The audit chain — REVEAL

[01 / THE PROBLEM]

The question is never what the model said. It is whether you can prove it later.

A regulator, a client, a works council or your own risk team will eventually ask how a number was produced. Most AI quality tools answer with a screenshot and a confidence percentage. That answer does not survive the second question.

WHAT USUALLY EXISTS

A score in a dashboard, with no record of which rule produced it.
A model that was silently upgraded, so last quarter and this quarter are not comparable.
A prompt nobody kept, in a version nobody logged.
A database row anyone with write access can change without leaving a mark.
Evidence that only exists inside the vendor's account.

WHAT REVEAL WRITES

The interaction as it arrived, with the redactions applied before anything read it.
The rule the output was judged against, quoted in full.
The model, its version and the prompt template that ran.
A hash over all of it, linked to the entry before it.
An export file you keep, and can check without us.

This is not a compliance feature bolted on at the end. It is the log layer everything else writes through, which is the only place a rule survives contact with a deadline.

[02 / CAPTURE]

It is sealed before anything reads it.

The chain starts at ingest, not at analysis. By the time the first module sees an interaction, the personal data is already out and the original is already fixed.

01 · REDACT

Personal data leaves first.

Names, account numbers, card fragments and contact details are removed on the way in, and the redaction map is part of the record. Constitution A4 makes this the default rather than a setting somebody has to remember.

02 · FIX

The input is hashed before analysis.

Channel, language, timestamps and the redacted text are hashed together. Every later entry refers to that hash, so a re-run on edited input is visible as a different chain rather than a quiet correction.

03 · ATTRIBUTE

Each step writes itself as it finishes.

Step results are committed one at a time, with their input, output, timing and confidence. A crashed run resumes from the last committed step, and the same table is what replays the run for an auditor.

[03 / THE CHAIN]

Every output signs itself.

Five records, written with every result, hashed together and linked to the entry before them. Change one field and the chain breaks visibly.

01 · INPUT
The interaction as it arrived. Channel, language, timestamps, and the redactions applied before anything read it.
02 · RULE
The policy line or scorecard item the output is judged against, quoted in full rather than summarised.
03 · MODEL
Which model ran, at which version, with which prompt template. Swap the model and the record says so.
04 · SIGNATURE
A hash over all of the above, linked to the entry before it. The break is in the chain, not in a warning email.
05 · EXPORT
The bundle leaves as a file. Verify it on your own machine, offline, without asking us for anything.

Confidence is written with the result rather than hidden behind it. Below the threshold you set, the output is marked for human review instead of being presented as settled, which is what Constitution A2 asks for.

[04 / CONSTITUTION]

Eight rules the platform enforces, not eight rules in a policy document.

A1 to A8 are checked at the log layer. An output that breaks one does not get written and quietly flagged. It does not get written.

A1

No unreviewed worker evaluation

Nothing that ranks or penalises a named person reaches anyone without a human in the path.

A2

Flag uncertainty

Below the confidence threshold, the result says so and asks for review rather than presenting itself as settled.

A3

No fabrication

Numbers come from a computation, quotes come from the transcript. Where there is nothing, the answer is nothing.

A4

PII protection

Personal data is redacted by default, on the way in, and the redaction is part of the signed record.

A5

Recommendations, not verdicts

The platform reports what the data indicates. The decision stays with the person who has to defend it.

A6

No intent attribution

What was said and what happened. Never what somebody was trying to do.

A7

AI disclosure

Every AI-produced line is marked as AI-produced, in the metadata and on the surface a person reads.

A8

Refuse harmful requests

Blacklists, discriminatory criteria, routes around human review. The platform declines and says why.

[05 / EXPORT]

Take the evidence with you.

The export is the point. A chain you can only inspect inside our product proves that our product agrees with itself. Verification has to work somewhere we are not.

5

RECORDS PER OUTPUT

Input, rule, model, signature, export. Written every time, on every module.

1

FILE

The bundle leaves as one signed archive, per interaction or per period.

0

CALLS BACK TO US

Verification runs offline, against the hashes in the file itself.

A1 → A8

CHECKED AT THE LOG

The constitution is enforced where the record is written, not in review.

The chain is included in every tier of the platform and in REVEAL Local, at no additional cost. It is not an enterprise add-on, because a result nobody can check is not worth selling at any tier.

Chain