Skip to content

Latest commit

 

History

History
62 lines (49 loc) · 3.81 KB

File metadata and controls

62 lines (49 loc) · 3.81 KB

Rollout Hard-Event Route Chain - 2026-06-12

This dated report records a public-safe production-like retrieval run over the synthetic rollout behavior hard-event fixture. It is a small route-chain and actionability calibration slice for #1197, #1195, and #309.

The fixture is checked in under CC0-1.0 and contains no private user history. The runner output is sanitized: it emits ids, hashes, aggregate metrics, and claim boundaries, but not raw rollout text, private source refs, local absolute paths, credentials, or provider responses.

Commands

python benchmarks\aippocampus\benchmark_vcs_future_event_recall.py --dataset benchmark_corpus\public_longitudinal_users\rollout_behavior_events_v1.jsonl --production-like-retrieval --source-disambiguation-top-k 1 --json --output docs\evidence\benchmarks\reports\public-longitudinal\rollout-hard-event-route-chain-topk1-2026-06-12.json
python benchmarks\aippocampus\benchmark_vcs_future_event_recall.py --dataset benchmark_corpus\public_longitudinal_users\rollout_behavior_events_v1.jsonl --production-like-retrieval --source-disambiguation-top-k 2 --json --output docs\evidence\benchmarks\reports\public-longitudinal\rollout-hard-event-route-chain-topk2-2026-06-12.json
python benchmarks\aippocampus\benchmark_vcs_future_event_recall.py --dataset benchmark_corpus\public_longitudinal_users\rollout_behavior_events_v1.jsonl --production-like-retrieval --source-disambiguation-top-k 3 --json --output docs\evidence\benchmarks\reports\public-longitudinal\rollout-hard-event-route-chain-topk3-2026-06-12.json

Result

Top K Gate Recall Precision Chain recovery Source failures Wrong-source evidence Stale source in top K Foreground action false positives
1 failed 0 / 3 0 / 3 0 / 3 3 3 / 3 0 / 3 0
2 passed 3 / 3 3 / 3 3 / 3 0 0 / 3 0 / 3 0
3 passed main gate, noisy 3 / 3 3 / 3 3 / 3 0 2 / 3 2 / 3 0

The useful operating point for this fixture is top-k 2:

  • top-k 1 finds the current failed behavior event but misses the paired route or reverted-edit source, so it cannot satisfy the source-backed chain;
  • top-k 2 recovers all three multi-source support chains and keeps narrative-only decoys out of foreground evidence;
  • top-k 3 keeps the main recall/precision gate green but starts carrying assistant narrative decoys in the top source set, which should stay diagnostic rather than foreground-evidence material.

What It Supports

  • The rollout hard-event runner can score future-event recall against a complete future window where silence produces false negatives.
  • Production-like retrieval can recover selected two-source rollout behavior chains without using required_past_source_ids as ranking input.
  • Successful current events can remain route candidates while the actionability gate suppresses foreground action, avoiding old-failure drag.
  • The top-k budget matters: broadening candidates after the complete chain is recovered can add source-looking noise without improving recall.

Boundary

This is a public synthetic contract slice, not live agent quality, private real-history quality, wild VCS corpus quality, external-model superiority, or source truth from assistant narrative alone. The sample is intentionally small: three positive hard-event chains, three anti-drift negatives, and three families (rejected_route, workaround_rationale, tacit_constraint). It does not close the representative public-quality requirements for #1197 or the broader #1195 benchmark promotion question by itself.

Machine-readable outputs: top-k 1, top-k 2, and top-k 3.