# python

.venv/

__pycache__/

.pytest_cache/

*.pyc

# model runs: weights are large; keep eval results and training logs (for the plot)

runs/*

!runs/kev/

runs/kev/*

!runs/kev/eval.json

!runs/logs/

runs/logs/*

!runs/logs/train_*.log

# logs elsewhere

*.log

!runs/logs/train_*.log

# os

.DS_Store

# research results: keep reports and comparisons, not raw prediction dumps

!runs/research-kev-v01/

runs/research-kev-v01/*

!runs/research-kev-v01/report.json

!runs/research-jev-v1/

runs/research-jev-v1/*

!runs/research-jev-v1/report.json

!runs/research-jev-v1/usage.json

!runs/kev-vs-jev-v1.json

!runs/transfer-kev-v01/

runs/transfer-kev-v01/*

!runs/transfer-kev-v01/report.json

!runs/transfer-jev-v1/

runs/transfer-jev-v1/*

!runs/transfer-jev-v1/report.json

!runs/transfer-jev-v1/usage.json

!runs/kev-vs-jev-transfer-v1.json

# vendored agent skills (modal skills install); reinstall with `uv run modal skills install`

.agents/skills/modal/

# Modal studies: keep results/ledger/provenance, not checkpoints or prediction dumps

!runs/backbone-v1/

!runs/ablation-v2/

runs/ablation-v2/*

!runs/ablation-v2/results.jsonl

!runs/ablation-v2/*/

runs/ablation-v2/*/*

!runs/ablation-v2/*/result.json

!runs/ablation-v2/*/provenance.json

!runs/ablation-v2/*/train.log

runs/backbone-v1/*

!runs/backbone-v1/results.jsonl

!runs/backbone-v1/*/

runs/backbone-v1/*/*

!runs/backbone-v1/*/result.json

!runs/backbone-v1/*/provenance.json

!runs/backbone-v1/*/train.log

!runs/research-kev2-dev/

runs/research-kev2-dev/*

!runs/research-kev2-dev/report.json

# v3 studies: results, ledger, provenance, comparisons, train logs only

!runs/v3-*/

runs/v3-*/*

!runs/v3-*/results.jsonl

!runs/v3-*/*/

runs/v3-*/*/*

!runs/v3-*/*/result.json

!runs/v3-*/*/comparison.json

!runs/v3-*/*/provenance.json

!runs/v3-*/*/train.log

# every study under runs/: keep ledgers, results, provenance, comparisons, training logs, Jev reports; never weights or prediction dumps

!runs/*/

runs/*/*

!runs/*/results.jsonl

!runs/*/report.json

!runs/*/usage.json

!runs/*/summary.json

!runs/*/comparison.json

!runs/*/calibration.json

# external-suite predictions (small): committed so anyone can refit a workload temperature or bootstrap (kev.calibrate)

!runs/kev-*-wanli-v1/rows.json

!runs/kev-*-typesafe-v1/rows.json

# long-state reads (round 4.12): the per-length paired report and the rows it is computed from

!runs/*/longstate.json

!runs/r4-*longstate*/rows.json

!runs/r4-*-ls/rows.json

# the soft-target Kev-9B release confirmation: every input scripts/release_confirm.py reads

!runs/rc-*/rows.json

!runs/*/*/

runs/*/*/*

!runs/*/*/result.json

!runs/*/*/comparison.json

!runs/*/*/provenance.json

!runs/*/*/train.log

!runs/*/*/summary.json

!runs/*/*/report.json

!runs/leaderboard.md

!runs/leaderboard.jsonl

!runs/autoresearch.jsonl

# frozen-suite partitions over ~10 MB live on the Hub mirror (jaredpalmer/kev-suites, pinned in kev/suite.py); manifests,

# development and test partitions stay in git and load_split fetches + sha256-verifies the rest on first use

evals/v5/decision-v5/train.jsonl

evals/v6/decision-v6/train.jsonl

evals/v4/decision-v4/train.jsonl

evals/public-pool-v5/train.jsonl

evals/public-pool-v6/train.jsonl

evals/public-pool-v4/train.jsonl

evals/v7/decision-v7/train.jsonl

evals/v8/decision-v8/train.jsonl

evals/round3/decision-r3/train.jsonl

# setuptools build metadata (kev is installed editable by uv sync since #14)

*.egg-info/

# kev-finetune skill: local data and pulled run reports when the scripts are run from the skill directory

skills/kev-finetune/data/

skills/kev-finetune/runs/

# round-4 delta data over ~10 MB (regenerated by scripts/build_long_states.py / build_soft_targets.py; sha256 in each manifest)

evals/round4/longstate-v1/train.jsonl

evals/round4/distill-9b-v1/soft.jsonl

evals/round4/distill-9b-v1/hard.jsonl

# release confirmation (scripts/release_confirm.py): the three arms' development rows its temperatures are fitted on

!runs/r4-soft/00-trial-0/development/

runs/r4-soft/00-trial-0/development/*

!runs/r4-soft/00-trial-0/development/rows.json

!runs/r4-deltas/01-trial-1/development/

runs/r4-deltas/01-trial-1/development/*

!runs/r4-deltas/01-trial-1/development/rows.json

!runs/night2-9b-du/00-trial-0/development/

runs/night2-9b-du/00-trial-0/development/*

!runs/night2-9b-du/00-trial-0/development/rows.json

# round-5 training data over ~10 MB (regenerated by scripts/build_long_states.py and concatenation; sha256 in each manifest)

evals/round5/longstate-v2/train.jsonl

evals/round5/combined-v1/soft.jsonl

evals/round5/combined-v1/hard.jsonl

# round-5 verdict inputs (experiments/rounds/r5.json, python -m kev.rounds readout): every read it scores

!runs/r5r-*/rows.json

!runs/r5-combined/00-trial-0/development/

runs/r5-combined/00-trial-0/development/*

!runs/r5-combined/00-trial-0/development/rows.json

!runs/r5-combined/01-trial-1/development/

runs/r5-combined/01-trial-1/development/*

!runs/r5-combined/01-trial-1/development/rows.json

!runs/r5-combined/02-trial-2/development/

runs/r5-combined/02-trial-2/development/*

!runs/r5-combined/02-trial-2/development/rows.json

!runs/r5-combined/03-trial-3/development/

runs/r5-combined/03-trial-3/development/*

!runs/r5-combined/03-trial-3/development/rows.json

!runs/r5-combined/04-trial-4/development/

runs/r5-combined/04-trial-4/development/*

!runs/r5-combined/04-trial-4/development/rows.json

!runs/r5-combined/05-trial-5/development/

runs/r5-combined/05-trial-5/development/*

!runs/r5-combined/05-trial-5/development/rows.json

!runs/night2-4b-du/00-trial-0/development/

runs/night2-4b-du/00-trial-0/development/*

!runs/night2-4b-du/00-trial-0/development/rows.json

!runs/night2-08b-du2/00-trial-0/development/

runs/night2-08b-du2/00-trial-0/development/*

!runs/night2-08b-du2/00-trial-0/development/rows.json

!runs/r5-verdict/

!runs/r5-verdict/*.json

# round 15's locked verdict (the released Kev-0.8B's confirmation): the rows tests/test_rounds.py reproduces it from offline

!runs/r15-08b/00-trial-0/development/

runs/r15-08b/00-trial-0/development/*

!runs/r15-08b/00-trial-0/development/rows.json

!runs/locked/kev-08b-r15-ungated/transfer/

runs/locked/kev-08b-r15-ungated/transfer/*

!runs/locked/kev-08b-r15-ungated/transfer/rows.json

!runs/locked/kev-08b-night2-du-ungated/transfer/

runs/locked/kev-08b-night2-du-ungated/transfer/*

!runs/locked/kev-08b-night2-du-ungated/transfer/rows.json

# documents-v1 train partition (23 MB; sha256 in evals/documents-v1/manifest.json; rebuild with scripts/build_documents_v1.py + label + freeze)

evals/documents-v1/train.jsonl

!runs/documents-v1-work/

!runs/documents-v1-work/adjudications.jsonl

!runs/documents-v1-work/spot_check.jsonl

!runs/documents-v1-work/spot_check_reviews.jsonl

!runs/documents-v1-work/labels/

!runs/documents-v1-work/labels/spend.json

!runs/documents-v1-work/candidates/

!runs/documents-v1-work/candidates/build.json

# documents-v2: private held-out suite; partitions live only in jaredpalmer/kev-private-evals (kev.suite mirror)

evals/documents-v2/*.jsonl

# breadth-v1: partitions live only in jaredpalmer/kev-private-evals (licences forbid public redistribution; kev.suite mirror)

evals/breadth-v1/*.jsonl

# sft-v1: Kev-27B's SFT suite; partitions live only in jaredpalmer/kev-private-train (kev.suite mirror; built by kev-sft assemble-v1)

evals/sft-v1/*.jsonl

# round-19 attribution arm's data: Kev-27B's own training set (47 MB; manifest from tag research-archive-2026-09-24, also sft-v1/components/b1v2-train in kev-private-train)

evals/round6/b1v2/train.jsonl

runs/documents-v2-work/

# hard-v1 train partition (23 MB; sha256 in evals/hard-v1/manifest.json; scripts/build_hard_v1.py regenerates the same bytes)

evals/hard-v1/train.jsonl