falsify · acceptance record AP-2026-001
illustrative · Falsify plays both roles · demo signatures
The record · 23 Sep 2026 · UTC
11:50:23Criterion C1 recorded5ce3c9c3484d
11:50:28Criterion C2 recorded464df2bee6b4
11:50:42Signed plan sealedda55fc4ac5cc
11:51:18Run start, C1 · C214bc38f2bd35 · 231970705630
11:51:37Run final, C1 · C286c224c760c7 · 74415372626b
11:52:20Signed report sealed86f76e446db7
11:55:31Verification statement sealedfbec1be1aa3a
Each line: a public registry record with a signed receipt, an RFC 3161 timestamp and a Rekor transparency-log entry.
What your client receives
at AI handover.
A completed acceptance record, in 60 seconds. A credit model on public data; every receipt on screen is a real public registry record.
1The acceptance plan, signed and sealed
Client and supplier agree the criteria and sign the plan; it is sealed before any run is registered. The criteria become canonical records with an independent time.
Acceptance plan AP-2026-001
Credit scoring model · one system version · one acceptance round
C1Impact ratio, female vs male≥ 0.80
C2Accuracy on the held-out split≥ 0.70
Client decision owner Supplier delivery lead
plan sha-256
sealed record receipted 11:50:42 UTC
2Tests run, results recorded
Each run is declared against its criterion record, then its result is recorded. The report is signed by both sides and sealed.
metricbarobserved
C1impact ratio (sex)≥ 0.80
C2accuracy≥ 0.70
Decision: accepted
report sealed receipted 11:52:20 UTC
3Someone moves the bar afterwards
The record does not stop an edit. It makes the edit visible: change one number and the criterion no longer matches what was receipted.
metric: impact_ratio
comparator: '>='
threshold:
dataset: uci-statlog-german-credit
receipted5ce3c9c3484d8db1ffbc67dc4f3958c9… 11:50:23 UTC
recomputed
4A gap, stated rather than hidden
Falsify checks the record and signs a statement of what was checked, what was not, and what is missing.
Verification statement
Falsify OÜ · signed and sealed 11:55:31 UTC · fbec1be1aa3a
Missing reference: no hash of the trained model artefact is recorded; the system version is identified by the training script’s hash and the seed.
5Anyone can check it, offline
No trust in the supplier, the client or Falsify: one script, no network.
Acceptance Evidence Pack
One acceptance decision, documented: the signed plan, the signed results, the agreed change conditions and a verification statement, each with an independent timestamp.
€3,000 fixed · one system · up to three criteria
first draft within 3 working days of receiving your documents
falsify.dev/acceptance-record
What the record does not show: that the criteria are appropriate, that the results are correct, or that the tests ran after the plan. The runner times are the supplier’s statement.