Certified runs
The registry
A place to put a contamination certificate so that other people's certificates can be counted alongside it. One certificate becomes one row, and the server checks the file before the row exists.
PLANNED There is no leaderboard page. The submission routes are live and the counts are published; the comparative view is separate work that starts when a cell meets the rule below.
The rule
Quorum
A cell reaches quorum when it holds at least five listed submissions signed by at least three distinct keys.
A cell is one pair: the run window the agents were scored over, and
the set of arms they were run under. Two runs land in the same cell
when both strings match exactly.
GET /v1/registry/cells reports the two counts and the
boolean for every cell holding at least one listed submission, and it
repeats needs on every row so the counts can be read
against the rule without leaving the response.
A count on its own would let the holder of one signing key reach quorum alone by submitting five of their own runs, so the three fingerprints sit beside it. A fingerprint is the first 16 hex of sha256 over a public key's raw bytes, and it tells you a distinct key signed a file. Three of them raise the floor on independence without establishing it.
Until a cell meets both halves there is nothing to compare, and once it meets them the cells route goes on answering with counts.
The thresholds live in internal/registry/quorum.go as
QuorumListed and QuorumFingerprints, and a
test fails if either moves without the documentation moving with it.
Lowering the bar is a change that shows up in a diff.
Listing
What a listing does, and what it does not
Private is the default
A private submission is yours. It appears in
GET /v1/registry/certificates for your account and
in no count, no aggregate and no page anyone else can reach.
Listed adds two things
One increment to listed and one fingerprint to
fingerprints, in that cell. The metrics are stored
and are served to nobody but you.
To take a submission back off the board, send the same certificate
again with ?visibility=private. The row does not move
otherwise: who registered it and when stay as they were first
written.
One certificate is one row. A second account sending a certificate somebody already registered gets a 409, because a certificate is a public file and counting it once per sender would let a cell reach quorum on one run forwarded five times.
Submitting
How to send one
$ curl -sS -X POST \
'https://api.pit.aqx.llc/v1/registry/certificates?visibility=listed' \
-H "Authorization: Bearer $PIT_KEY" \
-H 'Content-Type: application/json' \
--data-binary @certificate.json
-
The body is the certificate file, byte for byte. Use
--data-binaryrather than-d, which strips newlines, because the digests are taken over the bytes that arrive. - Any paid plan may submit, from an API key rather than a browser session, because the key id is recorded on the row.
- The file is capped at 1 MiB. The golden certificate in this repository is 23,347 bytes.
-
A refusal is a 422 with
certificate_unverifiableand the failure class inerror.param. Nothing is written, so the table holds only certificates this server checked itself. Every class is reproducible before you send anything, withcertverify verifyon your own machine.
The server runs three of the five checks: body_digest,
receipt_body and signature. The other two
compare a certificate against files you hold, and a server holds
neither, so a listing does not stand behind them. A run the harness
marked invalid is accepted; the row records
valid: false and the numbers that go with it.
The boundary
What a certificate proves
It proves
- The body has not changed since the digest was written over it.
- The embedded run receipt is the one the harness hashed, unedited.
- The file came from the key it names, and that key's fingerprint is recorded.
- The trajectory you hold is the one the run recorded, tool call for tool call.
- The run read the corpus release the manifest you hold describes.
It does not prove
- That the agent reasoned from the rows rather than from memory. The arm deltas, the positive control and the reasoning audit bound that.
- That the corpus release is a fair test of the task the agent was given. Nothing in the file scores that.
- That the key belongs to who you think it does. That is the key file's provenance, and a verifier cannot check where a file came from.
- Who a submitted key belongs to. The registry verifies against the copy the certificate carries, and records that key's fingerprint.
Contamination that arrives through a model's weights is bounded by
the control arms and the reasoning audit rather than by cryptography.
A model that remembers March 2023 shows up in the numbers the
certificate carries: a shifted-arm delta that collapses toward zero,
a probe that recovers the true date, an audit finding that quotes a
figure appearing in no row the run served. The signature covers those
numbers, so a leak is something you read in
arm_deltas and positive_control, not in the
signature check.