Improvement loop — weekly triage

[theboycoder]

Rolling tracker for the improvement-loop triage agent (runs Sun/Tue/Thu). Each run posts one consolidated comment: engine states, the PR opened (if any), and the queued code-class items with reproducible probes. The agent reads its own latest comment first to report deltas and spot recurring failure classes.

Scope reminder: the agent fixes only the safe data-registry class by PR (never merges), and queues everything touching routes / ranking / ingest / the OpenAPI contract for a human session. Precision over recall — every data change is verified from the project's own primary sources.

Reports below. ⬇️

[theboycoder]

🔁 Run 2026-07-19 (first run — baseline)

1. Since last run

First run — no prior report. This establishes the baseline. No open data/improvement-loop-* PR existed; the fork network to theboycoder is broken (isFork:false, cross-repo PR 422s), so this run's PR was pushed as a branch on this repo directly. Future runs: same-repo branch PRs.

2. Engine states

Engine	State
self-audit (local)	🟢 54/54 — the last CI red (wowmax Bridge-typed, empty supportedNetworks) self-cleared; it now carries ['evm'] on live.
golden (local)	🟢 45/46 (CI last: 46/46), forbidden-hits 0, above pass≥31 gate
retrieval health #436	🟢 golden 46/46, all recall floors held
engine-c-health (today's scheduled run)	🔴 by design — 250 probe failures = the fix queue (below), not an outage
coverage #562	3 llama-missing · 107 scf-missing · 52 NEW gaps / 0 resolved this month
liveness (engine-c B)	385 Live w/ no dated signal · 140 Live stale >365d (liveness-wave pool)
SCF cross-check	scfAwarded 0 overstated / 0 understated 🟢; round membership 14 overstated 🔴

3. PR opened

#585 — paper-trail catch-up only (no safe data-registry row this run). Adds improvements/waves/analytics-corpus-recency-2026-07-17-18.md covering #571–584 (corpus recency-measurability + first public /analytics surface) that merged after the quality-week wave left the folder at 07-16. No lessons-table row (patterns already recorded). Docs-only; biome clean.

Why no data fix: the audit red cleared on its own; the llama-missing entries (gami-labs / defa-invoicemate / near-intents) exist as projects but their Stellar chain slice is negligible (quality-week lesson: NEAR Intents = $11.5k on Stellar) — low-value speculative wiring, held as report-only.

4. Queued code-class items (priority order, with probes)

P1 — camelCase brand tokenization drops trailing single-capital token (new, confirmed)
StellarX (the brand's own casing) returns unrelated wallets; every other casing works.
```
curl -s "https://stellarlight.xyz/api/projects/search?q=StellarX&limit=3" # → [lobstr, xbull, reflector] ✗ (stellarx absent)
curl -s "https://stellarlight.xyz/api/projects/search?q=stellarx&limit=3" # → [stellarx, ...] ✓
curl -s "https://stellarlight.xyz/api/projects/search?q=STELLARX&limit=3" # → [stellarx, ...] ✓
```
Narrow: xBull/StellarTerm are fine — only a trailing lone-capital token (Stellar+X) breaks strict-AND. Fix in the search tokenizer/normalizer (route). Own-name P-KNOWN miss for a major Stellar DEX.

P2 — builders lookup by person name returns 0 (high real demand — Engine D)
```
curl -s "https://stellarlight.xyz/api/builders?q=tyler%20van%20der%20hoeven&limit=3" # → 0 results
```
Also 0 for strupey (17 hits), justin rice (9), tomer weller, david mazières, kalepail. Real consumers repeatedly ask /api/builders for named SDF/core people we don't index. Ingest/data-source gap (builders index doesn't cover named individuals).

P3 — natural-language phrase queries dilute the entity (P-PHRASE, 242 probes)
```
curl -s "https://stellarlight.xyz/api/projects/search?q=what%20is%20DIA&limit=3" # → [lobstr, xbull, idos] ✗ (dia absent)
curl -s "https://stellarlight.xyz/api/projects/search?q=is%20DFNS%20live&limit=3" # → [reflector, soroswap, dia] ✗ (dfns absent)
```
Stop-words + strict-AND let common tokens outrank the branded entity. Query-parse / stop-word handling (route). Same fetch-not-rank family as the #565–567 brand-lookup fix, but for /projects/search.

P4 — SCF round membership overstated: 14 records (scfAwardedRounds — ingest, not a safe registry field)
boundless ours [#36 #40] vs official [#40]; aida [#38 #39] vs [#39]; stride [#32 #33] vs []; freelii [#32 #34] vs [#32]. Cross-check vs communityfund.stellar.org/project/. Our round-attribution over-claims; source-of-truth is the official per-project page. Needs an ingest/reconciliation pass (12 more "unverifiable" — review by hand, never auto-accuse).

P5 — zk-snark on /api/repos/search → 0 results (4 real hits)
```
curl -s "https://stellarlight.xyz/api/repos/search?q=zk-snark&limit=3" # → 0
```
Hyphenated/synonym tokenization in repo search (route).

P6 — attr-implied ranking (P-ATTR, 7): evm / evm infrastructure don't surface band, gami-labs, warpdrive in top-10 despite their own supportedNetworks implying evm. Ranking should let structured attributes drive inclusion (F1 structured-fields-not-inclusion root).

P7 — corpus hygiene (engine-c S5/S6/S8): S5 junk-URL regression 1 🔴 (must be 0), S6 bad titles 106 docs, S8 mirrored chunks 129. Ingest/title-extraction + dedup.

5. Patterns

• Structured-fields-not-inclusion (F1) shows up again in P1/P3/P6 — brand/attr/phrase queries where the right record exists and is even structurally tagged, but tokenization or strict-AND keeps it out of top-k. Candidate for a structural fix (a normalizer + attr-inclusion lane) rather than per-query rows. Watch across next runs.
• Builders-by-name (P2) is the single largest real-demand miss; if it recurs next run it justifies a builders-index ingest change, not a queue line.
• Precision guardrail held: the only audit red self-cleared, and I declined 3 low-value llama wires rather than write speculative rows.

[theboycoder]

Update (same day): P1 and P2 from the queue above moved queued → PR in a same-day session (owner-directed).

• P1 (StellarX tokenization) — fixed in fix(search): StellarX name-lookup miss + honest /api/builders scope for person queries #587. Root cause confirmed as the camelCase collapse to the stellar corpus-stopword + candidate-window flood; tokenize() now falls back to the joined identity form. Live-route verified across all casings + brand/topic regressions.
• P2 (builders person-lookup) — refined and fixed in fix(search): StellarX name-lookup miss + honest /api/builders scope for person queries #587, but NOT as originally framed. The right fix is not to index SDF team/ecosystem people (Justin Rice = VP Ecosystem; Tomer Weller; Tyler) as GitHub "builders" — they aren't builders in the Passport sense. Instead /api/builders now returns an honest scope-boundary advisory on person-name queries and routes them to /api/research. Nobody is misclassified. (Open follow-up: named SDF people aren't cleanly retrievable via /api/research either — the team-page ingest doesn't surface clean person cards. That's a real gap worth a future coverage item, distinct from the builders scope fix.)

P3–P7 remain queued as reported.

[theboycoder]

Follow-up built: the open P2 gap I flagged above — "named SDF people aren't cleanly retrievable" — is now addressed in #588 (stacked on #587). New /api/people SDF team/people index (leadership / board / advisors, name→role→org, from stellar.org/foundation/team with provenance) + get_people MCP tool + OpenAPI/changelog/skill-doc wiring; /api/builders cross-links a name lookup to the person's roster record. Committed registry (DB-independent), spec 1.7.29.

[theboycoder]

🔁 Run 2026-07-21

1. Since last run

The 07-19 baseline queued P1–P7. Verified against live API today:

• P1 StellarX camelCase → CLEARED. q=StellarX now returns stellarx at rank 1 (was absent). Fixed by the shared synonym registry (feat(search): shared synonym registry — one core vocabulary, three surfaces #618).
• P3 NL-phrase dilution → CLEARED. q=what is DIA now surfaces dia; q=is DFNS live surfaces dfns. Same feat(search): shared synonym registry — one core vocabulary, three surfaces #618 lane.
• P2 builders-by-name → PARTIAL. /api/people org roster added (fix(search): StellarX name-lookup miss + honest /api/builders scope for person queries #587/feat(people): SDF team/people index — /api/people + get_people MCP tool #588) but /api/builders?q=tyler / ?q=strupey still return 0, and Tyler/hoeven aren't on the org roster. Individual/login lookup still a real-demand miss.
• P4 SCF round membership → PERSISTS (2nd consecutive run). boundless [36,40] vs official [40]; stride [32,33] vs [].
• P5 zk-snark repo search → PERSISTS. /api/repos/search?q=zk-snark still 0 (feat(search): shared synonym registry — one core vocabulary, three surfaces #618 only covered projects, not repos).
• Coverage: llama-missing 3→0 self-cleared.

2. Engine states

Engine	State
self-audit (local)	🟢 63/63 (grew 54→63)
golden (local)	🟢 46/48, forbidden 0, bad-title 0, thin 2 (misses capability-custody-identity, latest-protocol-release)
retrieval health #436	🟢 recall floors held (P-PHRASE 80%, P-ATTR 76%)
coverage #562	0 llama-missing · 107 scf-missing · 0 stale collections
liveness	34 dual-signal death candidates (report-only)
tag-mismatch	19 strong (‼) type-add candidates
engine-c-health	weekly — not re-run since 2026-07-19; 250-probe queue stands

3. PR opened

#648 — data/improvement-loop-2026-07-21, paper-trail catch-up only (docs-only). improvements/ was 2 days / ~35 PRs stale (last at #604) while #605–#645 merged; adds improvements/waves/onchain-identity-vocab-2026-07-20-21.md (7 arcs: on-chain metrics v1/v2, partner↔project identity, synonym registry, spec/guards, PG-award structured truth, i³ awards voting). No new lessons row — all durable classes already guarded.
No data-registry row this run: 34 liveness + 19 tag candidates held for owner review (per-project primary-source verification pending; precision guardrail). Consistent with #585.

4. Queued code-class (priority)

• P4 SCF round membership overstated — 2 runs → structural. boundless [36,40]→[40], stride [32,33]→[]. Needs a scfAwardedRounds reconciliation pass vs communityfund.stellar.org/project/<slug> + cross-check guard, not one-off rows.
• P5 zk-snark repo search → 0. curl "https://stellarlight.xyz/api/repos/search?q=zk-snark" → 0 (4 real hits). Extend synonym/hyphen tokenization to the repos surface.
• P2 builders-by-name → 0. curl "https://stellarlight.xyz/api/builders?q=tyler" → 0; ?q=strupey → 0. Individual/login lookup gap; /api/people covers only the SDF org roster.
• P6/P7 (engine-c) carry forward — weekly hasn't re-run.

5. Patterns

• F1 structural fix validated: the feat(search): shared synonym registry — one core vocabulary, three surfaces #618 synonym registry cleared multiple queued brand/phrase misses (P1, P3) at once. Keep investing in the shared-vocabulary lane over per-query rows — and extend it to the repos surface (closes P5).
• P4 SCF-rounds now recurring (2 runs) → promote from queue line to a scfAwardedRounds reconciliation ingester with a cross-check guard.
• Paper-trail went stale again during a heavy shipping window — data(improvement-loop): paper-trail catch-up 2026-07-21 (#605–#645) #648 is the catch-up; the recurring lag itself argues for a lighter per-day append discipline.

[theboycoder]

P5 update — addressed in #653.

Root cause was more specific than "repos surface lacks the registry" — the registry (#618) already reaches repo search; the zk family was asymmetric. zk → [snark, stark, groth16, …] existed, but snark, zksnark, and zk-snark were only values under zk, never keys, so a query FOR a family member never expanded back to the family:

q=zk        → 5 hits
q=zk-snark  → 0   (expanded only to [zk-snark, zksnark])
q=snark     → 0
q=zksnark   → 0

Fix: one shared ZK_FAMILY set in CORE_SYNONYMS, referenced by every entry point (zk/zkp/snark/stark/plonk/groth16/zksnark/zk-snark) → q=zk-snark ≡ q=zk. Loose proof/circuit stay repo-overlay-only (vocab test guards it). Verified: expansion now yields the full family for all entry points, reaching the zk/groth16 topics the live zk-SNARK repos carry (nethermindeth/stellar-private-payments, stellar-risc0-verifier); DB-match proven transitively via live q=zk. Regression test added; typecheck + biome + 147 lib tests green. Awaits merge + deploy to clear on live.

Queue status: P5 → in-review (#653). P4 (SCF rounds) and P2 (builders-by-name) remain open.

[theboycoder]

P4 update — structural guard shipped in #659; data fix is a human Action dispatch.

Investigated the recurrence. The reconciliation machinery already exists and is correct — scf-crosscheck.ts (detector) + scripts/data/fix-scf-rounds.ts / fix-scf-rounds.yml (fix). The ingest root cause was fixed 2026-07-11 (enrich now reads per-submission verdicts, not badge arrays). P4 persists purely because (a) the fix hasn't been executed on prod and (b) engine-c ran the crosscheck weekly as continue-on-error evidence nobody was paged on — so it silently rode two runs.

Current live state — 14 records, precision-verified (ran the crosscheck against live; spot-checked 3 directly against the official per-submission verdicts):

slug	ours	official awarded	offending round
boundless	[36,40]	[40]	#36 Prescreen Failed
stride	[32,33]	[]	#32 Prescreen Failed, #33 Draft
aida	[38,39]	[39]	#38
freelii	[32,34]	[32]	#34
token-tails	[26,30,41,42]	[26,30]	#41 #42
stellar-metamask	[19,32]	[19]	#32
swift-wallet-sdk	[30,31,36]	[31,36]	#30
easya	[17,23,30]	[17,23]	#30
noether	[41,44]	[41]	#44 Prescreen Failed
rendbit	[36,39]	[39]	#36
minah	[37,38,40]	[38]	#37 #40
alula	[33,34]	[34]	#33
untangled	[30,31]	[31]	#30
via-labs	[35,40]	[35]	#40

Shipped (#659): a drift guard that rides engine-c (reuses the /tmp/scf.json it already produces — no new workflow, no extra scrape/minutes) and promotes roundsOverstated > 0 to a self-closing "SCF round-membership drift" rolling issue via the #591 composite. So this can't silently persist to a 3rd run — the next engine-c run will open the tracked issue; running the fix auto-closes it.

Recommended human step to actually clear the 14 (I can't dispatch workflows or write prod): dispatch fix-scf-rounds.yml — dry-run first (inspect scf-rounds-plan.json), then --execute. It exact-syncs scf.awardedRounds to the official verdicts and handles stride (zero-awarded) via the boolean-flip allowlist. Then re-run the crosscheck to confirm 0.

Queue status: P4 → guard shipped (#659); data-fix awaiting fix-scf-rounds.yml dispatch. P2 (builders-by-name) still open.

[theboycoder]

P2 update — code-derived builder lookup shipped in #661.

Investigated the "builders-by-name → 0" gap. It's narrower than it first looked — /api/builders already routes multi-token person names well (justin rice → "Justin Rice — VP of Ecosystem (SDF Leadership)" + /api/people cross-link). The real miss is GitHub logins we have code evidence for but no Passport profile:

• kalepail owns 4 indexed Stellar repos (passkey-kit, kale-sc, smart-account-kit, stellarauth-serverless) → /api/builders?q=kalepail returned a bare 0 with a generic "broaden your filter" advisory (a single-token login doesn't trip the person-lookup steer).
• strupey → 0 repos in our index (no code evidence), so correctly nothing to surface — the "17 hits" wasn't repo ownership.

Shipped (#661): a precise code-derived fallback — only when Passport matched none, the query is handle-shaped (one login-charset token, not a skill term), and it exactly matches an indexed repo owner, we synthesize one builder row from repo ownership (match.basis:"repo-owner", bio/roleTitle null, evidence in codeEvidence). q=rust/q=soroban stay skill searches; person names still route to /api/people. Contract bumped to openapi@1.8.16 (basis enum + changelog). Gate + synthesis extracted as pure fns and unit-tested (5 tests); typecheck + biome + 458 lib tests green. Live path confirms post-deploy (q=kalepail → basis repo-owner).

Queue status: P2 → shipped (#661). With P1/P3 cleared earlier (#618), P5 in-review (#653), P4 guarded (#659), the improvement-loop queue from the 07-19 baseline is now cleared or in-flight.

[theboycoder]

📋 Queue status — consolidated (2026-07-21, updated)

The 07-19 baseline queue (P1–P7) after this session's work:

#	Item	Status	Where
P1	StellarX camelCase tokenization	✅ cleared (live)	#618 synonym registry
P3	NL-phrase dilution (what is DIA)	✅ cleared (live)	#618 synonym registry
P5	zk-snark repo search → 0	🟡 in review	#653 — bidirectional zk family
P2	builders-by-name (kalepail → 0)	🟡 in review	#661 — repo-owner code-derived rows
P4	SCF round membership overstated (14)	🟡 guarded; data-fix pending human	#659 guard shipped · dispatch fix-scf-rounds.yml to clear the 14
P6	attr-implied ranking (P-ATTR)	⚪ carry-forward	engine-c weekly (not re-run since 07-19)
P7	corpus hygiene (S5/S6/S8)	⚪ carry-forward	engine-c weekly; S6 titles already 11→0 via #624

Open PRs from this session (all awaiting human merge):

• #648 — paper-trail wave (Status re-verification queue (weekly) #605–fix(guards): raven-drift grace window — catalog lag is cadence, not drift #645), docs-only
• #653 — P5 zk family
• #659 — P4 drift guard
• #661 — P2 builders repo-owner lookup

Only remaining human actions: merge the 4 PRs; then dispatch fix-scf-rounds.yml (dry-run → execute) to clear P4's 14 records. P6/P7 re-evaluate on the next weekly engine-c run — and P4's new guard will open/auto-close a rolling issue from that run onward.

[theboycoder]

P6 update — addressed in #667.

P-ATTR: chain queries didn't let structured supportedNetworks drive ranking.

q=evm            → band #19, gami-labs & warpdrive past #50
q=evm infrastructure → warpdrive #6, gami-labs #7, band #9   ✅

All three carry supportedNetworks=[stellar, evm, …] but no "evm" in prose. The queries diverge because q="evm infrastructure" has a token (infrastructure) that hits the identity zone via category, flipping anchorIdentity true — bare q=evm had none.

Root cause (F1): identityZone() — the "what the record IS" signal anchorIdentityHit keys the top-tier sort on — folded in types and coverage but not supportedNetworks, so a chain token could only reach identity through prose.

Fix: fold supportedNetworks into identityZone (exactly as types/coverage already are). Now q=evm hits identity off the structured field → the genuinely-evm projects join the top tier with the evm bridges (ordered within it by score → authority) instead of ranking below every prose-mentioner. Chain values are specific tokens, so it fires only for chain-named queries. Ranking-only — no OpenAPI/contract change (so no version-bump conflict thrash). Regression test added; 440 lib tests green.

Queue status: P6 → in-review (#667). Only P7 (engine-c corpus hygiene) remains, carry-forward to the next weekly engine-c run.

[theboycoder]

P7 update — html-entity slice addressed in #669.

Corpus-health S6 = 113 bad-title docs, broken down: 47 html-entities, 56 sentence-like, 9 overlong, 1 bare-date. (S5=1 junk URL, S8=130 mirrored chunks.)

Fixed the cleanest deterministic slice — the 47 html-entity titles (raw '/& the ingesters never decoded):

• src/lib/decode-entities.ts — shared decoder (decimal/hex/named; unknown refs left verbatim), unit-tested.
• ingest-sdf-blog.ts — cleanTitle decodes at write time (stops regression).
• scripts/fix-research-entity-titles.ts — dry-run/--apply back-fill; local-corpus dry-run: 122 rows / 33 distinct titles decode cleanly. Human runs --apply via Action to drop S6 by 47.

Remaining P7 (follow-ups, not #669):

• S6 sentence-like (56) / overlong (9) — mis-extraction; needs per-source title-selection heuristics (the fix-dev-docs-titles RETITLE pattern), not entity decode.
• S8 mirrored chunks (130) — dedup pass (data cleanup / Action).
• S5 junk URL (1) — single prune row.

---

Queue status — every actionable P-item now has an open, CI-green PR

#	Status	PR
P1, P3	✅ live	#618
P2	🟡 in-review	#661
P4	🟡 guard in-review + data-fix pending dispatch	#659
P5	🟡 in-review	#653
P6	🟡 in-review	#667
P7 (S6 entities)	🟡 in-review	#669

Plus paper-trail #648. All six MERGEABLE. Then: dispatch fix-scf-rounds.yml (P4 data) + fix-research-entity-titles.ts (P7 data).

[theboycoder]

✅ All queued P-items merged to main (2026-07-21)

Every actionable item from this triage is now shipped:

#	Item	PR (merged)
P1, P3	brand/NL-phrase tokenization	(already live, #618)
P2	builders-by-name → repo-owner code-derived rows	#661
P4	SCF round-membership drift guard	#659
P5	bidirectional zk synonym family	#653
P6	structured supportedNetworks = identity	#667
P7	research title entity-decode (S6 slice)	#669
—	paper-trail wave #605–#645	#648

Spec now at openapi@1.8.22.

Two data-side steps remain (owner-run Actions — code is in):

1. P4: dispatch fix-scf-rounds.yml (dry-run → execute) to clear the 14 overstated round-membership records. The new drift guard (feat(guards): SCF round-membership drift guard — closes the P4 recurrence loop #659) will then keep it green.
2. P7: run scripts/fix-research-entity-titles.ts --apply (via Action) to decode the ~47 existing entity titles. The ingester (fix(corpus): decode HTML entities in research titles — P7 (S6 hygiene) #669) keeps new ones clean.

Still open (P7 follow-ups, not blocking): S6 sentence-like (56) / overlong (9) mis-extraction, S8 mirrored chunks (130), S5 junk URL (1) — per-source heuristics / dedup, tracked for a future run.

Watch the Vercel deploy on the merge wave; re-run scripts/self-audit.ts after deploy to confirm green.

[theboycoder]

P4 data fix — DONE. Dispatched fix-scf-rounds.yml dry-run → execute+surgical.

• 14/14 writes applied (13 exact-sync + 1 surgical stride [32,33]→[33], dropping the proven-negative Create Hackathon events collection and calendar page #32, keeping the unverdicted Add post-hackathon project status tracking #33). 0 failed.
• Live-verified: boundless [40], stride [33], noether [41], via-labs [35], token-tails [26,30] — all reflect on the live API immediately (projects endpoint reads the DB directly).
• Crosscheck re-run: roundsOverstated 14 → 0, overstated 0. The 11 roundsUnverifiable are ambiguous/in-flight rounds correctly left for hand review (never accused).

P4 fully closed: drift guard (#659, merged) keeps it at 0 going forward.

Only P7's data step remains: run scripts/fix-research-entity-titles.ts --apply to decode the ~47 existing entity titles (dry-run verified 122 rows / 33 titles).

[theboycoder]

P7 data step — DONE. Added a dispatchable workflow (fix-research-titles.yml, #671) and ran the apply against prod.

• 152/152 titles written, 0 failed (306 rows scanned, 152 carried decodable entities). Foundation's 2023 Strategy → Foundation's 2023 Strategy, On & Off Ramps → On & Off Ramps, "…" → "…".
• Live-verified: research-title queries no longer surface any &#/&/" — the S6 html-entity class is cleared. The ingester (fix(corpus): decode HTML entities in research titles — P7 (S6 hygiene) #669) keeps new titles clean.

---

🏁 Queue closed

All seven triage items shipped and both data steps executed:

#	Item	Code	Data
P1, P3	tokenization	✅ #618	—
P2	builders-by-name	✅ #661	—
P4	SCF round membership	✅ #659 (guard)	✅ 14 records corrected, roundsOverstated 14→0
P5	zk-snark repos	✅ #653	—
P6	attr-implied ranking	✅ #667	—
P7	corpus S6 entities	✅ #669 + #671	✅ 152 titles decoded

Remaining (non-blocking, future runs): P7's S6 sentence-like (56) / overlong (9) mis-extraction, S8 mirrored chunks (130), S5 junk URL (1) — per-source heuristics / dedup.

[theboycoder]

🔁 Improvement loop — 2026-07-23

1. Since last run (07-21 close-out)

The 07-21 comment closed P1–P7 with both data steps executed. Re-probed live today — the fixes hold:

item	probe today	verdict
P2 builders-by-name (#661/#690/#691)	/api/builders?q=tyler van der hoeven → 1 row, displayName: "Tyler van der Hoeven"	✅ cleared
P5 zk-snark repos (#653)	/api/repos/search?q=zk-snark → total 78, top-3 nethermindeth/stellar-private-payments, Shreshtthh/Zk-Seep, nethermindeth/stellar-risc0-verifier	✅ cleared
P7 html-entity titles (#669/#671)	golden BAD-TITLE 0	✅ cleared
P4 SCF rounds	roundsOverstated 14 → 0 (07-21 apply)	✅ cleared

Did NOT clear (both carried into the queue below): the P-ATTR q=evm probe (still no band/warpdrive/gami-labs in top-10 after #667) and strupey on /api/builders (still 0, 17 real hits).

2. Engine states

• self-audit (local, today): 61 passed / 0 warnings / 2 failures — and both failures are false. @stellar-light/api-client@1.5.0 and 1.5.1 were reported "missing"; the registry returns 200 for both on re-probe (1.4.0, checked seconds later in the same run, passed). The check treats any non-ok status as absent, so a transient npm 504 reads as an uninstallable published artifact. Effectively green. → queued as Q1.
• golden (local, today): 48/48 pass, forbidden-hits 0, JUNK 0, BAD-TITLE 0, THIN 3.
• retrieval health (Retrieval health — weekly tracker #436): 🟢 today's post-deploy run — all 8 recall floors held, Engine E no silent params.
• engine-c-health: last run 07-21 (not re-dispatched — 0 Actions minutes spent this run). All 8 Engine A buckets above floor; 226 probe failures, 217 of them one bucket (P-PHRASE "is X live"). Engine B: 1 domain-keyed dupe (fixed below), 2 prose⇄structure divergences, 385 Live rows with no dated signal.
• status queue (Status re-verification queue (weekly) #605): 500/500 Live rows carry statusAsOf, median age 133d, 485 with no statusSourceUrl. Top-10 worked this run.
• liveness (Liveness watch — monthly review queue #465) / coverage (Coverage watch — ecosystem gap report #562): no new comments since last run — no delta.

3. PR opened — #699 (not merged)

data/improvement-loop-2026-07-23 · registry rows only, no code, no status flips.

row	evidence
DUPE_MERGES boss-revolution → boss-pay	Both store bossmoney.com. Leaf's own site: "Leaf's app is now BOSS Money"; stellar.org's SDP post names "IDT's Boss Money" and "also available for non-smartphone users" — verbatim the dupe row's prose. BOSS Revolution is IDT's separate calling/top-up brand with zero Stellar signal. Class 10.
DESCRIPTION_FIXES flipside	All flipsidecrypto.* marketing URLs 200-redirect to edisyl.com; its patent block reads "Assignee: Flipside Crypto Inc (edisyl)", blockchain work in past tense. Description only — status left Live, since a redirect to a moved site is not evidence of death. Owner call queued on #605.
STATUS_FIX ×7 Live→Live	nodies, warmancer, scorechain, spiko, ultra-stellar, katagames, stellar-passport — each re-fetched by hand today, verbatim <title> quoted in the row. Converts source-inherited + no URL → dated site-liveness + URL. wisdomtree excluded (403 behind Cloudflare = not observable either way).

Paper trail (improvements/ had not recorded #646–#696): new wave ledger-and-raven-loop-2026-07-22-23.md, plus lessons class 31 — build-only breakage (route.ts helper export passed tsc + 514 vitest tests, failed next build, froze deploys; guards #675 + #676).

Verified locally: biome clean, typecheck clean, vitest 514 passed (1 pre-existing local DB failure), golden 48/48. After merge: curate-projects.yml dry-run → inspect → execute → re-run self-audit.ts.

4. Queued — code class (not touched this run)

Q1 — self-audit reads a 5xx as "version missing" (highest priority: it makes the daily green/red signal lie).

curl -s -o /dev/null -w "%{http_code}\n" https://registry.npmjs.org/@stellar-light/api-client/1.5.0

Expected 200 (it is). Today's audit logged "the npm registry returns 504 — published changelog must never advertise an uninstallable version" for 1.5.0 and 1.5.1. scripts/self-audit.ts:658-666 does if (r.ok) ok() else bad(). Fix: retry once, then split 404 → bad (genuinely unpublished) from 5xx/network → warn (registry unavailable ≠ artifact absent).

Q2 — P-ATTR q=evm still short after #667.

curl -s "https://stellarlight.xyz/api/projects/search?q=evm&limit=10"

Expected (per the engine-c probe): band, warpdrive, gami-labs in top-10 via their own supportedNetworks. Actual: allbridge, axelar, circle-cctp…, stronghold, rozo, bim-exchange, via-labs, usdc-swap, tricorn, transfuse (total 91). Note the actual top-10 is defensible — those are real EVM-bridging projects. So this needs a decision, not a reflex fix: either identity-zone supportedNetworks isn't reaching bare single-token queries, or the probe's expectation is wrong and should assert presence in top-N-of-91 rather than top-10.

Q3 — /api/projects accepts and silently ignores unknown filters.

curl -s "https://stellarlight.xyz/api/projects?slug=boss-pay" | head -c 200

Expected: the boss-pay row, or a 400. Actual: the unfiltered first page (returned "Stellar Flutter SDK"). A consumer reading that concludes the wrong record is boss-pay. Same class as Engine E's invalid-accepted 1, and the read-side twin of the silent-drop write class.

Q4 — identity-less demand queries answer 0 with no honesty signal. strupey (17 hits) returns 0 on /api/builders and 0 on /api/repos/search; github.com/strupey is a 404, so there may be nothing to find. Same shape: saleem, kutana, sorted fall back to matchMode=semantic on /api/projects/search. Needs a "no such builder — here's why" answer, not an empty array.

Q5 — Engine B doesn't name its rows. The health report says "2 records name a chain/country/SEP their structured fields lack" without printing the slugs, so the safe data class cannot act on it. One-line fix in the reporter; unblocks a SUPPORTED_NETWORKS row next run.

Q6 — corpus hygiene carry-forward (unaddressed since 07-21): S6 sentence-like titles 56 · overlong 9 · S8 mirrored chunks 130 · S5 junk URL 1.

5. Patterns

• Detectors that cry wolf are now the recurring class — 2 runs in a row. fix(quality): engine-e optional-absent is spec-compliant — close contract wave (0%→3% closing rate) #686 last run (engine-e flagged spec-compliant optional-absent as drift, closing rate stuck at 0%); Q1 this run (self-audit calls a live npm version missing on a 504). Both make a green/red signal untrustworthy, which is worse than a missed finding because it trains everyone to skim past the red. Candidate structural fix: a shared probeExternal() helper for every detector that touches a third party — retry once, and classify 404/NXDOMAIN → finding vs 5xx/timeout/403 → unverifiable, with unverifiable counted in its own column that never fails the gate. This run alone hit it three times (npm 504, wisdomtreeprime.com 403, app.flipsidecrypto.com 403) — I had to make the same judgment call by hand each time.
• P-PHRASE is 217 of 226 probe failures and has been the dominant bucket for two runs while sitting above its floor (75.8%, floor 60%). The floor is hiding the ecosystem's single biggest miss shape ("is X live"). Worth either raising the floor or splitting the bucket so it can't hide.
• Status provenance is a population problem, not a top-10 problem. This run fixed 7 rows by hand; 478 remain unsourced. Hand-verification will never catch up — the durable fix is an automated liveness prober writing site-liveness provenance on a cadence, with human review reserved for anything that would demote.

[theboycoder]

Q1 — fixed in #701.

self-audit.ts collapsed every non-ok registry response into bad("advertised version missing"), so today's transient npm 504 on api-client@1.5.0/@1.5.1 reported two published, installable versions as absent — while 1.4.0, probed seconds later in the same loop, passed.

Now retries once, then classifies: 404/410 → bad (genuinely unpublished — the class-12 finding this check exists for, still gating), 5xx/429/network → warn (unverifiable; an upstream outage can't fail our gate or file a false daily issue). All four branches exercised against real endpoints (live 200, real 404, local 503 server, dead socket).

Self-audit is now 63 passed / 0 warnings / 0 failures, exit 0 — genuinely green, where this morning's 61/0/2 was green-with-two-lies.

Remaining queue is unchanged: Q2 (q=evm ranking-vs-probe decision), Q3 (/api/projects silently ignores unknown filters), Q4 (identity-less demand queries), Q5 (Engine B doesn't print the slugs it flags), Q6 (corpus hygiene carry-forward). The structural probeExternal() from pattern 1 is still open — #701 fixes the one instance, not the class.