REPRODUCTION GUIDE — source snapshot, 3 November 2026

This package contains the code used for a new, explicitly reconstructed run of
experiments 1–8. It is not the authors' missing private harness or a turnkey
installer. Interactive experiments and long GPU evaluations were still running
when this source snapshot was packaged. Consult the current replication report
and individual case states for completed results; a prepared case is not a result.

The accompanying reproduction-source.zip contains a Git archive of the reviewed
reconstruction repository, the initial one-shot launch/audit scripts, this guide,
and compact provenance. archived-runtime/ retains the exact older Exp4 parser
and Exp7 training orchestrator used before later review-only guard improvements.
MANIFEST.json pins the Git commit and every member's
SHA-256/size. It excludes model weights, datasets, credentials, CLI authentication,
and live scratch directories. Existing review JSON files describe this recorded
run; they do not authorize a changed installation. Preserve upstream licenses.

SOURCE AND MODEL PINS

Published repository: https://github.com/RikiyaT/lm_understand-logs
Commit: 4d09ba6d6a7e891dd8063974dfe81b3d78e3494f
Kinetix: https://github.com/FLAIROx/Kinetix
Commit: 80ee9c292837c825cafaf9889323fdcc212efd1b; jax2d 1.0.1.
Qwen/Qwen3.5-0.8B: 2fc06364715b967f1860aea9cf38778875588b17.
google/gemma-3-270m-it: ac82b4e820549b854eebf28ce6dedaf9fdfa17b3.

Qwen's text-only export is built by exp8/prepare.py from the pinned full snapshot
and released configuration. All 320 text tensors retain original names/values.
Full source checkpoint SHA-256:
04b1c301231dd422b8860db31311ab2721511346a32cb1e079c4c4e5f1fe4696
Exported model.safetensors SHA-256:
978661a573971be7257815c2554a28c74913c7b7ea51b0a0b9503fd11eb0d2f8
Gemma base model.safetensors SHA-256:
700b710a9a99c295ed546647aa81cacf9f81f4c573ea2be613a0e2517a44afab

The compact dataset manifest pins all ten Exp8 evaluation sets and their prepared
checksums. The official MMLU-Pro helper is pinned to commit
3016805055f78c55ec54a2f9be9fd5f89be95e0a; official IFEval checkers to
e6890f85757dd84e27ca6df2dd30651dafad28e0. GPQA Diamond comes from the
authors' public GitHub dataset.zip at commit
56686c06f5e19865c153de0fdb11be3890014df7, using the download instructions
in that repository's README. Its extracted CSV SHA-256 is
41d1213cd7a4998605a26c2798500652572007161b3a92817ba46b35befcd305;
its Git blob matches the official gated HF revision
83022cefff930aea54f654c0b282e74b9eeda5c6. No third-party dataset mirror is used.

ACTUAL HOSTS AND SETUP REQUIREMENTS

The controller ran on macOS 15.6 arm64, using codex-cli 0.166.0 at
<local-home>/.local/bin/codex, with a logged-in Codex account. The GPU host was
the SSH alias <gpu-ssh-alias>: NVIDIA DGX Spark/GB10, 121 GiB unified RAM, aarch64 Linux
6.17.0-1031-nvidia, Docker 29.6.2. Paths below are recorded host paths, not portable
defaults. Configure your own SSH connection and authentication outside this bundle.

Local workspace (ROOT):
<workspace>
Remote task root (TASK): <gpu-project>
Local immutable source (SRC): ROOT/work/lm_understand-logs-<source commit>
Remote immutable source: TASK/source
Local reconstructed checkout (CODE): ROOT/work/reconstructed-evaluation

Extract reconstruction/ to ROOT/work/reconstructed-evaluation and bootstrap/ to
ROOT/work; these relative locations matter to the scripts' BASE calculations.
Fetch the published repository at its exact commit separately. Place only one
lm_understand-logs-* checkout under ROOT/work because several drivers select it
by glob. Create ROOT/work/replication-empty and the required output directories.
Keep each fresh rerun in a separate workspace/output tree. Never overwrite the
recorded runs or run two controllers against the same case.

For example, with actual paths substituted:
  git clone https://github.com/RikiyaT/lm_understand-logs SRC
  git -C SRC checkout --detach 4d09ba6d6a7e891dd8063974dfe81b3d78e3494f
  git clone https://github.com/FLAIROx/Kinetix PINNED_KINETIX_DIRECTORY
  git -C PINNED_KINETIX_DIRECTORY checkout --detach 80ee9c292837c825cafaf9889323fdcc212efd1b
Download HF model snapshots at the revisions above with your own authorized
account when required; verify their file hashes before export/training/inference.
Exp8 Gemma validation expects download-manifest.json recording revision and each
file's name/size/SHA-256; its exact recorded schema is included in metadata.

Several modules contain the recorded SSH alias, absolute remote paths, Python
paths and image IDs. Search these before porting: <gpu-ssh-alias>, <gpu-home>, /Users/,
REMOTE, ROOT and VENV. Record every path/environment adaptation as a new revision,
rerun preflights, and obtain a new hash-bound startup review. Do not edit the
recorded production manifests to make changed code appear identical.

The initial macOS experiment-cli.sb denies reads of two specific personal AGENTS
files. Adapt those two paths to the actual controller account. sandbox-exec and
the CLI rollout location ~/.codex/sessions are macOS/runtime dependencies, not
portable assumptions. The actual context audit, rather than flags alone, must
confirm absence of injected personal instructions and unexpected tool calls.

Remote Python environments used:
  TASK/.venv/bin/python: CPU NumPy/MinAtar; MinAtar 1.0.15.
    The reconstructed Exp5 PPO campaign additionally uses Torch 2.15.0+cpu,
    NumPy 2.5.4 and Python 3.12.3 in this task-owned environment.
  TASK/.venv-kinetix/bin/python: Python 3.12, JAX 0.9.0, Flax 0.12.3,
    CUDA 13; see reconstruction/exp4-requirements.lock.
  <gpu-python-environment>/bin/python: Python 3.12.13,
    Torch 2.14.0+cu130, Transformers 5.20.2, vLLM 0.33.0.
    This pre-existing environment was used read-only, not modified.
  TASK/reconstructed/exp6/deps: PGX 2.6.0, svgwrite 1.4.3 overlay.
  Exp7 reader CPU environment: NumPy 2.2.6; other ML packages unavailable.
  TASK/reconstructed/exp8/cpu-deps: task-owned scientific Python/math grader
    dependencies, including math-verify 0.9.0 and latex2sympy2_extended 1.11.0.
  TASK/reconstructed/exp8-bench/deps and nltk_data: evaluator-only dependencies
    and language-tokenization resources, bound by benchmark provenance.

Use the included requirements/provenance records to provision equivalent
task-owned environments; the bundle does not provision CUDA wheels, Docker,
NLTK corpora, HF downloads or these virtualenvs. Some recorded CUDA packages are
platform-specific. A package list alone cannot recreate the historical binary
runtime. Preserve library/source hashes and report unavoidable differences.

CPU sandboxes require Docker and readable Python runtime mounts. The recorded
Ubuntu image ID is sha256:5b8c0c14690ed170da4e663fe0bae0d58efe59661e791296ffab28ed2113b650.
An image ID is an audit identity, not a registry download tag. If it is unavailable
on another host, record the replacement and rerun isolation checks. Readers must
not see source answers, labels, held-out data, credentials, a Docker socket or GPU.
The active Exp8 v1 conversations use 0.35 GPU memory utilization and 32 concurrent
sequences, retaining full token caps; these runs can take hours per pass. A separate
64-sequence fork has no active editing or benchmark cases; its separately approved
zero-answer Qwen and Gemma capacity checks are queued and pending. Every heavy
pass, held-out shard and new capacity worker uses
TASK/reconstructed/exp8/gpu.lock. Leave unrelated host workloads untouched.

CODEX SETTINGS AND CONTEXT AUDIT

Every scored conversation requests model gpt-6-astra, context window 872000 and
auto-compaction threshold 1000000. Effort: Exp1/2/4 high; Exp3/5/6/7 xhigh;
Exp8 max. Detailed reasoning summaries are requested for Exp1/2/5/7; Exp3/4/6/8
do not add that override. Exact argv are retained in each run/turn record.

Launches use --no-daemon exec --strict-config --ignore-user-config --ignore-rules
--skip-git-repo-check --json, the released harness/codex_config.toml with tools
disabled, original neutral model instructions, project_doc_max_bytes=0,
skills.include_instructions=false, skills.bundled.enabled=false and memories=false.
Turns resume their actual new thread UUID. Actual saved CLI rollouts are checked
for exact user messages, model/effort, unexpected developer instructions and tools.
The current model alias/backend is not a pinned historical model snapshot.

Early Exp1/4 attempts with unintended local context were quarantined under
outputs/replication-invalid-initial. Current scored cases use audited clean runs.
There is one valid scored reply per one-shot case, with no performance-based
regeneration or tuning. Failed attempts and recovery evidence remain separate.

MAIN COMMANDS

These are entry-point examples after provisioning, staging, path review and
preflight. ROOT/SRC/CODE/TASK below are literal placeholders for the paths above;
substitute actual paths in every command. CPU/GPU Python must match each experiment.
Do not run this whole section against the existing active output tree.

Exp1/2/4/5 one-shot prompts, on the controller:
  python3 ROOT/work/run_published_prompts.py \
    exp1_which_game/runs/breakout exp1_which_game/runs/freeway \
    exp1_which_game/runs/pacman exp1_which_game/runs/pong \
    exp2_diffusion/runs/circle exp2_diffusion/runs/square \
    exp2_diffusion/runs/triangle \
    exp4_fix_car/runs/oneshot_repair/reply \
    exp5_write_from_scratch/runs/asterix/reply \
    exp5_write_from_scratch/runs/breakout/reply \
    exp5_write_from_scratch/runs/seaquest/reply
  python3 ROOT/work/audit_fresh_runs.py
The launcher refuses existing case destinations. It copies exact published
prompt/instruction bytes and settings. The audit is the original early-stage
one-shot utility: run it in a clean one-shot workspace before adding interactive
results. It recursively assumes every run.json has that schema/source layout and
must NOT be run indiscriminately over today's mixed outputs. Its historical
not_claimed/status text does not describe later reconstructed evaluations.
Exp1/2 scoring is documented label/alias comparison, not an unpublished author
scorer. Exp2 category agreement does not validate the incorrect hollow-shape prose.

Exp3 (controller, from CODE, after sandbox staging/preflights):
  python3 -m unittest test_exp3_score.py
  python3 -u exp3_runner.py
  python3 exp3_score.py
Six independent conversations use the exact 40-array released blocks.npz files.
exp3_sandbox.py is staged on the Spark with opaque archive paths. Its required
sandbox-preflight.json and sandbox-fresh-process-check.json must be generated by
actual local isolation tests, not copied from this bundle as evidence for a new
host. See EXP3_PROTOCOL.md for the required tests and feedback contract.

Exp4 (Spark, dedicated JAX environment):
  python evaluate_exp4.py --source TASK/source \
    --fresh-reply PATH_TO_FROZEN_FRESH_ANSWER \
    --kinetix PATH_TO_PINNED_KINETIX --output NEW_RESULT_DIRECTORY
The rounded base comes from the prompt, not the full-precision PPO checkpoint.
Validate the original reply against released edited_actor.npz, then run base,
published edit and fresh edit: 256 search plus 1024 test episodes per policy.

Exp5 (Spark, CPU environment):
  python exp5_evaluate.py prepare --source TASK/source \
    --fresh TASK/reconstructed/exp5/fresh --out NEW_RESULT_DIRECTORY
  python exp5_evaluate.py evaluate --source TASK/source \
    --out NEW_RESULT_DIRECTORY --episodes 1000 --workers 8 \
    --seed 20261102 --max-steps 10000
Stage immutable fresh replies under fresh/runs/GAME/reply/answer.md.
Validate all three published hand.npz controls before evaluating six
policies. Same-argument resume preserves completed episodes and verifies protocol.

Exp5 reconstructed PPO baseline campaign (prepared, NOT launched at this snapshot):
The original training code and PPO checkpoint bytes remain unavailable. The new
32-run campaign uses the published actor 600->32->32->6/tanh, lr 5e-4 with annealing,
64 environments, 128 rollout steps, 4 epochs, 8 minibatches, gamma .99, GAE lambda .95,
PPO clip .2 and entropy coefficient .01. Its separate 600->64->64->1 critic, Adam
epsilon 1e-5/betas .9,.999, orthogonal initialization, value coefficient/clip,
gradient clip .5, advantage normalization and precise linear schedule are explicit
reconstructions guided by CleanRL commit fe8d8a03c41a7ef5b523e2e354bd01c363e786bb.
The upstream MIT license is reconstruction/LICENSE-CleanRL; full defaults and
reference source hashes are in metadata/exp5-ppo-campaign-plan-v5.json.

At 100M nominal steps: Breakout and Asterix use seeds 0–9, Seaquest seeds 0–2.
At 500M nominal steps: all three games use seeds 0–2. Each budget starts anew;
500M is not a continuation of 100M. Full 8192-step updates produce 99,999,744 or
499,998,720 actual steps: 6,799,982,592 total across 32 runs (6.8B nominal).
Training uses native MinAtar terminal/reset behavior, no invented training time
cap, no observation/reward normalization, all 6 actions and float32 CPU Torch.
Evaluation separately uses the already frozen 1000-episode/10000-step protocol.
Retain all 32 final policies/scores; report the maximum seed mean only as a labeled
best-seed selection, alongside every seed. Its episode SE does not account for
selection over seeds.

Conditional startup approval exists in metadata/exp5-ppo-startup-review.json,
bound to plan SHA 7750dc4fa3522386e46be40259ffb67c7cd8b8cb1c8dae1e585e9f15622811b3.
Launch must wait until all 15 Exp7 readers finish and reader CPU workers clear.
This is scheduling/review state, not a completed training result. The approved
Spark schedule has at most 4 workers, each single-threaded on a distinct core from
5–8, niceness 10, with GPU disabled. These are measured 3.9GHz Spark cores, not
portable core numbers or exclusive reservations. A different host needs a new
runtime/affinity review and plan. GPU editing can continue independently.

On the Spark, after that launch condition is verified:
  mkdir -p TASK/results/reconstructed-exp5-ppo
  nohup TASK/.venv/bin/python -u \
    TASK/reconstructed/exp5-ppo-feasibility/exp5_ppo_campaign.py run \
    --plan TASK/results/exp5-ppo-feasibility/campaign-plan-v5.json \
    --output TASK/results/reconstructed-exp5-ppo --workers 4 \
    > TASK/results/reconstructed-exp5-ppo/campaign-run.log 2>&1 < /dev/null &
After all 32 final actors are frozen, evaluate with the same durable launch form:
  nohup TASK/.venv/bin/python -u \
    TASK/reconstructed/exp5-ppo-feasibility/exp5_ppo_campaign.py evaluate \
    --plan TASK/results/exp5-ppo-feasibility/campaign-plan-v5.json \
    --output TASK/results/reconstructed-exp5-ppo --workers 4 \
    > TASK/results/reconstructed-exp5-ppo/campaign-evaluate.log 2>&1 < /dev/null &
Use unique log filenames on a later launch to retain previous logs. The campaign
lock is inherited by children, so an orphan worker still excludes a second
launcher. Same-plan restart skips frozen cases and resumes full actor/critic,
optimizer, RNG and 64-environment state from the latest durable checkpoint.
Checkpoints occur every 100 updates; up to 100 updates may replay if interruption
occurs after update 100 computes but before its checkpoint pointer publishes. Latest/previous payloads and all immutable checkpoint receipts are
retained. Full campaign timing is not guaranteed by the short feasibility sample.

Exp6 (controller; see EXP6_PROTOCOL.md and preflight entry point):
  python3 exp6_preflight.py
  python3 exp6_runner.py --initialize-only
  python3 -u exp6_runner.py --budget 0
Repeat independent cases for budgets 10,25,50. Final scoring on the Spark:
  JAX_PLATFORMS=cpu PYTHONPATH=TASK/reconstructed/exp6/deps \
    python exp6_evaluate.py --candidate FROZEN_POLICY_DIRECTORY \
    --opponent HELD_OUT_RELEASED_POLICY_DIRECTORY --games 2000 --seed 31337 \
    --output NEW_RESULT_DIRECTORY
Omit --opponent for uniform random. Keep practice seed1 separate from held-out
same-size PPO seeds0/2, and score published controls separately from fresh policies.
Policies are directories containing arch.json and the 14 tensor .npy files.
The additional bootstrap/independent_exp6_scoring.py accepts budget10,25 or50
after that budget's six evaluations are complete. Install it at ROOT/work and
run python3 ROOT/work/independent_exp6_scoring.py 10. It checks episode aggregates,
outcome/action consistency, paired seeds and provenance against the prior audited
budget0; it does not replay all boards through a second Go engine.

Exp7 training (Spark CUDA environment, six sequential fine-tunes):
  python exp7_retrain.py --source TASK/source --base PINNED_GEMMA_DIRECTORY \
    --output TASK/results/reconstructed-exp7
  python exp7_audit.py --source TASK/source \
    --output TASK/results/reconstructed-exp7
Uses unchanged author train.py/eval.py/make_release.py, splits/seeds100/101/102,
batch8, epochs3, lr2e-5, answer-only loss, frozen embedding/output and bf16 autocast.
Source/base/model/evaluation hashes guard resume. The resulting six models are new
checkpoints; they are not the authors' unavailable originals. Wrong-follow rates
measure teacher-forced total-answer-logprob preferences, not free-generation rates.

Exp7 readers (controller, after role-hidden staging and preflight):
  python3 exp7_preflight.py
  python3 exp7_attention_preflight.py
  python3 -u exp7_runner.py --workers 2 --round weights_only_1
  python3 exp7_score.py
Other round names: weights_only_2, weights_only_3, weights_only_4, attention_tool.
The setup creates three opaque A/B cases per round. Readers receive only their
authorized release/prelude; never the private roles, facts, training logs or evals.
The attention sidecar returns raw attention only. For causal follow-up, after all
ten eligible final selections are frozen: exp7_patch_freeze.py,
exp7_patch_gpu.py --requests FROZEN_REQUESTS, and exp7_patch_audit.py. Follow
EXP7_PATCH_PROTOCOL.md, including unchanged-baseline parity before any transplant.

Exp8 (controller, from CODE; prepare-only stages inputs without a model call):
  python3 exp8/controller.py --case p10c --base ROOT --source SRC \
    --output ROOT/outputs/replication/reconstructed-exp8 --prepare-only
  python3 -u exp8/controller.py --case p10c --base ROOT --source SRC \
    --output ROOT/outputs/replication/reconstructed-exp8
Published cases are p10c,p25c,p50c,g05b,g10b,g25b,g50b; each starts from the unedited
base. Current active p10c,p25c,g05b conversations stay on v1 for their full budgets.
The other four remain unstarted; do not launch them with either profile until
the specific new startup review and existing three-kernel scheduling gate allow it.
Stage approved exp8 code, task-owned dependencies, pinned models and practice
arrays on the Spark first. exp8/prepare.py builds the Qwen text export. Native
teacher/generation/activation/edit-drop parity and actual kernel-restart tests
must pass. Hash-bound startup-review.json is required before launching a case.

Exp8 held-out data (trusted Spark process, never an editor mount):
  python exp8_bench/prepare.py --source TASK/source --output BENCH_DIRECTORY
  python exp8_bench/prepare_gpqa.py --source TASK/source --bench BENCH_DIRECTORY
The GPQA command expects the separately downloaded/verified CSV at
BENCH_DIRECTORY/raw/gpqa/gpqa_diamond.csv and download.json beside it. The
included metadata/gpqa-download.json records the required schema, official URLs
and exact source/archive/CSV hashes; regenerate it from your verified retrieval.
prepare.py handles the other nine downloads through huggingface_hub, reading the
retrieved files with PyArrow or JSON.
Use the exact CLI options in --help for control rebuild, checkpoint assembly,
data/provenance audits and scorer checks; retain all generated manifests. Rebuild
seven published controls with unchanged author apply_edit.py before comparing.
After a separate benchmark startup review, a bounded example is:
  python exp8_bench/evaluate.py --profile gemma --model PINNED_GEMMA_DIRECTORY \
    --bench BENCH_DIRECTORY --set math500 --shard 0/63 --source TASK/source \
    --runtime TASK/reconstructed/exp8/code --output NEW_SHARD_DIRECTORY \
    --review BENCH_DIRECTORY/startup-review.json
The active v1 evaluator has at most 32 answers per shard; all Gemma final comparisons
keep this partition. Problem sample counts and full token caps remain unchanged.
The first Gemma MATH500 shard 0/63 completed and passed all 32 actual-record checks,
same-code grading/tokenizer replay and independent root review. These checks use
bootstrap/audit_exp8_shard.py, as specified by that script's --help. Aggregate
complete shards with exp8_bench/aggregate.py. Build fresh checkpoints only from
hash-verified frozen submissions using exp8_bench/build_checkpoint.py. The
aggregation checks completeness/identity; it is not an independent regrading.

The approved ready Gemma campaign uses exp8_campaign/campaign.py on Spark. It covers
398 fixed shards and 12,638 answers: base MATH500+GSM8K and MATH500 for four published
controls. The identical completed first 32-answer shard is verified and reused,
leaving 397 new shards and 12,606 answers. Fresh Gemma families wait frozen submissions.
The single dispatcher started at 2026-11-03T12:26:40Z; it verified that first receipt
and queued published-g05b MATH500 shard 0/63. No new campaign worker had started at launch
verification. This is a running campaign, not a complete benchmark score.

  CUDA_PYTHON TASK/reconstructed/exp8-gemma-campaign/code/campaign.py prepare \
    --task TASK --output NEW_CAMPAIGN_MANIFEST
  CUDA_PYTHON TASK/reconstructed/exp8-gemma-campaign/code/campaign.py run \
    --manifest EXACT_CAMPAIGN_MANIFEST --review EXACT_CAMPAIGN_APPROVAL

The reviewed task manifest is under
TASK/results/reconstructed-exp8/campaign-preparation/gemma-ready-v1; state, logs
and immutable per-job receipts are under
TASK/results/reconstructed-exp8/campaigns/gemma-ready-v1. Exact checkpoint weights,
full metadata and variant output namespaces are checked before invocation and
cache reuse. Base/published identities stay separate even if hashes coincide.
The dispatcher holds a campaign lock and passes both that FD and the canonical GPU
lock FD to the unchanged v1 evaluator's existing --gpu-worker entry point. It
releases the GPU lock between shards. Completed outputs recover without inference;
incomplete/ambiguous v1 attempts stop for review and are never automatically rerun.
Five implementation tests and six independent negative tests passed, including a
real surviving-worker/parent-exit exclusion test. The limitation concerning killed
workers with surviving internal vLLM children still applies. Historical approval,
exact model/partition manifest and root launch verification are bundled as metadata.

SEPARATELY STAGED EXP8 RESOURCE PROFILE — NOT AN EXPERIMENT RESULT

The source snapshot at 66f3553f358dfeace0421ff851d2eb3ad5edf765 includes
exp8_s64/ and exp8_bench_s64/, plus plans/exp8-s64-v2/. The original exp8/ and
exp8_bench/ production files remain unchanged. The new namespaces on Spark are
TASK/reconstructed/exp8-s64-v2/code and TASK/reconstructed/exp8-bench-s64-v2/code.
They reuse the existing read-only CUDA environment, CPU/scoring overlays, NLTK
resources, pinned models, practice files and immutable Docker image. They do not
replace active controllers or copy existing conversations into a new profile.

At this package snapshot, the original deployment review and a separately reviewed
JSONL correction bind all deployed source bytes, all 24 unchanged v1 production
hashes, practice/base hashes and dependency/vendor/NLTK/source context. Qwen-only
and Gemma-only zero-answer capacity checks were queued at 2026-11-03T10:50:15Z and
12:14:56Z; neither capacity report existed at the latest recorded check. The real
first Gemma 32-answer shard logged 2,789,059 KV tokens and 272.37x capacity at 10,240 context;
that legitimate 32-sequence observation is distinct from a 64-sequence initialization.
No v2 editor or Qwen benchmark case has launched. The three live editor kernels stay
on v1. The first 32-record benchmark gate passed; the separate ready-Gemma v1 campaign
is described above. Reviews and one-time capacity approvals are historical evidence,
not transferable approval for another host or source revision.

The new resource profile uses 64 concurrent sequences at the same 0.35 GPU fraction,
BF16, eager execution, 4096 batched tokens, chunked prefill and no prefix caching.
All prompts, instructions, model revisions, token/pass/sample budgets, temperatures,
thinking/presence settings and seeds remain unchanged. Runtime identity:
  config e93f222c25ba92276fd15c6577671e65d608aefd24b2c63a143566e2f05c92e0
  source aaed712c22d05000d8a437b709e7b8a60f97fe7a0892916cfa86aa7e2e44b918
Benchmark context:
  f98539a3d602df0cbc145a83604d22cb5001b1bc93af67dcd8eb9dfe7d64c706
The full context binds every runtime/configuration/vendor hash, dependency version,
NLTK file and original grader/helper source. Any changed path or runtime requires
new provenance/review; do not modify these manifests to bypass their guards.

The benchmark context changed only for two physical-JSONL-line readers before any
v2 benchmark launched. A valid GPQA string contains a raw Unicode U+2028 character;
str.splitlines() incorrectly treated it as another record. The corrected readers
preserve every original byte/record, all 10 datasets and all 206 grouping proofs.
The v1 scientific readers were unaffected. A matching read-only auditor correction
fixed the same issue in practice GSM text; its original failed audit is retained,
and all 256 first-forward records were subsequently checked without new inference.
Corrected forward/s64 audit helpers and regression tests are in bootstrap/.

Only untouched p50c,g10b,g25b,g50b may later use the new editor profile. Per-case
startup approval must bind the initial prompt, source/config, base and actual
architecture-specific capacity. Qwen evidence cannot approve Gemma. Prepared v1
manifests remain intact, with an external supersession index for any later approved
replacement. The global three-kernel limit remains a root-managed scheduling gate.
New local case results use ROOT/outputs/replication/reconstructed-exp8-s64-v2.

The reviewed capacity utility receives an exact model/profile/output-bound approval:
  CUDA_PYTHON TASK/reconstructed/exp8-s64-v2/code/test_capacity.py \
    --profile qwen --model PINNED_QWEN_TEXT_DIRECTORY \
    --output UNIQUE_CAPACITY_REPORT --review EXACT_CAPACITY_APPROVAL
It initializes an engine but calls no generation API. The recorded task launched
it through bootstrap/run_exp8_capacity.py, which retains queue/start/exit times,
PIDs and logs under TASK/results/reconstructed-exp8-s64-v2/capacity. Actual engine
context must cover the complete requested context and its reported cache must fit
the requested concurrency. Failure stops before generation; tokens and samples
are never shortened automatically. Queued Qwen and Gemma requests require 34816 and
10240 context respectively; use the matching model/profile-specific approval.

The actual locked file descriptor is inherited by each new GPU worker. A real
killed-parent/surviving-worker test verifies exclusion until that worker exits.
This test does not establish FD inheritance by every vLLM-internal descendant
following a simultaneous worker SIGKILL; installed vLLM was not modified. The old
v1 lock implementation stays frozen. New runtime/recovery tests 26 and benchmark
tests 14 passed on CPU with CUDA hidden; the narrower capacity-approval checks also
passed after the final two test-prefixed files changed. Historical v1 GPU smoke
reports remain historical. The JSONL correction additionally passed 13 benchmark/
recovery checks, six auditor checks and exact pinned-data/grouping checks, without
inference. No new v2 experimental generation result is claimed.

All seven Qwen final variants may later use the fixed new 64-answer partition, only
if the entire comparison family remains unstarted and receives a new review. The
actual-ID grouping proof retains every ID/sample/seed tuple and contains 206 shards
per variant rather than 406. Every Gemma final remains on v1: 650 shards total. The
combined plan has 2092 shards and 110910 answers, with every sample count unchanged.
New Qwen evaluation outputs use TASK/results/reconstructed-exp8/evaluations-s64-v2.
The new evaluator requires --profile qwen --variant REVIEWED_VARIANT and the exact
set/index/stride from its grouping proof; --output must match that fixed namespace.
It rejects legacy Qwen jobs, mixed partitions, changed checkpoint hashes and missing
shards. Larger shards can generate up to 2097152 tokens and delay queued editor
passes longer. No measured throughput gain is claimed.

The new benchmark persists and fsyncs full returned answers before constructing the
grader. Raw receipts bind job, source, model, prompts, sample order and payload.
A grading interruption reuses the exact answers without loading a generation model;
an integration test proves a single generation call across interruption/resume.
Grading attempts and incomplete outputs remain separate. Completed publication is
recovered from immutable files; differing existing bytes are never overwritten.
If a crash loses returns before durable capture, the run stops for diagnosis. An
explicit retry review must bind the job and every retained incomplete-attempt hash;
it cannot replace complete retained answers. Aggregates verify raw and graded hashes.

Different batching can change generated tokens despite equal seeds. Cross-budget
editor comparisons consequently include a resource-profile difference if new cases
activate. All Qwen final comparisons will use one new profile, and all Gemma final
comparisons keep v1. Record this limitation with every associated result.

REVIEW, RESUME AND ARTIFACT CONTRACT

Interactive Python is independently reviewed before execution, with approval
bound to the exact code SHA-256. This review applies experiment rules without
editing generated code, supplying answers or changing budgets. Do not approve
code merely because its syntax parsed. Context audit failure stops the case.
Existing successful turn/pass receipts are reused only after identity checks;
do not delete state to obtain another model answer. Exp8 restart keeps files and
edits, loses variables, starts a new queue epoch and notifies the model. Final
edits are copied to a unique frozen submission directory with per-file hashes.

Exp7's narrowly defined CLI empty-output recovery accepts only one retained
nonempty completed message, matching the actual rollout, followed by an empty
message; it preserves failed artifacts and uses the exact existing response.
It never regenerates. Other failures need diagnosis; no generic blind retry rule
is authorized. For any new runtime adaptation, retain old code/artifacts and
record a fresh review and provenance, rather than rewriting old hashes.

Local delivered artifacts: ROOT/outputs/replication/exp{1,2,4,5}_*/runs and
reconstructed-exp3 through reconstructed-exp8, plus reconstructed-exp7-readers.
Per-turn folders retain prompts, answers, CLI events, context audits, code reviews
and raw tool outputs. Per-experiment protocols, summaries and independent reviews
identify which measurements are completed. Remote large weights, frozen edits,
activations and full evaluation records remain under TASK/results; Exp7 models
remain under TASK/results/reconstructed-exp7. Never expose held-out records or
private role metadata to ongoing reader/editor conversations.

MATERIAL DIFFERENCES TO KEEP IN EVERY REPRODUCTION CLAIM

* Original private evaluators/drivers and some checkpoints were not released.
  Current Codex/backend and GPU binaries differ; exact historical randomness is
  not recoverable. Container limits and recovery behavior are documented choices.
* Exp4 uses JAX0.9/Python3.12 instead of author JAX0.4.38/Python3.10. Stored edit
  values are float64; inference is float32. Reconstructed published tensors are
  numerically equal; 3145 entries differ only by the sign bit of zero. This is
  not a claim of byte-identical reconstruction of that NPZ.
* Exp5 uses reconstructed seeds and simulator wrapper. Original PPO checkpoints
  are unavailable; their archived scores remain references. The separate full
  32-run PPO reconstruction is prepared and conditionally approved, with explicit
  unknown-default choices. It must not be described as recovered author weights
  or completed training until actual frozen outputs and evaluations exist.
* Exp6 substitutes disclosed practice/held-out opponents and preserves prompt
  diffs. PGX version and PRNG splitting are reconstructed.
* Exp7 retrains once per seed using unchanged published code. Checkpoint bytes
  and behavioral rates can differ. Reader driver/nudge thresholds, attention
  precision and causal transplant/control semantics are documented reconstructions.
* Exp8 reconstructs seed mapping, persistent kernel, forward service, handoff
  every40 replies, explanation reminder ceiling and truthful notice adaptations.
  Generation activations replay actual consumed tokens through Transformers;
  float16 captures aggregate in float32. Final hidden residuals are pre-final-norm
  following the prompt; the unavailable original implementation is ambiguous.
  Current vLLM consumes two Gemma BOS tokens from rendered prompt strings; actual
  IDs are retained. The authors' historical BOS count cannot be established.
* Exp8 final evaluation reconstructs private MMLU/GPQA prompt helpers and fixes
  langdetect seed0. Three original normalized MMLU/practice overlaps are retained
  and reported, not filtered. Bounded shard grouping differs while full sample
  counts/caps remain. The separate 64-sequence candidate and 64-answer Qwen partition
  require prospective review and may introduce batching differences; active cases
  and Gemma evaluations stay on v1. Published controls, fresh edits and archived reference
  values always remain distinct. Full details are in the bundled protocols.
