Research log / RL-005

RL-005 Grounded follow-ups and typed search outcomes

Evidence grade AB narrow

Field Value
Date 2026-09-10
Harness version 2026-09-10-grounded-followups-3
Models qwen2.5-coder:7b (live use); deterministic paths in simulation
Evidence grade A (narrow: 2 scenarios × 3 repeats) for routing/state; B for the search diagnosis
Pre-registered no (backfilled 2026-09-13)
Source record FOLLOWUP-REVIEW.md, grounded-followup-evaluation.json

Observation

im bored was treated as a name introduction; describing a workspace invented gameplay details ("guards, inmates") for prisonidle; empty searches were indistinguishable from failed ones. A pasted review asserted the empty searches must have been transport or parser failures.

Hypotheses (reconstructed)

Intervention

Mood-statement guard; name corrections as conversation-local facts with provenance; catalogue-ID-grounded descriptions that say when only metadata is available; deterministic search queries; typed error categories (HTTP_403, HTTP_429, INVALID_JSON, TIMEOUT, …); failed searches clear cached results.

Method

npm run sim -- grounded-followups 3 with a synthetic workspace catalogue. Direct live check of the search integration with the query fun games to play.

Results

Threats to validity

Simulated paths are deterministic, so they establish routing and state, not improved model reasoning. The diagnosis ran on the current implementation, not the one that produced the original empties.

Conclusion

Follow-ups

Acceptance-criteria checking of generated content (→ RL-010, RL-011).

Reproduce

npm run sim -- grounded-followups 3