Protocol · Proof of AI
Execution evidence.
Independent validation.
Proof of AI is a validation approach, not a truth oracle. It is intended to make computational work attributable, testable and permanently traceable — so a result can be argued with rather than simply believed.
What POAI proves
The claims we are willing to make
- 01
A job was specified
The request, boundary and success criteria existed before the work.
- 02
Work was performed
A pseudonymous participant executed the packet within permitted resources.
- 03
Evidence was returned
Inputs, outputs, versions and execution evidence were recorded and signed.
- 04
It was independently tested
One or more validators reproduced or compared the work.
- 05
The outcome was recorded
Agreement, disagreement, challenge and resolution all persist.
What POAI does not prove
The claims we refuse to make
- 01
That an answer is objectively true
Validation tests process and reproducibility, not metaphysical truth.
- 02
That a forecast will be correct
Probabilistic work is evaluated over time, not certified in advance.
- 03
That a model is unbiased
Provenance exposes what was used; it does not sanitise it.
- 04
That every job class is verifiable
Some work is far easier to reproduce than other work. Scalar treats that as a design constraint.
Validation methods
Different work needs different proof
No single validation method fits every job class. The intended design selects a method per job class and records which method was used.
- 01
Deterministic re-execution
StrongestRe-run the packet from recorded inputs and versions and compare hashes.
- 02
Independent reproduction
A different participant performs the same specified task and results are compared.
- 03
Cross-model comparison
Multiple approaches are compared for agreement and divergence.
- 04
Structural checks
Schema, constraint and consistency validation against the job specification.
- 05
Evidence audit
Cited sources, retrieved data and intermediate artefacts are checked.
- 06
Outcome resolution
Over timeWhere reality later resolves the question, recorded predictions are scored against it.
Challenge path
Disagreement is a feature of the protocol
If validators cannot disagree meaningfully, validation is decoration.
- Result returnedwith evidence
- Validationreproduce / compare
- Challengeraised on divergence
- Resolutionadditional validators
- Final stateaccepted / rejected
- Recordedeither way
Rejected work is not deleted. A network that hides its failures cannot be evaluated, and failure data is part of what makes accepted work meaningful. Penalty and incentive mechanics for repeated bad-faith behaviour are in design and are not active.
Next
Provenance is what makes proof reusable
Braid records the lineage; Cortex turns accepted lineage into context applications can use.