Dispatch: source record
Check every recorded answer against the original verifier. No model run is needed.
Replay the verification
Use curl, shasum and Node.js 22 or later. The commands create a new directory and verify all 21 files (64,595 bytes). Review the downloaded code before executing it.
Download the reviewed files and verify their hashes
set -eu
# Use a new directory. Existing directories are never overwritten by this recipe.
mkdir 'dispatch-review'
cd 'dispatch-review'
mkdir -p 'baseline'
mkdir -p 'baseline/archive'
mkdir -p 'baseline/data'
mkdir -p 'baseline/exports'
mkdir -p 'baseline/policy'
mkdir -p 'candidate'
curl --fail --silent --show-error --max-time 30 'https://evals.ax/research/artifacts/dispatch-2026-09-08/README.md' --output 'README.md'
curl --fail --silent --show-error --max-time 30 'https://evals.ax/research/artifacts/dispatch-2026-09-08/baseline--AGENTS.md' --output 'baseline/AGENTS.md'
curl --fail --silent --show-error --max-time 30 'https://evals.ax/research/artifacts/dispatch-2026-09-08/baseline--README.md' --output 'baseline/README.md'
curl --fail --silent --show-error --max-time 30 'https://evals.ax/research/artifacts/dispatch-2026-09-08/baseline--archive--dispatch-2025.md' --output 'baseline/archive/dispatch-2025.md'
curl --fail --silent --show-error --max-time 30 'https://evals.ax/research/artifacts/dispatch-2026-09-08/baseline--data--catalog.json' --output 'baseline/data/catalog.json'
curl --fail --silent --show-error --max-time 30 'https://evals.ax/research/artifacts/dispatch-2026-09-08/baseline--exports--customer-holds.csv' --output 'baseline/exports/customer-holds.csv'
curl --fail --silent --show-error --max-time 30 'https://evals.ax/research/artifacts/dispatch-2026-09-08/baseline--exports--order-lines.csv' --output 'baseline/exports/order-lines.csv'
curl --fail --silent --show-error --max-time 30 'https://evals.ax/research/artifacts/dispatch-2026-09-08/baseline--exports--orders.csv' --output 'baseline/exports/orders.csv'
curl --fail --silent --show-error --max-time 30 'https://evals.ax/research/artifacts/dispatch-2026-09-08/baseline--exports--payments.csv' --output 'baseline/exports/payments.csv'
curl --fail --silent --show-error --max-time 30 'https://evals.ax/research/artifacts/dispatch-2026-09-08/baseline--exports--stock.csv' --output 'baseline/exports/stock.csv'
curl --fail --silent --show-error --max-time 30 'https://evals.ax/research/artifacts/dispatch-2026-09-08/baseline--policy--dispatch.md' --output 'baseline/policy/dispatch.md'
curl --fail --silent --show-error --max-time 30 'https://evals.ax/research/artifacts/dispatch-2026-09-08/candidate--AGENTS.md' --output 'candidate/AGENTS.md'
curl --fail --silent --show-error --max-time 30 'https://evals.ax/research/artifacts/dispatch-2026-09-08/expected.json' --output 'expected.json'
curl --fail --silent --show-error --max-time 30 'https://evals.ax/research/artifacts/dispatch-2026-09-08/final-answers.json' --output 'final-answers.json'
curl --fail --silent --show-error --max-time 30 'https://evals.ax/research/artifacts/dispatch-2026-09-08/manifest.json' --output 'manifest.json'
curl --fail --silent --show-error --max-time 30 'https://evals.ax/research/artifacts/dispatch-2026-09-08/plan.md' --output 'plan.md'
curl --fail --silent --show-error --max-time 30 'https://evals.ax/research/artifacts/dispatch-2026-09-08/replay.mjs' --output 'replay.mjs'
curl --fail --silent --show-error --max-time 30 'https://evals.ax/research/artifacts/dispatch-2026-09-08/result.json' --output 'result.json'
curl --fail --silent --show-error --max-time 30 'https://evals.ax/research/artifacts/dispatch-2026-09-08/summary.json' --output 'summary.json'
curl --fail --silent --show-error --max-time 30 'https://evals.ax/research/artifacts/dispatch-2026-09-08/validation.mjs' --output 'validation.mjs'
curl --fail --silent --show-error --max-time 30 'https://evals.ax/research/artifacts/dispatch-2026-09-08/verify.mjs' --output 'verify.mjs'
shasum -a 256 --check <<'EVX_REVIEWED_SHA256'
41d9db6f0f41676c726c0ba310907228faca458946a2ff27663e97358fc0b38a README.md
2673888cc0aac381b9c3e59bfb847febdc993270315a5f18032fc3036541f79b baseline/AGENTS.md
3144b5851625244274b536252d2e90217be61a77870be50fac1fac11eac26a00 baseline/README.md
aace478d430571107e13253883fb528b4a5d12fbd7cfecb48a0aabe6e94d2109 baseline/archive/dispatch-2025.md
79144510fea8bb4b84b934e926c2a39d513bab4fee2d135a6e58f951b4719025 baseline/data/catalog.json
a544702766a913412659a7b7fee438354f21479df2823e026663a53a8cfee85e baseline/exports/customer-holds.csv
c15ed7bfcf63c872eb261998bd5ad386f1d2e8ea7135f4c2aacb0d902f4a5624 baseline/exports/order-lines.csv
06c9aa9b84a38167cbadd886010b88a6483d516e7018de04d6f4fe8f7b4bcaf0 baseline/exports/orders.csv
9e2e52f354555c0581caecb718d09db1914e9672a8470f181a6a275c3b544ff7 baseline/exports/payments.csv
efd8bf3568ce151d47136bdda6bb39c0f2cec4e826210f45d5bc03f9a58ff681 baseline/exports/stock.csv
ee06e30b385be75e1438d74970fa29d3945d54f9856024795eab17faddb6ddeb baseline/policy/dispatch.md
417642d35cb4d35280856ad183fd72fb66dcfe22ff49863be1ee5010894f6995 candidate/AGENTS.md
4768cefed0a580224985aa870bee2629300eefb6e080c6e5dcb643bbfb4076e4 expected.json
b7309497689155e2d818820b80034420bf32dc32ab8bb5d7f6c1602a129aea8d final-answers.json
8b98cb941603690b7d7c5690d63c76ace9d6f30987c1dbd905db7e7cedef1b6f manifest.json
fa89035748bdf5ae6612cbd0be4b03b7c386d52bed638e319e3134cf4354bf22 plan.md
a17b66424a5edc297f61e22028396a45c8cb3491c432bbfbb5fbf73a1ddc0e10 replay.mjs
bddfb0083a4c4d21515a6921a294fc3ceb3ec6c5f1e489d2e82252e14ea848a9 result.json
c96c1597b682954c51a17ad3c23c21fa2c0ffdaf861a44ccd452a9b2e2c31b1b summary.json
19a27d16ac0a9951101aa83c63d06135f8d7c7ba1a83057d37ee56f56865234d validation.mjs
2de997c9296f184b94e8320d9c9402f6256ccacf818fa2e81b1d0cbe3f311779 verify.mjs
EVX_REVIEWED_SHA256In the same shell, after the download and hash checks succeed:
node replay.mjs > replay.jsonThe four declared controls must match. All eight final answers reproduce their recorded verdicts: 4/4 baseline and 4/4 candidate. The original files stay unchanged. If any download, hash or replay check fails, stop and inspect it; retry downloads from a new directory.
This confirms a reproducible acceptance record. It does not establish a benefit from the source index. The four-pair study reached a ceiling; its declared conservative interval spans −100 to +100 percentage points.
Keep the verifier versions separate
These files retain version 1. The replay also demonstrates its known defect: an equivalent leading-zero money cell is rejected. None of the eight recorded answers used that spelling. The current runnable example uses version 2 and additional controls; those later changes do not rewrite this study.
Original verification record
The final-answer collection is a later extraction, checked against the original trace and output hashes. The result and its verdicts are unchanged.
- expected.json1,975 bytes
SHA-256
4768cefed0a580224985aa870bee2629300eefb6e080c6e5dcb643bbfb4076e4 - final-answers.json12,741 bytes
SHA-256
b7309497689155e2d818820b80034420bf32dc32ab8bb5d7f6c1602a129aea8d - manifest.json8,400 bytes
SHA-256
8b98cb941603690b7d7c5690d63c76ace9d6f30987c1dbd905db7e7cedef1b6f - plan.md2,031 bytes
SHA-256
fa89035748bdf5ae6612cbd0be4b03b7c386d52bed638e319e3134cf4354bf22 - result.json22,116 bytes
SHA-256
bddfb0083a4c4d21515a6921a294fc3ceb3ec6c5f1e489d2e82252e14ea848a9 - summary.json2,444 bytes
SHA-256
c96c1597b682954c51a17ad3c23c21fa2c0ffdaf861a44ccd452a9b2e2c31b1b - validation.mjs2,943 bytes
SHA-256
19a27d16ac0a9951101aa83c63d06135f8d7c7ba1a83057d37ee56f56865234d - verify.mjs694 bytes
SHA-256
2de997c9296f184b94e8320d9c9402f6256ccacf818fa2e81b1d0cbe3f311779
Synthetic warehouse inputs
Only AGENTS.md differed. All other candidate files matched the baseline copies supplied here. These are fictional operational records.
- baseline/AGENTS.md128 bytes
SHA-256
2673888cc0aac381b9c3e59bfb847febdc993270315a5f18032fc3036541f79b - baseline/README.md320 bytes
SHA-256
3144b5851625244274b536252d2e90217be61a77870be50fac1fac11eac26a00 - baseline/archive/dispatch-2025.md195 bytes
SHA-256
aace478d430571107e13253883fb528b4a5d12fbd7cfecb48a0aabe6e94d2109 - baseline/data/catalog.json105 bytes
SHA-256
79144510fea8bb4b84b934e926c2a39d513bab4fee2d135a6e58f951b4719025 - baseline/exports/customer-holds.csv172 bytes
SHA-256
a544702766a913412659a7b7fee438354f21479df2823e026663a53a8cfee85e - baseline/exports/order-lines.csv300 bytes
SHA-256
c15ed7bfcf63c872eb261998bd5ad386f1d2e8ea7135f4c2aacb0d902f4a5624 - baseline/exports/orders.csv433 bytes
SHA-256
06c9aa9b84a38167cbadd886010b88a6483d516e7018de04d6f4fe8f7b4bcaf0 - baseline/exports/payments.csv886 bytes
SHA-256
9e2e52f354555c0581caecb718d09db1914e9672a8470f181a6a275c3b544ff7 - baseline/exports/stock.csv423 bytes
SHA-256
efd8bf3568ce151d47136bdda6bb39c0f2cec4e826210f45d5bc03f9a58ff681 - baseline/policy/dispatch.md2,039 bytes
SHA-256
ee06e30b385be75e1438d74970fa29d3945d54f9856024795eab17faddb6ddeb - candidate/AGENTS.md867 bytes
SHA-256
417642d35cb4d35280856ad183fd72fb66dcfe22ff49863be1ee5010894f6995
Later reproduction tools
- README.md2,037 bytes
SHA-256
41d9db6f0f41676c726c0ba310907228faca458946a2ff27663e97358fc0b38a - replay.mjs3,346 bytes
SHA-256
a17b66424a5edc297f61e22028396a45c8cb3491c432bbfbb5fbf73a1ddc0e10
Provenance and limits
The input and verifier revision is 5255a0cc9b8989fb07e34704d827baf955baca8b. The README records the source of every file category. Publication is limited to exact reviewed bytes.
The original local Mac run was not externally observed. Complete model context and access outside the task directory were not attested. This replay cannot recover them. Hashes identify files; they do not attest execution. Dollar cost remains unknown.
No raw provider traces, host configuration or private authentication source are included. This is a verification download, not a new agent-run distribution.