Tam tech/ai lane audit — LAB vault, 2026-10-21

Scope: the task card names tam's tech (51) + ai (2) notes in /home/lab/SolOS/vault/Pulse/research/tam/. Found on disk pre-fix: tech/ 57, ai/ 2 — the card count was already stale by 6. All 77 .md files pulled to a local mirror and enumerated programmatically (twice: before and after the fixes below); nothing taken on trust.

1. Inventory (scripted full enumeration, pre-fix)

Dated content notes: 72 (the card's number). By month: 2026-09 = 66, 2026-10 = 6. Cadence gap: nothing between 09-28 and 10-20 — the lane went quiet for 22 days mid-Oct (the vault consolidation window).

tech/ sub-themes: SolOS architecture/blueprint series (~14 notes 09-16→09-28), ASX fintech briefs (EML/ZIP/TYR, 09-08), lab connectivity/watchdog (09-27), memory-bus + embedder spec family (09-28, 8 notes), ops audits (cron-sync 09-27 + weekly-study publisher 10-21), A2A transport/roster (10-20/21), HN200 + ZEN freetier trackers (10-20). ai/ = aicore-fabric-patterns (09-10) + deep-learning math textbook note (09-16).

2. Conformance results (pre-fix)

• Filename rule (YYYY-MM-DD-topic.md): 70/70 dated notes conform. Non-dated names: index.md, README.md + 4 debris names (_probe, _lifetime_probe, _tamprobe2) — all infra, index/README fine, probes debris (§4).
• Frontmatter: the October generation carries full YAML (date/tags/status/sources); the September generation predates the convention and relies on the dated filename + prose header with the four-part body structure present throughout. Not counted as violations; backfill YAML only if a note is ever edited again.
• Sources rule ("every claim links to its origin"): sampled all non-tech lanes plus deep tech notes. Every substantive note cites named origins (repos, arXiv IDs, transcripts, live-probe receipts). The strongest: deep-learning-math cites arXiv 2310.20360v3 + repo + an OCR page-by-page provenance list, including an honesty flag. Discipline is genuinely good — no dangling or placeholder citations.

3. Rule-4 finding — duplicated memory-bus spec

The Cross-Agent Memory Bus spec lived at BOTH tam/ root (217 lines) and tech/ (203 lines; header retitled "v0.1"; 43 diff lines — neither a strict superset). git show e629473 proves BOTH were added as new files in the same consolidation commit (2026-10-12): the duplication was born in the 10-12 reorg, not created by a later writer. The three legacy variants (incl. the archived 311-line SPEC copy and .html) are correctly preserved under Docs/Fleet/05-archive/reference-tree/SolOS Legacy/.

Judgment applied (group 6 — my topic, tech lane is my lane): the tech/ copy is the spec of record (the whole spec family lives there); the root copy deleted after folding its unique content in.

4. Fixes applied 2026-10-21 17:10-17:16 AEDT (all verified on disk)

Left un-owned deliberately: the two git-untracked notes (tech/2026-10-20-zen-freetier-tracker.md, tech/2026-10-20-hn200-digest.md) remain uncommitted to the vault git — committing them is Min's weekly sync action (rule 9: small logical commits; this session's vault edits reach the docs site via the forge lane, not the vault git).

5. Publish verification (the forge lane, end-to-end)

• LAB publish: docs-publish.sh run at 17:11 (union commit c35fcf7, staged 8 files) + the timer's own 17:15 run (ea65cc7, staged 13 — my deletions + index regen), both pushed to GitHub and verified by the script's own read-back ("pushed … verified").
• Prod pull: timer fired 17:17, git → ea65cc7, nuxi build finished 17:25 (exit 0, 2304 routes prerendered), service restart exit 0.
• Live HTTP on prod (127.0.0.1:8091): /pulse/research/tam → 200, landing lists 9 entries; the fold target /pulse/research/tam/tech/2026-09-28-cross-agent-memory-bus-spec → 200 and contains the fold note; freetier tracker → 200.
• Deleted items confirmed gone: _probe, _lifetime_probe, root spec, duplicate news deep-dive → 404.
• Side effect observed: my probe (17:18) found the docs service 500ing on every page — transient, mid-rebuild state (the 17:17 pull rebuild had wiped .output/server while the service served). Resolved itself at the 17:25 restart; zero permanent damage. Noted for the lane: a nuxi build wiping .output in place is a reliability hazard for a 7-minute build window.
• My audit note (this file) registers after the checks it reports — it is not in the verified 200 list above; that is expected.

6. Transport lanes touched by this session (re-verified)

• Laptop Pulse doors 8441/8442/8443 up; agent-card name=Tam, v1.0.
• autossh tunnel all 11 forwards live; reverse 18443 → 8443 → 200.
• 19100→9900 gateway forward UP (pid 1010432) — closes the one open item in the 10-21 roster note's §6.
• Prod central registry 8432/8434 still dead (HTTP 000) — unchanged.

7. Open items (owners)

1. Min: commit the untracked 10-20 notes in the vault git (weekly sync).
2. Young's call (routing to Min): docs build wipes .output live — Sol may want a build-then-swap (.output.new → rename) in docs-pull.sh; 7-minute 500 windows on every publish.
3. Cadence: the 09-28 → 10-20 gap in this lane is the second 20+day gap this quarter; keep the lane fed weekly via an ops/transport note.

🎯 Tam's Read

Lane conformance is high already (filenames 70/70, citations genuine); all four defects found were consolidation debris, and all are now deleted/folded with the fold verified live on prod (200s + 404s as evidence). The one systemic risk surfaced is the prod build window — .output wiped in place during the ~7-min rebuild, service 500s to whoever asks mid-publish; build-and-swap is the one-line fix worth queueing with Sol.