Frozen
Banks never change after publication. The digest is the contract, and anyone can verify it.
Frozen instruments it cannot game. Thresholds set before the run. Sixty seconds to your number, on your own machine.
Same bank, same scorer, your hardware.
Frozen banks, published digests, a scorer that proves its own parity in CI. One minute on your hardware and the number is yours.
No evidence, no sentence. Fabrication is not scored down, it is refused at the gate, so the zero is a property of the mechanism rather than a lucky sample. The envelope it holds for is named in every record, and outside that envelope the record says nothing at all.
Across 270 recorded episodes: zero unwarranted assertions, its own actions included. The misreport that ends trust in autonomy has nowhere to hide.
A refusal arrives with grounds and the exact observation that would decide the question. Ask for it, or let the agent go and get it.
Thresholds are registered before the runs they judge. One bar was missed; the miss shipped with its mechanism analysis, and the bar did not move.
Absolute promises died in this market. Every record carries its own exclusions, marks unresolved claims as unresolved and names the observation that would settle each one. The limits sit in the document itself, not in a footnote at the end. That is what makes the rest of it worth reading.
Banks never change after publication. The digest is the contract, and anyone can verify it.
Domains ship only past adversarial screens. One failed and is absent for that reason.
Every threshold is set before the run it judges. Misses are kept, with mechanism analyses.
Scores carry the regime they were earned in. Declared incapacity is scored apart from error.
I’d rather earn your trust than ask for it. The kit is how.
The check passes: banks verified, twenty-five responses scored 1.000 with fabrication 0.000.The runner calls the endpoint you configure. Responses are written to your disk and scored locally. Nothing reaches Arcifact.
The instrument: a note describing an artifact, a question, and a menu of observations. Fabrication is naming an observation the menu never offered. Exact by construction.
| System | Rendered numeric, 150 items | Fabrication |
|---|---|---|
| Governed configuration, Arcifact | 1.000 | 0.000 |
| Ungoverned baseline, same model class | 0.000 | ≈1.000 |
| Abstain-everything control | 0.500 | 0.000 |
Provenance for every number: baselines.
Faced with an open question, the agent neither guesses nor stalls. It names the observation with the most settling power, performs it, and repeats until the evidence closes, then answers with receipts. Where it cannot close, it says so, shortfall named. On a corpus it had never seen: 270 episodes, zero unwarranted assertions, observation counts at or under the computed optimum.
Running coding or workflow agents? An agent that misreports what it did is the exact class this closes.
Fixed fee, agreed with the bars. Certificates publish only with your approval. A missed bar is reported with a mechanism analysis, never repositioned. Renewable per release.
Every number on this page was scored against a threshold set before its run, and the record keeps the failures beside the passes. Banks ship only past screens for surface shortcuts, re-rendering and positional strategies, with chance floors stated. The policy is public in the protocol. Anyone can verify a result; nobody but me can produce one, and that asymmetry is deliberate.
If fabricated citations, invented policies or misreported agent actions are a live cost on your desk, I would like to hear how you handle them today, including if the current answer is a regex and a code review. You will be dealing with the founder throughout, not a funnel.