docs/evaluation_protocol.md

Evaluation protocol

The library evaluates imported evidence; it does not reinterpret a historical rollout as a score for a new target model. Every canonical attempt is retained and receives one row in data/library/evaluation/per_trajectory.jsonl.gz. Aggregates are regenerated from that table and data/library/index.sqlite.

A. Dataset and trajectory quality

scripts/build_complete_library.py validates every canonical record, then records the following independent fields:

• structural validity: canonical IDs, roles, ordering and tool-call linkage;
• action-observation completeness: the fraction of supervised decisions with source-derived context since the preceding decision;
• outcome availability: success/failure is known versus genuinely unknown;
• verification reliability: independently reproduced, source evaluator, source-reported or unverified;
• structural SFT eligibility: valid complete record, at least two assistant decisions and full action-context coverage;
• training eligibility: structural eligibility plus the source registry gate.

Protected benchmark rollouts can be structurally suitable while remaining ineligible for training. Final-answer-only BFCL outputs are excluded from the trajectory population entirely; their audited rejection count remains in the coverage manifest. Unknown outcomes are never converted to failures.

Quality tiers and within-task ranks use the policy in trajectory_schema.md. Ranks are recomputed within the same task_key and compatible environment/version. Exact duplicate interactions retain all contributing trajectory IDs.

B. Recorded-agent performance

The baseline uses the outcome and verification evidence published with each rollout. A success is independently reproduced only when the canonical evidence explicitly says so. In this build no benchmark container was replayed, so every measured outcome is labelled source_evaluator or source_reported; independently_reproduced is zero.

The persisted summary reports:

• every attempt and unique task, including unsuccessful alternatives;
• success, failure and unknown denominators separately;
• outcomes by source, model, scaffold, task family and quality tier;
• task/environment attempt-count distributions;
• fraction of tasks solved at least once;
• empirical success-at-k for groups with at least k recorded attempts, ordered by the immutable source locator. This is descriptive and is not estimated pass@k;
• action counts for all records and cost/latency/token fields where upstream rows publish them.

Cost and latency are not aggregated across incompatible benchmarks or units. Raw values remain queryable in SQLite and the per-trajectory result file.

C. Model imitation

The two causal SFT views and exact structured next-action scorer are implemented. trajdata evaluate-actions compares per-turn assistant text and tool calls without using later observations. A scalable held-out export is:

.venv/bin/python scripts/export_sft_views.py data/library \
  --view assistant-turn --selection all --split validation \
  --output outputs/validation-turns.jsonl.gz
trajdata evaluate-actions heldout-canonical.jsonl predictions.jsonl

Teacher-forced token negative log-likelihood requires a specified model, tokenizer/chat template and inference runtime. None is configured in this repository or environment, so it was not fabricated. The machine-readable status and reason are persisted in data/library/evaluation/model_imitation.json.

Leakage controls

Task split assignment hashes task_key, not trajectory ID. All attempts, sources and derived full/turn views for the same task therefore share a split. Sources registered as evaluation-only are assigned protected_evaluation; scripts/export_sft_views.py additionally requires training_eligible=1, so protected rows cannot enter its SFT output. Outcome, reward, reference answers, later decisions and future observations are metadata and never appear in a turn's model_input.

Reproduction

After completing the pinned imports listed in trajectory_sources.md, rebuild all scores with:

.venv/bin/python scripts/build_complete_library.py \
  data/full/canonical/seed_evaluation/official_tau3_toolbench.jsonl.gz \
  --input-dir data/full/canonical/swegym_openhands_full \
  --input-dir data/full/canonical/swegym_sampled_v4_pinned \
  --input-dir data/full/canonical/swebench_verified_v2 \
  --input-dir data/full/canonical/swebench_multilingual \
  --input-dir data/full/canonical/terminalbench_v3 \
  --input-dir data/full/canonical/agentsuite_tau2_v3 \
  --input-dir data/full/canonical/browsecomp \
  --input-dir data/full/canonical/nvidia_v1_2_openhands_original \
  --input-dir data/full/canonical/terminaltraj_full \
  --input-dir data/full/canonical/mind2web_full \
  --input-dir data/full/canonical/swesmith_tool_full \
  --output data/library

.venv/bin/python scripts/build_coverage_manifest.py data/library \
  --output data/coverage/trajectory_sources.json
.venv/bin/python scripts/render_complete_gallery.py data/library \
  --output data/library/examples.html

AgentSuite BFCL is deliberately absent from the build command because its 115,466 records have no returned observations and are final-call evaluation outputs rather than trajectories.