Loading
Please wait...
Loading
Please wait...
HOW IT WORKS
Multi-perspective deliberation is the easy part. Aequitas adds the hard part: a verification layer that decides whether a verdict has earned the right to be certified, and says “not certified” when it hasn’t.
Every argument entering the pipeline is screened for internal inconsistency before it can shape a conclusion. Flawed inputs are flagged and set aside, so a single bad premise can’t quietly corrupt the result.
A traceable reasoning chain with a machine-computed reliability certificate
Statistical feature-importance scores
See the path from input to conclusion. No black box, no hand-waving.
Every judgment comes with its own record of what was checked and found.
A deep, multi-stage pass designed for regulated review, not a single shot.
The reasoning structure of every judgment is checked for internal consistency with formal methods rather than statistical pattern-matching. When values genuinely conflict, the trade-off is labeled outright instead of being smoothed into a false consensus.
Formal logical-consistency checks across the full argument set
Keyword and rule matching
Goes past surface wording to examine the structure of the reasoning itself.
Flags where an argument contradicts itself across a chain of reasoning.
Genuine value conflicts are labeled, not resolved into false agreement.
Every proposed judgment runs an independent adversarial audit that actively tries to break it. Verification findings outrank the model’s own opinion, a grounding gate demands positive evidence before certification, and an engine-health check means a degraded run can never quietly certify. The result is labeled CERTIFIED or PROVISIONAL with documented reasoning.
An independent audit that can demote a judgment to PROVISIONAL
Manual review after the fact
Independent auditors stress-test each judgment from multiple angles.
When an audit fails, the judgment is revised and re-checked, or returned as provisional.
Every result is graded: CERTIFIED, or PROVISIONAL with documented caveats.
Every decision passes through all three layers in sequence, and the gate at the end has two exits. That fork is the product: certification is earned, or honestly declined.
Bring these to every vendor evaluation, including ours. Our answers are on the record here and verifiable in the product.
| The question | Aequitas’ answer | What to watch for |
|---|---|---|
| What happens when it can’t verify an answer? | It declines to certify and discloses why: an honest “not certified,” never a confident guess | Most tools answer anyway, in the same confident tone |
| How is a conclusion actually checked? | A formal proof engine checks every claim in a certified verdict: mathematically, not by another model’s opinion | A second model grades the first model’s homework |
| What if the system itself degrades? | An engine-health check refuses to certify anything produced in a degraded state | Silent fallbacks keep serving as if nothing changed |
| Is the confidence score a real probability? | Ours is disclosed honestly, and labeled uncalibrated until calibration is fitted and validated | A formula output dressed up as a percentage |
| Can I audit the result afterwards? | Every certified verdict carries its basis of certification and a content hash | Reports reconstructed after the fact |
Walk through a live judgment on your own material: the reasoning, the conflicts it surfaces, and the certificate it produces.