harb/planner/MEMORY.md

Planner Memory

2026-03-22 — Run 5 summary

Bottleneck: Contract safety (unchanged since tree creation)

Contract safety is the singular bottleneck blocking all of Phase 1, Phase 2, and Phase 3. Four prerequisites remain:

• #1031 — int256 overflow in ThreePositionStrategy._computeFloorTickWithSignal
• #1019 — catch block skips clamping that try block applies
• #1067 — floor ratchet attack not yet defeated (red-team data from #1085 run exists but fix not implemented)
• #1054 — OptimizerV3 has no dedicated test file

Only #997 (overflow guard slots 1-7) resolved across 5 runs — merged 2026-03-22. Priority labels added to #1031, #1019, #1067 to signal dev agent urgency.

Evidence pipeline

• Red-team: has baseline data (evidence/red-team/2026-03-20.json)
• Evolution: best.push3 exists on disk but NOT committed to master — local-only
• Resources, protocol, holdout, user-test: all empty on master
• All 6 formulas exist on disk. Evidence gaps are not on the critical path — contract safety must clear first.

Patterns

• Predictor over-signals on infrastructure/dark-factory noise — most predictions dismissed as not harb-specific
• Contract safety work velocity: 1 prereq resolved per day when dev agent engages. At this rate, remaining 4 items could clear in ~4 days.
• Evolution pipeline produces output but doesn't persist to git — evolution commit-via-PR (#1047) was fixed but evidence still not appearing on master
• Three stale unlabeled issues (#517, #383, #147) noted by prediction #1105 — gardener has not triaged them across multiple runs

Strategic context

• Phase 1 (quality gate) is entirely blocked on contract safety
• Phase 2 (launch) is blocked on Phase 1
• Phase 3 (operations) is blocked on Phase 1
• The prerequisite tree is stable — no new objectives or prerequisites discovered in runs 2-5
• Priority label system introduced in run 5 to direct dev agent attention to bottleneck issues