Results (paired cells, re-aggregated from raw cell_result.json):
Object 28/30=93.3% vs 29/30=96.7%
Spatial 3/30=10.0% vs 1/30= 3.3%
Goal 5/30=16.7% vs 6/30=20.0%
Long 3/15=20.0% vs 3/15=20.0% (in progress)
TOTAL 39/105=37.1% vs 39/105=37.1% -- dead heat at 3-10x fewer VLM calls
Adds:
- TIMELINE.md: dated progress timeline, including corrections to earlier claims
- results/all_cells.json, logs/ (gate history + archived validator rejections)
- patches/: two new validator rules found by this evaluation
W9 parent conditional_edges must cover every subgraph exit
W10 router_field must be a non-empty string
both opt-in on the codegen path; each previously produced 3 dead trials
with 0 control steps that scored as legitimate task failures
- v3_perception/: relation-aware re-ranking + ambiguity guard (14 unit tests)
- friend_run/: self-contained reproduction bundle
- scripts/: per-task YAMLs for all four suites + quota-aware resumable drivers
Al…
