You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
fix(bench): modern DataHub seeding, enrichment-volume reporting, and the first published pilot results (#951)
Standing the a2 arm up on a current DataHub (v1.6.0 quickstart on
OpenSearch) surfaced two harness defects, both exercised live after the fix:
- The MCE emitter wrote the legacy datahub put --file bulk format, which
current datahub CLIs no longer accept. Aspects are now emitted in the MCP
file format (GenericAspect {"json": ...}), and bench-seed-datahub drives
datahub ingest with a generated file-source recipe (BENCH_DATAHUB_GMS
overrides the sink). Seed artifacts regenerated.
- Suite reports summed enrichment_tokens_dedup, which is zero when
enrichment runs in full mode, so the a2 arm's enrichment volume read as
zero. Metrics and reports now carry enrichment_tokens_full (the delivered
volume) alongside dedup, and the suite table prints the full volume.
docs/reference/benchmarks.md is the new living results page (mkdocs nav +
llms pair updated): phase 1 pilot manifest, arm-by-suite and trap-class
tables, honest caveats, and a result-history table that future phases
append to. Raw pilot run JSON (claude-sonnet-5, k=3, 60 episodes, zero
harness failures) is committed under bench/results/v1.102.0-pilot/ as the
first history entry.
Pilot headline recorded on #930: knowledge-trap accuracy 60% (a0) vs 100%
(a2) with the policy-revenue tasks going 0/6 vs 6/6, median knowledge-suite
tool calls 16 vs 8, wall clock 67s vs 25s.
0 commit comments