Flywheel Stage-3 gate reading due 2026-07-17

[bejranonda]

The flywheel-repair spec gates Stage 3 on the loop-health panel over a rolling 2 weeks: ≥60% of sessions closed-with-learnings AND ≥40% injection→used. The used-rate signal accrues from v1.13.0 (2026-07-03), so 2026-07-17 is the first honest full window.

Baseline (2026-07-07, read-only SQL against prod):

• 14-day window: 17 opened / 16 closed / 10 with learnings (59%) / used 4 (24%, polluted by pre-v1.13.0 sessions)
• Honest window since v1.13.0: 4 opened / 3 with learnings (75%) / 4 injected / 3 used (75%) — both metrics above gate, small n
• Duplicate project identities: 0 · Knowledge: 152 active, 58 with outcomes (38%)

How to read the gate on/after 2026-07-17: dashboard → Show everything → Loop health, or the read-only SQL in this issue's provenance (any agent session in the BrainPlatform checkout can run it — ask it to 'read the Stage-3 gate').

If the gate passes: (1) shrink MEMORY.md/file memory to bootstrap stubs (spec §5), (2) run the generation-uplift benchmark #126, (3) then pool widening #146. If it fails: diagnose which half starved — do not proceed anyway (spec §5).

Baseline note: the Brain-first protocol + Stop-hook were installed in Dementia-Daily / Nutri-Vision-AI / Start-Idea on 2026-07-07; sessions in those repos feed the same gate.

[bejranonda]

V2.0 shipped dark (2026-07-07/08, v2.0.0): meeting/doc intelligence landed via #151/#152/#155/#154 and is deployed, but both features are behind default-off flags (V2_ACTION_ITEMS, V2_ORACLE_TASKS). Per the amended spec §6, this gate governs flag enablement: read the panel on/after 2026-07-17 — pass → flip the flags (add both ="true" to the prod .env + ./scripts/reload.sh mcp-server web); fail → flags stay off and flywheel diagnosis takes priority. The teach path (meeting-miner protocol, docs/protocols/) works regardless of the flags and can be dry-run now.

🤖 Generated with Claude Code

[bejranonda]

Spec §6.1 dry-run executed against the live server (2026-07-08, v2.0.4, flags dark) — PASS on all six contract points:

Fictional meeting mined into isolated project V2-DryRun-Fixture via the meeting-miner protocol (real MCP calls against prod):

1. ✅ Decisions + supersession: D2 (Postgres+Timescale) taught with supersedesKnowledgeId=D1 → D1 soft-deleted, parentKnowledgeId links back; retrieval serves only D2 with the full instead chain rendered in the Decisions block.
2. ✅ Action-item tag contract: 3 rows (blocker, plain for:<email>, open-question) persisted with type=action_item, visibility=project, exact tag set.
3. ✅ Semantic-retrieval exclusion: a retrieve prompt literally matching the blocker text ("fix the staging database") returned only the decision — tasks are invisible to rule retrieval in prod.
4. ✅ Dark flags: a fresh session in the fixture project carried relevantKnowledge (the decision) and no openActionItems despite an item addressed to the caller — V2_ACTION_ITEMS off gates correctly.
5. ✅ Resolve lifecycle: resolvedActionItemIds at close retired exactly the named item (resolvedActionItems: 1); the other two stay open.
6. ✅ Metric purity: zero SessionKnowledgeApplication rows exist for any action_item DB-wide — this gate's injection→used rate is untouched by the task machinery.

The fixture project is left in place (tagged v2-dryrun) as the live test bed for the flag-flip on/after 2026-07-17: enabling the flags should surface the open blocker + question at the next session start there. Remaining pre-enable validation that needs the operator: the same run on a real meeting transcript (extraction quality), per spec §6.1.

🤖 Generated with Claude Code

[bejranonda]

Mid-window preview (2026-07-09, read-only SQL, same methodology as the 07-07 baseline):

Honest window (sessions started ≥ 2026-07-03 / v1.13.0):

• Opened 16 · closed 15 · closed-with-learnings 13 → 87% (gate ≥60%)
• Sessions with injection 15 · reporting knowledge used 13 → 87% (gate ≥40%)

Both metrics well above threshold with a real sample (baseline was 75%/75% at n=4). The two learning-less closes are the V2 §6.1 dry-run fixture sessions (intentionally minimal). This is a preview, not the reading — the official read stays 2026-07-17 per the spec.

Stack state at preview time: v2.0.4 deployed and drift-clean (scheduled prod-drift green), lockdown PASS, nightly backup 03:00 OK, V2 flags still dark.

🤖 Generated with Claude Code

[bejranonda]

Per-project split of the gate window (2026-07-10, read-only) — important for an honest 07-17 reading:

Project	closed sessions	with learnings
Brain Platform (meta-work on the platform itself)	14	14
V2-DryRun-Fixture (validation fixtures, learning-less by design)	4	0
Default	1	0
Dementia-Daily / Nutri-Vision-AI / Start-Idea (external repos)	0	0

The aggregate (74% closed-with-learnings) still clears the 60% bar, but the window so far measures almost exclusively the platform's own construction sessions. Zero sessions have flowed from the three external repos where the Brain-first protocol was installed on 2026-07-07. Recommendation for the official reading: report the external-repo count alongside the aggregate — if it is still ~0, treat the gate as technically passed but unproven on independent workloads, and let Stage-3's file-memory shrink proceed only for this repo, not globally.

Also done today: backup restore drill passed (nightly dump restored into an isolated pgvector container with zero errors; row counts match prod minus post-03:00 activity; all 228 embeddings intact) — and a 3-row corpus prune of KEA validation meta-noise (fixture-validated, idempotent, applied to prod under the data-repair carve-out).

🤖 Generated with Claude Code