.gitignore

# 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
# MLX long-state serving reads (scripts/mlx_long_states.py): one report per size
!runs/mlx-long-states/*.json
!runs/mlx-long-states/summary.md
# 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; on the kev-suites mirror, load_split fetches it; 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
# longdoc-v1: ~100 MB partitions live only in jaredpalmer/kev-private-evals (kev.suite mirror)
evals/longdoc-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; on the kev-suites mirror, load_split fetches it; scripts/build_hard_v1.py regenerates the same bytes)
evals/hard-v1/train.jsonl
# round 19's record (experiments/rounds/r19.json; tests/test_rounds.py reproduces its read-out): the read-out, the breadth
# report, every arm read's report + rows, the three trials' result, provenance, in-trial transfer rows and calibration
# temperature (never their 48 GB checkpoints), the parent's transfer-r3 test read. The trials' development rows carry
# sft-v1 record ids and option keys: they are in jaredpalmer/kev-private-train (runs/r19-readout/private-rows.json)
!runs/r19-readout/
!runs/r19-readout/round19.json
!runs/r19-readout/b-vs-c.json
!runs/r19-readout/private-rows.json
!runs/r19-breadth-report/
!runs/r19-breadth-report/report.md
!runs/r19-P27-r3test/
runs/r19-P27-r3test/*
!runs/r19-P27-r3test/report.json
!runs/r19-P27-r3test/rows.json
!runs/r19-27b-*/
runs/r19-27b-*/*
!runs/r19-27b-*/report.json
!runs/r19-27b-*/rows.json
!runs/r19-27b-*/00-trial-0/
runs/r19-27b-*/00-trial-0/*
!runs/r19-27b-*/00-trial-0/result.json
!runs/r19-27b-*/00-trial-0/provenance.json
!runs/r19-27b-*/00-trial-0/transfer/
runs/r19-27b-*/00-trial-0/transfer/*
!runs/r19-27b-*/00-trial-0/transfer/rows.json
!runs/r19-27b-*/00-trial-0/calibration/
runs/r19-27b-*/00-trial-0/calibration/*
!runs/r19-27b-*/00-trial-0/calibration/temperature.json
# Kev-27B's locked transfer-v4 rows (the parent side of rounds 19 and 20's locked stage)
!runs/locked/kev-27b-v2-ungated/transfer/
runs/locked/kev-27b-v2-ungated/transfer/*
!runs/locked/kev-27b-v2-ungated/transfer/rows.json
# round 20 (experiments/rounds/r20.json): read-out, verdicts, every read's report + rows, the interpolation reports; the
# interpolated checkpoints stay on the kev-runs volume (/runs/r20-wise/<name>/checkpoint)
!runs/r20-readout/
!runs/r20-readout/round20.json
!runs/r20-verdict/
!runs/r20-verdict/*.json
!runs/r20-27b-*/
runs/r20-27b-*/*
!runs/r20-27b-*/report.json
!runs/r20-27b-*/rows.json
!runs/r20c-*/
runs/r20c-*/*
!runs/r20c-*/report.json
!runs/r20c-*/rows.json
!runs/r20-wise/
runs/r20-wise/*
!runs/r20-wise/*/
runs/r20-wise/*/*
!runs/r20-wise/*/interpolation.json
!runs/locked/kev-27b-r20-ungated/
!runs/locked/kev-27b-r20-ungated/summary.json
!runs/locked/kev-27b-r20-ungated/transfer/
runs/locked/kev-27b-r20-ungated/transfer/*
!runs/locked/kev-27b-r20-ungated/transfer/rows.json
# round 20's record: report-only extras (served metrics, breadth report) and the scienthoon drift analysis, with the two
# scienthoon reads it uses that were not in git (Jev's, converted by scripts/freeze_scienthoon.py; round 17 arm (a)'s).
# scienthoon-v1 and its scripts were removed on 2026-09-27 (kev.suite.REMOVED_SUITES); these rows stay as the record.
!runs/r20-readout/served.json
!runs/r20-breadth-report/
!runs/r20-breadth-report/report.md
!runs/r20-scienthoon/
!runs/r20-scienthoon/analysis.md
!runs/r20-scienthoon/drift.json
!runs/jev-scienthoon-v1/rows.json
!runs/r17-27b-r10k-lr2e5-scienthoon/
!runs/r17-27b-r10k-lr2e5-scienthoon/rows.json
# longdoc-v1 probe (report-only): each read's report + rows (no document text; the suite is a private mirror), Jev's
# refusals and usage, the serving/parity measurement and the by-length read-out
!runs/longdoc-v1-*/
runs/longdoc-v1-*/*
!runs/longdoc-v1-*/report.json
!runs/longdoc-v1-*/rows.json
!runs/longdoc-v1-*/rejected.json
!runs/longdoc-v1-*/usage.json
!runs/longdoc-v1-report/report.md
!runs/longdoc-v1-report/README.md
runs/longdoc-v1-jev-r3/
runs/longdoc-v1-jev*.failed-503/
# tasksource-heldout-v1: eval-only held-out datasets; partitions and family list live only in jaredpalmer/kev-private-evals / the private kev-sft manifest
evals/tasksource-heldout-v1/*.jsonl
# sft-v2 (the extended SFT corpus), sft-v2-r21 and sft-v2-r22 (rounds 21 and 22's training suites): partitions live only in jaredpalmer/kev-private-train (kev.suite mirror; built by kev-sft assemble-v2)
evals/sft-v2/*.jsonl
evals/sft-v2-r21/*.jsonl
evals/sft-v2-r22/*.jsonl
# sft-v2-r25 (round 25's training suite, derived from sft-v2-r22 by kev-sft assemble/derive_r25.py): partitions only in jaredpalmer/kev-private-train
evals/sft-v2-r25/*.jsonl
# sft-v2-r26 (round 26's training suite: sft-v2-r25 + 13,000 tasksource-v1 records, from sft-v2-r21 by kev-sft assemble/derive_r26.py): partitions only in jaredpalmer/kev-private-train
evals/sft-v2-r26/*.jsonl
# eval-only suites set aside by sft-v2's components; partitions live only in jaredpalmer/kev-private-evals
evals/ood-v2/*.jsonl
evals/agents-ood-v1/*.jsonl
evals/guardrails-ood-v1/*.jsonl
# round 22 (experiments/rounds/r22.json): the read-out, its private-rows manifest and the monitor logs; every public-suite
# read's report + rows; the ood-v2 / agents-ood-v1 / guardrails-ood-v1 reads' reports (aggregates only); the final trial's
# provenance, redacted result (result-public.json), in-trial transfer rows, in-trial calibration temperature and snapshot
# records; the breadth report; the scienthoon drift. Private (jaredpalmer/kev-private-train, runs/r22-readout/private-rows.json,
# scripts/private_rows.py restore): every tasksource-heldout-v1 read (its rows and report name the private families), the
# ood / agents-ood / guardrails-ood rows (held-out records of sft-v2's private components), the trial's development rows
# and its full result.json (its per-task table names tasksource-v1's families). Round 21's four parent reads, which rounds
# 22 and 23 reuse, the same way: reports of ood / agents-ood / guardrails-ood here, rows and the tsheld read private.
!runs/r22-readout/
!runs/r22-readout/round22.json
!runs/r22-readout/private-rows.json
!runs/r22-readout/monitor.log
!runs/r22-readout/final.log
!runs/r22-breadth-report/
!runs/r22-breadth-report/report.md
!runs/r22-scienthoon/
!runs/r22-scienthoon/drift.json
# round 24 (experiments/rounds/r24.json, a retrospective selection over existing reads). Private (never in git): the
# tasksource-heldout exclusion list (runs/r24-private/; jaredpalmer/kev-private-train runs/r24/, manifest
# runs/r24-readout/private-exclude.json), every tasksource-heldout-v1 read (its rows and report name the private families)
# and the ood / agents-ood / guardrails-ood rows (held-out records of sft-v2's private components; manifest
# runs/r24-readout/private-rows.json). Committed: the read-out, the manifests, and the public reads it scores: round 22's
# and the 2026-09-27 sweep's report + rows (OOD reads: reports only), round 22's final trial's in-trial transfer rows and
# provenance (as round 22's record keeps them).
runs/r24-private/
!runs/r24-readout/
!runs/r24-readout/round24.json
!runs/r24-readout/readout.txt
!runs/r24-readout/private-exclude.json
!runs/r24-readout/private-rows.json
!runs/sweep-*/
runs/sweep-*/*
!runs/sweep-*/report.json
!runs/sweep-*/rows.json
runs/sweep-*-tsheld/
runs/sweep-*-ood/rows.json
runs/sweep-*-agentsood/rows.json
runs/sweep-*-guardood/rows.json
!runs/r22-27b-*/
runs/r22-27b-*/*
!runs/r22-27b-*/report.json
!runs/r22-27b-*/rows.json
runs/r22-27b-*-tsheld/
runs/r22-27b-*-ood/rows.json
runs/r22-27b-*-agentsood/rows.json
runs/r22-27b-*-guardood/rows.json
runs/r21-P27-tsheld/
!runs/r22-27b-lr2e6/00-trial-0/
runs/r22-27b-lr2e6/00-trial-0/*
!runs/r22-27b-lr2e6/00-trial-0/provenance.json
!runs/r22-27b-lr2e6/00-trial-0/result-public.json
!runs/r22-27b-lr2e6/00-trial-0/transfer/
runs/r22-27b-lr2e6/00-trial-0/transfer/*
!runs/r22-27b-lr2e6/00-trial-0/transfer/rows.json
!runs/r22-27b-lr2e6/00-trial-0/calibration/
runs/r22-27b-lr2e6/00-trial-0/calibration/*
!runs/r22-27b-lr2e6/00-trial-0/calibration/temperature.json
!runs/r22-27b-lr2e6/00-trial-0/snapshots/
# round 23 (experiments/rounds/r23.json): the read-out, its table and private-rows manifest; each candidate's
# interpolation report (runs/r23-wise/<arm>/interpolation.json; the checkpoints stay on the kev-runs volume); the read
# launch records; every public-suite read's report + rows; the ood-v2 / agents-ood-v1 / guardrails-ood-v1 reads' reports
# (aggregates only). Private (jaredpalmer/kev-private-train runs/r23/, runs/r23-readout/private-rows.json,
# scripts/private_rows.py restore): every tasksource-heldout-v1 read (its rows and report name the private families) and
# the ood / agents-ood / guardrails-ood rows. Confirmation (runs/r23-verdict/, runs/r23c-*): as round 24's.
!runs/r23-readout/
runs/r23-readout/*
!runs/r23-readout/round23.json
!runs/r23-readout/readout.txt
!runs/r23-readout/tables.md
!runs/r23-readout/private-rows.json
!runs/r23-wise/
!runs/r23-wise/*/
runs/r23-wise/*/*
!runs/r23-wise/*/interpolation.json
!runs/r23-reads-*.json
!runs/r23-27b-*/
runs/r23-27b-*/*
!runs/r23-27b-*/report.json
!runs/r23-27b-*/rows.json
runs/r23-27b-*-tsheld/
runs/r23-27b-*-ood/rows.json
runs/r23-27b-*-agentsood/rows.json
runs/r23-27b-*-guardood/rows.json
# tasksource-heldout-v1 test reads (rounds 23 and 24's confirmations): private, never in git
runs/r24c-*-tshtest/
runs/r23c-*-tshtest/
# round 23's confirmation of 27b-k-w85: the verdicts, the T-interval table and the private-rows manifests; the launch
# records (runs/r23-reads-*-{tests,locked}.json, above); the public test reads (rows for hard-v1, devtools-v1 and
# documents-v1, reports for breadth-v1, documents-v2 and longdoc-v1); the locked read; the test breadth report; the serving
# reports. Private (jaredpalmer/kev-private-train runs/r23/): tasksource-heldout-v1 test rows and report, breadth-v1 test
# rows, documents-v2 rows (runs/r23-verdict/private-rows.json) and longdoc-v1 test rows (private-rows-longdoc.json).
!runs/r23-verdict/
runs/r23-verdict/*
!runs/r23-verdict/27b-tests.json
!runs/r23-verdict/27b-locked.json
!runs/r23-verdict/27b-tests-t-interval.json
!runs/r23-verdict/private-rows.json
!runs/r23-verdict/private-rows-longdoc.json
!runs/r23c-27b-cand-*/
runs/r23c-27b-cand-*/*
!runs/r23c-27b-cand-*/report.json
!runs/r23c-27b-cand-hardtest/rows.json
!runs/r23c-27b-cand-devtest/rows.json
!runs/r23c-27b-cand-docs1test/rows.json
!runs/r23-breadth-report/
runs/r23-breadth-report/*
!runs/r23-breadth-report/report.md
!runs/r23-breadth-report/report.json
!runs/serving-27b-r23/
runs/serving-27b-r23/*
!runs/serving-27b-r23/report.json
!runs/serving-27b-r23-long/
runs/serving-27b-r23-long/*
!runs/serving-27b-r23-long/report.json
!runs/locked/kev-27b-r23-ungated/
runs/locked/kev-27b-r23-ungated/*
!runs/locked/kev-27b-r23-ungated/summary.json
!runs/locked/kev-27b-r23-ungated/transfer/
runs/locked/kev-27b-r23-ungated/transfer/*
!runs/locked/kev-27b-r23-ungated/transfer/rows.json
runs/r23c-*-tshtest/
# round 25 (experiments/rounds/r25.json): the read-out (round25.json, readout.txt, tables.md), its report-only extras
# (vs-w85*, breadth-ece-by-t.json, transfer4-vs-parent.json and the scripts that wrote them) and private-rows manifest;
# the read launch records; every public-suite read's report + rows; the ood-v2 / agents-ood-v1 / guardrails-ood-v1 reads'
# reports (aggregates only); each trial's provenance.json, result-public.json (result.json without the per-task table),
# in-trial transfer rows, in-trial calibration temperature, training metrics and the snapshot.json records. Private
# (jaredpalmer/kev-private-train runs/r25/, runs/r25-readout/private-rows.json, scripts/private_rows.py restore): every
# tasksource-heldout-v1 read (rows and report), the ood / agents-ood / guardrails-ood rows, both trials' development rows
# and full result.json. Checkpoints (~51 GB each) stay on the kev-runs volume.
!runs/r25-readout/
runs/r25-readout/*
!runs/r25-readout/round25.json
!runs/r25-readout/readout.txt
!runs/r25-readout/tables.md
!runs/r25-readout/vs-w85.md
!runs/r25-readout/vs-w85-*.json
!runs/r25-readout/breadth-ece-by-t.json
!runs/r25-readout/transfer4-vs-parent.json
!runs/r25-readout/private-rows.json
!runs/r25-readout/r25-*.py
!runs/r25-reads-*.json
!runs/r25-27b-*/
runs/r25-27b-*/*
!runs/r25-27b-*/report.json
!runs/r25-27b-*/rows.json
runs/r25-27b-*-tsheld/
runs/r25-27b-*-ood/rows.json
runs/r25-27b-*-agentsood/rows.json
runs/r25-27b-*-guardood/rows.json
!runs/r25-27b-lr1e6/00-trial-0/
!runs/r25-27b-lr2e6/00-trial-0/
runs/r25-27b-lr*e6/00-trial-0/*
!runs/r25-27b-lr*e6/00-trial-0/provenance.json
!runs/r25-27b-lr*e6/00-trial-0/result-public.json
!runs/r25-27b-lr*e6/00-trial-0/transfer/
runs/r25-27b-lr*e6/00-trial-0/transfer/*
!runs/r25-27b-lr*e6/00-trial-0/transfer/rows.json
!runs/r25-27b-lr*e6/00-trial-0/calibration/
runs/r25-27b-lr*e6/00-trial-0/calibration/*
!runs/r25-27b-lr*e6/00-trial-0/calibration/temperature.json
!runs/r25-27b-lr*e6/00-trial-0/checkpoint/
runs/r25-27b-lr*e6/00-trial-0/checkpoint/*
!runs/r25-27b-lr*e6/00-trial-0/checkpoint/training_metrics.json
!runs/r25-27b-lr*e6/00-trial-0/snapshots/
!runs/r25-27b-lr*e6/00-trial-0/snapshots/*/
!runs/r25-27b-lr*e6/00-trial-0/snapshots/*/checkpoint/
runs/r25-27b-lr*e6/00-trial-0/snapshots/*/checkpoint/*
!runs/r25-27b-lr*e6/00-trial-0/snapshots/*/checkpoint/snapshot.json
# round 26 (experiments/rounds/r26.json): the read-out (round26.json, readout.txt, tables.md), its report-only extras
# (vs-r25* / vs-w85* against round 25's lr 1e-6 arm and round 23's 27b-k-w85, breadth-ece-by-t.json and the scripts that
# wrote them, incl. the spend gate and launch helpers) and private-rows manifest; the read launch records; every public-suite
# read's report + rows; the ood-v2 / agents-ood-v1 / guardrails-ood-v1 reads' reports (aggregates only); the trial's
# provenance.json, result-public.json (result.json without the per-task table), in-trial transfer rows, in-trial calibration
# temperature, training metrics and the snapshot.json records. Private (jaredpalmer/kev-private-train runs/r26/,
# runs/r26-readout/private-rows.json, scripts/private_rows.py restore): every tasksource-heldout-v1 read (rows and report),
# the ood / agents-ood / guardrails-ood rows, the trial's development rows and full result.json. Checkpoints (~51 GB each)
# stay on the kev-runs volume.
!runs/r26-readout/
runs/r26-readout/*
!runs/r26-readout/round26.json
!runs/r26-readout/readout.txt
!runs/r26-readout/tables.md
!runs/r26-readout/vs-*.md
!runs/r26-readout/vs-*.json
!runs/r26-readout/breadth-ece-by-t.json
!runs/r26-readout/both-bars-by-t.json
!runs/r26-readout/private-rows.json
!runs/r26-readout/r26-*.py
!runs/r26-readout/r26-*.sh
!runs/r26-reads-*.json
!runs/r26-27b-*/
runs/r26-27b-*/*
!runs/r26-27b-*/report.json
!runs/r26-27b-*/rows.json
runs/r26-27b-*-tsheld/
runs/r26-27b-*-ood/rows.json
runs/r26-27b-*-agentsood/rows.json
runs/r26-27b-*-guardood/rows.json
!runs/r26-27b-lr1e6/00-trial-0/
runs/r26-27b-lr1e6/00-trial-0/*
!runs/r26-27b-lr1e6/00-trial-0/provenance.json
!runs/r26-27b-lr1e6/00-trial-0/result-public.json
!runs/r26-27b-lr1e6/00-trial-0/transfer/
runs/r26-27b-lr1e6/00-trial-0/transfer/*
!runs/r26-27b-lr1e6/00-trial-0/transfer/rows.json
!runs/r26-27b-lr1e6/00-trial-0/calibration/
runs/r26-27b-lr1e6/00-trial-0/calibration/*
!runs/r26-27b-lr1e6/00-trial-0/calibration/temperature.json
!runs/r26-27b-lr1e6/00-trial-0/checkpoint/
runs/r26-27b-lr1e6/00-trial-0/checkpoint/*
!runs/r26-27b-lr1e6/00-trial-0/checkpoint/training_metrics.json
!runs/r26-27b-lr1e6/00-trial-0/snapshots/
!runs/r26-27b-lr1e6/00-trial-0/snapshots/*/
!runs/r26-27b-lr1e6/00-trial-0/snapshots/*/checkpoint/
runs/r26-27b-lr1e6/00-trial-0/snapshots/*/checkpoint/*
!runs/r26-27b-lr1e6/00-trial-0/snapshots/*/checkpoint/snapshot.json
# release candidate Kev-27B v2 (internal id kev-27b-r23; PLAN.md "Release candidate: Kev-27B v2"): the release numbers,
# the staging record (volume copy, hashes, temperature fit, private Hub upload and its verification) and the verification
# read of the private Hub checkpoint (runs/rel27-hub-semif: report + rows)
!runs/release/kev-27b-r23.json
!runs/release/kev-27b-r23-staging.json
!runs/rel27-hub-semif/rows.json
# released Kev-27B v2 (PLAN.md "Released: Kev-27B v2"): the public upload + anonymous verification record and the verification
# reads of the public repository (runs/rel27-public: the resolutions in verify.json, each read's report + rows)
!runs/release/kev-27b-r23-published.json
!runs/rel27-public/verify.json
!runs/rel27-public/*/
!runs/rel27-public/*/report.json
!runs/rel27-public/*/rows.json
# Kev 1.0 (PLAN.md "Released: Kev 1.0"): the release record (Hub card commits and v1.0 tags, GitHub assets and hashes,
# verification), the assets' SHA256SUMS + manifest, kev-family's retired notes and the anonymous 2-record smoke reads
!runs/release/kev-1.0.json
!runs/release/kev-1.0-SHA256SUMS.txt
!runs/release/kev-1.0-manifest.json
!runs/release/kev-family-notes-retired.md
!runs/release/kev-1.0-verify/
!runs/release/kev-1.0-verify/*/
!runs/release/kev-1.0-verify/*/report.json
!runs/release/kev-1.0-verify/*/rows.json
# kev-deploy at KEV_REF 0a58b69 against Kev-27B v2: the check (a ~16k-token state, admitted and answered) and the served
# semif-v1 development read it compares with runs/rel27-public (the long request itself is rebuilt by make_long_state.py)
!runs/kev-deploy-27b-v2/deploy_check.json
!runs/kev-deploy-27b-v2/make_long_state.py
!runs/kev-deploy-27b-v2/semif-v1/rows.json
# round 27 (experiments/rounds/r27.json): the read-out, verdicts and launch records; round 18's development reads and the
# parent's reads it names (rows + report), the new transfer-r3 calibration reads, the confirmation reads (runs/r27c-*: rows
# + report, no document text) and the locked read; the release numbers and the staged-checkpoint check.
!runs/r27-readout/
!runs/r27-verdict/
!runs/r27-reads-*.json
!runs/r27-9b-*/
runs/r27-9b-*/predictions.jsonl
!runs/r27c-9b-*/
runs/r27c-9b-*/predictions.jsonl
!runs/r18-9b-joint-*/
runs/r18-9b-joint-*/predictions.jsonl
!runs/hv1-P9/
!runs/dt1-P9/
!runs/docs1-P9/
runs/hv1-P9/predictions.jsonl
runs/dt1-P9/predictions.jsonl
runs/docs1-P9/predictions.jsonl
!runs/rel9-staged-semif/
runs/rel9-staged-semif/predictions.jsonl
!runs/r27-readout/*
!runs/r27-verdict/*
!runs/r27-9b-*/*
!runs/r27c-9b-*/*
!runs/r18-9b-joint-*/*
!runs/hv1-P9/*
!runs/dt1-P9/*
!runs/docs1-P9/*
!runs/rel9-staged-semif/*
runs/r27-9b-*/predictions.jsonl
runs/r27c-9b-*/predictions.jsonl
runs/r18-9b-joint-*/predictions.jsonl
runs/hv1-P9/predictions.jsonl
runs/dt1-P9/predictions.jsonl
runs/docs1-P9/predictions.jsonl
runs/rel9-staged-semif/predictions.jsonl
runs/r27c-9b-*-docs2/rows.json
# drift-v1 (runs/drift-v1/REPORT.md): why Kev-27B v1's reads moved between 2026-09-23 and 2026-09-30; the probe, its
# comparison scripts and every read it made (per row: id, question, raw logits; no record text)
!runs/drift-v1/*.py
!runs/drift-v1/*.json
!runs/drift-v1/REPORT.md
!runs/drift-v1/*/
!runs/drift-v1/*/*.json
# rounds 28 and 29 (experiments/rounds/r28.json, r29.json; phase A of the development-read sweep): the read-outs, their
# private-rows manifests; every public-suite read's report + rows (breadth-v1, hard-v1, devtools-v1, transfer-r3); the
# ood-v2 reads' reports (aggregates only); the existing reads the read-outs name that were not yet in git. Private
# (jaredpalmer/kev-private-train runs/r28/, runs/r29/, scripts/private_rows.py restore): every tasksource-heldout-v1 read
# (rows and report name the private families) and the ood-v2 rows.
!runs/r28-readout/
!runs/r28-readout/*
!runs/r29-phase-a/
!runs/r29-phase-a/*
!runs/r28-*/
!runs/r29-*/
!runs/r28-*/rows.json
!runs/r29-*/rows.json
# round 28's validated context read-out (PLAN.md > Round 28 (registered) > validated context length): longdoc_report.py's report.md too
!runs/r28-context*/report.md
runs/r28-*/predictions.jsonl
runs/r29-*/predictions.jsonl
runs/r28-*-tsheld/
runs/r29-*-tsheld/
runs/r28-*-ood/rows.json
runs/r29-*-ood/rows.json
# phase B (2026-10-01): the agents-ood-v1 / guardrails-ood-v1 rows are private like the ood-v2 rows (held-out records of
# sft-v2's private components; jaredpalmer/kev-private-train, runs/r28-readout/private-rows.json); reports stay here.
# r28-4b-r10-agentsood-t10800 is the volume name of the re-read copied to runs/r28-4b-r10-agentsood.
runs/r28-*-agentsood/rows.json
runs/r29-*-agentsood/rows.json
runs/r28-*-guardood/rows.json
runs/r29-*-guardood/rows.json
runs/r28-4b-r10-agentsood-t10800/
# round 29's read-out (aggregates only; no tasksource-heldout family names)
!runs/r29-readout/
!runs/r29-readout/*
!runs/fam-4b-breadth/rows.json
!runs/fam-08b-breadth/rows.json
!runs/fam-9b-breadth/rows.json
!runs/fam-9bnew-breadth/rows.json
!runs/r10-4b-skills-*/rows.json
!runs/r15-08b-a-*/rows.json
!runs/r7-9b-s*-*/rows.json
!runs/r9-9b-*-*/rows.json
!runs/r11-9b-s*-*/rows.json
!runs/r16-9b-r10k-*/rows.json
runs/fam-*/predictions.jsonl
# the trials these read-outs serve: provenance, in-trial transfer rows (the transfer4 panel; excluded from the temperature
# pool) and, for round 28's parents, the development rows their shipped T is fitted on (decision-v7, public)
!runs/r10-skills/00-trial-0/provenance.json
!runs/r15-08b/00-trial-0/provenance.json
!runs/r10-skills/00-trial-0/development/
runs/r10-skills/00-trial-0/development/*
!runs/r10-skills/00-trial-0/development/rows.json
!runs/r10-skills/*-trial-*/transfer/
!runs/r15-08b/*-trial-*/transfer/
!runs/r7-docs/*-trial-*/transfer/
!runs/r9-docs/*-trial-*/transfer/
!runs/r11-docs/*-trial-*/transfer/
!runs/r16-9b/*-trial-*/transfer/
!runs/r18-9b/*-trial-*/transfer/
runs/r10-skills/*-trial-*/transfer/*
runs/r15-08b/*-trial-*/transfer/*
runs/r7-docs/*-trial-*/transfer/*
runs/r9-docs/*-trial-*/transfer/*
runs/r11-docs/*-trial-*/transfer/*
runs/r16-9b/*-trial-*/transfer/*
runs/r18-9b/*-trial-*/transfer/*
!runs/r10-skills/*-trial-*/transfer/rows.json
!runs/r15-08b/*-trial-*/transfer/rows.json
!runs/r7-docs/*-trial-*/transfer/rows.json
!runs/r9-docs/*-trial-*/transfer/rows.json
!runs/r11-docs/*-trial-*/transfer/rows.json
!runs/r16-9b/*-trial-*/transfer/rows.json
!runs/r18-9b/*-trial-*/transfer/rows.json
!runs/r7-docs/*-trial-*/provenance.json
!runs/r9-docs/*-trial-*/provenance.json
!runs/r11-docs/*-trial-*/provenance.json
!runs/r16-9b/*-trial-*/provenance.json
!runs/r18-9b/*-trial-*/provenance.json