/cross-check phase 2 — semantic edge annotation (unblock for phase 3)

Context. Notion ticket "Revamp /cross-check to use graph queries (phase 3)" (page 342f7d49-36d1-81ea-98b4-c8d29c545c4e) was stuck because phase 3 (the graph-traversal rewrite of /cross-check) depends on phase 2 — typed semantic edges (validates / contradicts / supports-position) annotated onto synthesis docs — and phase 2 had never actually been executed, despite the annotator script existing.

What was actually found

The curation work for phase 2 was already done — ~/.claude/scripts/graph-annotate.py contains a fully hand-curated seed set: 6 Cluster definitions (including harness-thesis-dissent, the ticket's named unblock cluster, and agent-architecture, its 13-doc superset), 8 named Positions, 22 validates edges, 6 contradicts edges, and 29 supports-position edges, each with a one-line evidence quote.

It had never been run against the live graph. ~/.claude/state/graph.duckdb had zero Cluster vertices and zero semantic edges prior to today — graph-query cluster-members harness-thesis-dissent returned "no Cluster vertex found." The graph-query SKILL.md's claim that "the annotator also runs automatically at the end of graph-ingest.py" is inaccurate — graph-ingest.py contains no reference to graph-annotate.py or any --no-annotate flag. This should be corrected in the skill doc as a follow-up (not done today — out of scope, flagging only).

Bugs found and fixed in graph-annotate.py

Running --dry-run surfaced 7 unresolvable doc stems (out of ~45 unique docs referenced). All 7 were filename-convention mismatches — the annotator was written against a YYYY-MM-DD-slug stem pattern, but the actual files use a book-slug-YYYY-MM-DD or synthesis-slug-YYYY-MM-DD or internal-review-slug-YYYY-MM-DD convention:

All 8 were corrected in place in graph-annotate.py (find/replace on the stem string, 26 occurrences across CLUSTERS/VALIDATES/CONTRADICTS/SUPPORTS_POSITION). Post-fix --dry-run resolves cleanly with zero unresolved stems.

Annotation run

Ran for real (not dry-run):

~/.claude/scripts/graph-db-venv/bin/python ~/.claude/scripts/graph-annotate.py

Result: 6 Cluster vertices, 8 Position vertices, 39 part-of-cluster edges, 22 validates, 6 contradicts, 29 supports-position edges written to ~/.claude/state/graph.duckdb, all tagged source='annotate' (idempotent re-run).

Verified via /graph-query cluster-members harness-thesis-dissent: now returns 4 real member docs (was: cluster not found). /graph-query author-authority now returns a populated authority table (was: presumably empty/error before, table structure existed but no edges to score).

harness-thesis-dissent cluster — the ticket's named unblock criterion

4 docs (the ticket's task notes said "at least 12" for the full cluster and estimated "~15 total" for the whole seed set — see caveat below):

• 2026-03-25-seattle-data-guy-know-nothing-and-be-happy.md — comprehension-atrophy critique
• 2026-03-31-semistructured-data-layer-does-the-work.md — Natkins, data-layer-is-the-moat
• 2026-04-10-jaya-gupta-anthropic-moat.md — trust/permission as the scarce asset
• synthesis-harness-thesis-dissent-2026-04-12.md — the anchor dissent synthesis doc itself

Caveat on the "12 docs" / "~15 total" estimate in the ticket notes: the actual hand-curated seed set covers 6 clusters totaling ~45 unique doc references, but the harness-thesis-dissent cluster specifically has only 4 members as curated. The closely-related agent-architecture cluster (13 members, anchored by the 2026-04-12-cross-check-agent-architecture cross-check doc) is the one that's closer to "12 docs." It's likely the ticket's "12 docs" note was describing agent-architecture or the combined harness-thesis + agent-architecture superset, not harness-thesis-dissent alone. Flagging rather than silently reconciling — a human should confirm which cluster the original ticket note meant before treating either as the definitive "phase 2 done" signal.

Known gap discovered (not fixed — flagging for follow-up)

2026-04-12-cross-check-agent-architecture (the agent-architecture cluster's anchor doc, at 06-reference/cross-checks/2026-04-12-cross-check-agent-architecture.md) is not ingested into the graph at all — it resolves as a stub (path IS NULL) even after today's re-ingest. The file exists on disk with correct frontmatter, isn't newer than the last reingest, and no exclusion for cross-checks/ appears in graph-ingest.py or the graph-reingest SKILL.md's find command. Root cause not diagnosed — recommend a follow-up check of the last graph-ingest.py --files run's input list.

Similarly, 01-projects/phdata/2026-04-15-career-moat-synthesis.md (referenced by a validates edge in the phdata-vs-mg-decision cluster) resolves as a stub. The file exists, has correct frontmatter, and 42 other 01-projects/phdata/ docs ARE in the graph — so this looks like an isolated ingestion miss rather than a directory-level exclusion. Same "not diagnosed, flagging" status.

Neither gap blocks the annotation edges from existing (stub vertices still carry the edge for citation-tree purposes per the graph-query SKILL.md's documented stub behavior), but both mean path-dependent queries (open-the-file-from-result) will fail for these two docs until a future re-ingest picks them up.

Spot-check (5 edges, re-read source docs directly)

1. contradicts: synthesis-harness-thesis-dissent-2026-04-12 → 2026-04-11-garry-tan-thin-harness-fat-skills — CONFIRMED. Dissent doc text: "the harness matters but is necessary-not-sufficient; data is the durable moat... Garry Tan's framework is correct as architecture guidance but incomplete as moat theory."
2. contradicts: 2026-03-25-seattle-data-guy-know-nothing-and-be-happy → 2026-04-11-garry-tan-thin-harness-fat-skills — CONFIRMED. SDG's 2030 vignette ("did anyone check it? No... you will know nothing and be happy") is a direct comprehension-atrophy counter to Tan's "thin harness is trustworthy because skills encode judgment" framing.
3. validates: book-adwd-master-synthesis-2026-04-13 → 2026-03-30-founder-data-quality-framework — CONFIRMED. Doc text explicitly: "Corr's central methodological claim: profile source data BEFORE designing target schemas... his framing of profiling-as-testing validates the approach."
4. supports-position: 2026-03-31-semistructured-data-layer-does-the-work → data-is-the-moat — PLAUSIBLE, not fully line-verified (doc title alone strongly implies the position; didn't locate the exact "moat" sentence in a quick grep — recommend a closer read before treating this one as airtight).
5. validates: 2026-04-15-career-moat-synthesis → 2026-04-11-phdata-vs-mg-decision-analysis — CONFIRMED on content (doc explicitly frames itself as distilling Chin's 5 Career Moats articles "into a single decision framework for the phData counter-offer," i.e., feeding the structural decision analysis) — but see the stub-vertex gap above; the edge exists in the graph pointing at a stub, not the real Document vertex, until that ingestion gap is fixed.

Remaining work (phase 3, explicitly NOT attempted here)

The actual /cross-check rewrite to use graph-query-style traversal instead of a linear vault scan is untouched. Phase 2's blocking condition — semantic edges existing in the graph for at least the harness-thesis-dissent cluster — is now satisfied (4/4 members annotated, verified queryable). Two ingestion gaps (cross-check-agent-architecture stub, career-moat-synthesis stub) should be resolved before phase 3 leans on those specific docs, but they don't block starting the rewrite for the harness-thesis-dissent cluster itself.