Harryis/LWM_Phase1

RobustAgentWorldBench

benchmark/ contains one current formal release: rawb-v2. Phase 1 is closed.

Canonical Phase-1 artifacts

Purpose	Path
Frozen questions	data/questions/release_v2/
Release contract	specs/RELEASE_V2.md
Oracle and three-model results	results/release_v2/
Phase-1 close report	results/release_v2/PHASE1_REPORT.md
Model comparison	results/release_v2/MODEL_COMPARISON.md

Rawb-v2 contains 600 model inputs: A=200, B=200, C=200. This is 400 logical units and 1,314 next-observation calls per evaluated model. Do not edit the frozen JSONL files or their manifest/checksums in place.

Active implementation

• environments/: executable Terminal, MCP, and Controlled-Web adapters;
• protocol/: AgentWorld-compatible message and carrier handling;
• evaluation/: runners, scoring, aggregation, audits, and close verification;
• generation/: reproducibility code for the accepted construction pipeline;
• tests/: current release tests plus retained provenance regression tests.

Historical material

All superseded pilots, candidate grids, rawb-v1, calibration runs, smoke outputs, and accepted construction provenance are grouped under archive/phase1/. Nothing was deleted during the 2026-09-01 cleanup. See archive/phase1/README.md before using an archived artifact.

The links data/questions/{a_boundary,b_commitment,c_adversarial} and results/release_v1 exist only because frozen rawb-v2 metadata records those original provenance paths. They point into the archive and are not additional current releases.

Verification

python -m benchmark.evaluation.run_release \
  --oracle --output-dir /tmp/rawb-v2-oracle
python -m benchmark.evaluation.audit_frozen_release \
  --release benchmark/data/questions/release_v2 \
  --b-results benchmark/results/release_v2/oracle/track_b.jsonl \
  --output /tmp/rawb-v2-audit.json
python -m unittest discover -s benchmark/tests -v
python -m benchmark.evaluation.verify_phase1_close

Phase 2 must treat rawb-v2 as held out. The LWMs remain frozen environment simulators; Phase 2 trains Agent policies in matched real versus LWM-backed environments and evaluates them only in real executable environments.