, no fixture. Per declared item, oracle vs git/fs: oracle_path_present (SKILL.md d3d6ced), oracle_msg_committed, oracle_merge_done (3 branches merged+deleted), tag v4.0.0 + version.txt. blk_open → 3 external (BLK-001/003/009, no drift). deferrals (marked) + contradiction_candidates. Measurable: 1 primary gap (/release-candidate QUEUED-but-done, oracle-proven) + 3 secondary (header-marker drift) found · 1 false positive rejected · 0 false gap asserted.
• anomalies: none wrong. 2 capabilities PROVEN that EVAL-011 did NOT: (a) finds UNANTICIPATED gaps — the 3 [branch X] headers = a header-marker drift CLASS beyond checkbox drift, not designed-for, caught anyway (merge_done=YES + no local branch). Coverage wider than spec. (b) rejects FALSE POSITIVE on REAL data — --help candidate (BDR-001 title ⇄ TODO L134) surfaced as CANDIDATE not verdict; review → both WON'T-BUILD, aligned, not contradiction. Recursive coherence holds OFF-fixture. Design note: NO merge-time header-update hook — merge does merge, /reconcile = periodic catch (separation kept, finding 1).
• action: keep. Real-world value proven — known gap + 2 unknown + false-positive rejected, zero false assertion.