Offbrief receipts

Offbrief — "production traffic is the only witness your agent didn't write"

Offbrief is a release gate for changes that coding agents make to our synthetic patient API. A human opens one issue; GitLab Duo flows write the contract, the code and the verdict; the warden runs the sandbox replay and egress check, merges, deploys through a canary and records every step in a hash-chained ledger. This page is generated from that ledger.

At a glance (29 loops)

Start here: three recorded runs

Production loops (12)

Each row is one issue. Links go to the GitLab artifacts; the loop number opens the full timeline. Stopped loops halted fail-closed (early development loops among them) and never reached 100%.

Benchmark runs (17)

Fixture merge requests from benchmark/fixtures/, evaluated by the same sandbox and gate and never merged. A blocked benign fixture would be a false alarm; a passed harmful one a miss (docs/benchmark.md has the full table).

Verify it yourself

Every ledger record is JSON with a seq, the prev_hash of the record before it and its own hash: the sha256 of the record without hash, encoded as canonical JSON (sorted keys, no spaces, UTF-8). Changing any record breaks every hash after it. Each loop page links its full chain as loops/<id>.json, the records exactly as stored.

1. Check a chain with nothing but Python 3:

import hashlib, json, sys
doc = json.load(open(sys.argv[1]))          # a downloaded loops/<id>.json
prev = "0" * 64
for i, rec in enumerate(doc["records"], 1):
    body = {k: v for k, v in rec.items() if k != "hash"}
    canon = json.dumps(body, sort_keys=True, separators=(",", ":"), ensure_ascii=False, allow_nan=False)
    assert rec["seq"] == i and rec["prev_hash"] == prev, f"broken link at seq {i}"
    assert hashlib.sha256(canon.encode()).hexdigest() == rec["hash"], f"hash mismatch at seq {i}"
    prev = rec["hash"]
assert prev == doc["head_hash"]
print("loop", doc["loop_id"], "ok:", len(doc["records"]), "records, head", prev)

Or from a clone of the repository: .venv/bin/python -m offbrief ledger verify --export 14.json. To check a receipt against its evidence package and the chain: .venv/bin/python -m offbrief receipt verify --receipt receipt.json --ledger-export 14.json (it downloads the evidence package the receipt names and recomputes its sha256).

2. Check the root. The ledger root is the sha256 of these sorted loop:head lines joined by newlines; maintenance.sh freeze records it at the deadline. Current root: b602a93c783200b73abe339c22457433dead0564bd42e4d3f41e66397aa00f1c

1:c67e1b3d50cda758095d6800e09f6bc533636d6e08b8a4a105a4ee23b53c6412
10:a3d87ebc4d1b784c8954668a676d9ebe1645cd688e295dfe3bd23fc5a5110930
12:77e5a41a0780e1f956e46ecb3a1004118d22bf4d122ad7c9a7323c9d564623dc
14:6ea1a5f0bd37b062592a2fcd80f055e485078b668e81135f46a2fdb4aef84708
16:17d2792081cfe9acc6e057be7371f7dfc1d63e20034773fe3b521458a7ad0bd8
18:250b9538ab9e3616d390cee42832dfde785227ce308d4107f89db1be4f576ffc
19:4c23265b503ac8cf63755011332fddc2d8e08b66df9578e25f4ea8333f43127a
2:f5c1970634720e0493f6bf0ef13a92cc66f68193512477a50b21e9baaa5f233c
20:cb46a74c71437d1786a8fbcfd64fc8d6852a58e17dc80ac9960825c24b996c88
22:18a389a0d3993009b52f52c6cb35f389332324a6a32d724837b1bbedcea82e7e
23:c9ef819665d93dd11ad9dbd2fdcc378fd918284aa3a848c20e1db0da8a2abe43
24:d86134cb451f54d2f09c0a5d983dd1b0cdb37cd5e0da559f59c8ad01b6a2db3e
25:97f2caf27eb6fca765831fed8e052539dcae88515716f954a183d3adb140da30
26:d87b140bcd2cfa65048b4182c7fb7000adbba735dabda0a13f0e0b2f4d286b8b
27:6d7f2df39edfa1cf918f2210c44406a97b1c863f791782610318d44e97129e99
28:78ee7470c90b7191d064408209dd61e410c93673145bb9318835db7daab7c8e5
29:5c6ee40d5b9041e326347585a1d8f427b04573bcd64ec4d053ca9f0fc5e4a880
30:02b39ea25289873a63391113db0cb573f3e8bcf1de807f797b29b59eb1038489
31:37357ff48c95fc9e72a420e81aee18b92605e67c9cb82e50af7d7e754406e072
32:39ef9631f9f09a6d695a401f1f2c1892ea11c6d5b7b82653f0eed0fe3b667566
33:b1ad170f25d6fa2c1cf4d978c01f9817e36fac027c6231f157ed034820f6797d
34:72746033bed6961f67fd63f6287764f696609d5c18891b1a4c7d6b4f22153f91
35:7082ba1d783d46e5e073386c21f4d57d0de99c867b7952073d0cbab06b0b61bd
4:ade3b79bf771e5ee1a83c392223b4451c5d12e55a91bdc448abef3f1c23c4a37
5:6dcdaff0ceaad0c641efe120085c7c37ed3f2702b0974cbb8c05433a0d1eb40d
6:b55962043667baa170cbe2599a12a4ee6381f8c3d98905a3c98495457dd35abd
7:76a2e8985f125d720035581c9f1ee94c9d86812c2ff4cac42c68d28adc31cd9f
8:d7af04fe198e3e6caece8eb2e18fb2a198d401bd276b99a98d774ce1ceff1b2f
9:3b02c89971ba58e2185e99c1c520d3879d164589911fa75cd7f7059a85294bc8
global:13a86d927254d13b8e3ebd7411b89fd7cd4b3e5d5252c976cd333e50f758a374

import hashlib
lines = open("heads.txt").read().split()     # the loop:head lines above
print(hashlib.sha256("\n".join(sorted(lines)).encode()).hexdigest())

Chains published: 29. The global chain (maintenance anchors) is not published: its records name the operator's cloud account. Its head hash is in the list above and its maintenance rows are summarised below.

Global ledger

Head seq 39, hash 13a86d927254d13b8e3ebd7411b89fd7cd4b3e5d5252c976cd333e50f758a374; stable digest sha256:2ae6b1539e302d29b41db26e63d4cf159e5f89dfca26d3aaea17190a258a2f40.

Maintenance pushes (history anchors)

Rejected authenticated calls (1)