Solutions
Produce evidence for every AI action
Turn AI activity into an auditable record mapped to the controls you already report against, instead of reconstructing what happened after the fact.
What you stop doing
Reconstructing what happened. Screenshotting a console. Asking an engineer which version of a rule was live in March. Writing a memo explaining why a control probably operated.
Records to assemble
0
Decisions recorded
240
Written for whoever assembles the evidence
The problem is not that the controls do not work. It is that you cannot show it.
Most AI controls in most organizations do broadly what they were meant to. What nobody can produce is the record — so a quarter goes into reconstructing intent from three systems and a chat history, and the output is an assurance rather than evidence.
- An assurance
- Somebody stating that a control operated, supported by a screenshot taken this week of a setting that may have changed since.
- Evidence
- The control having written down what it did at the time, including the times it refused, sealed before the response returned.
- Why the difference is expensive
- An assurance is challenged and then defended with more assurance. Evidence ends the conversation, which is the entire value of having it.
- The reconstruction trap
- Anything assembled afterwards is a reconstruction, however careful. A reviewer who has seen one before knows that, and discounts it accordingly.
In the sample estate, 240 decisions produced 240 sealed records — 215 authorized, 20 held for a person and 5 refused. The refusals are recorded as completely as the authorizations, which is what makes the set usable: a reviewer’s first question is what the system tried to do, not what it managed.
What the ledger looks like
Sealed when the decision was made, chained so a gap is detectable.
Sealed means no endpoint can edit it — there is no code path that amends a record, which is what makes the word a property rather than a promise. Chained means each entry carries the digest of the one before, so a removed record fails verification rather than simply being absent.
Evidence
6 records| Seq | Record | Action | Decision | Policy | Previous | This record |
|---|---|---|---|---|---|---|
| 7735 | EV-7735 | close.ticket | Decision: Authorized | write.window v2 | 1c310ca2 | 0b46589958bd78db… |
| 7736 | EV-7736 | close.ticket | Decision: Authorized | write.window v2 | 0e2386e8 | 773e9be89f8f6e38… |
| 7737 | EV-7737 | update.record | Decision: Authorized | write.window v2 | 497d8d4d | 20db948841aad218… |
| 7738 | EV-7738 | adjust.stock | Decision: Authorized | write.window v2 | 9957a401 | b34da714573e047c… |
| 7739 | EV-7739 | read.contract | Decision: Authorized | pii.export_block v11 | e3b7375d | c3307766ab4bf592… |
| 7740 | EV-7740 | adjust.stock | Decision: Authorized | write.window v2 | f113ccde | 8aa688e428297eec… |
One record, in full
- Record
- EV-7512
- Action
- ACT-7512
- Decision
- authorized
- Rule version in force
- v7
- Version published
- 30 Jun 2026
- Previous digest
- d5bdfde76116b2a5cf853ad280bb989659848df5e0ab78af500ef78f968fb61d
- This digest
- d7ffd0b3bab589c9f5e902bddacf4f875e6f108129d6fb13e0e5eea8b1f6b278
And the rule’s own history
A version is published, never edited. Editing a rule destroys the record of what it used to say, which makes every past decision it produced unexplainable.
v7tightened after a review
30 Jun 2026 · in force
v6scope widened to a second system
11 May 2026
v1first published by Finance
24 Mar 2026
Which means the honest answer to “what was the limit in March” is a lookup rather than an archaeology exercise. Replay goes further and re-evaluates the action itself.
IllustrativeIllustrative records from the OpsAI sample estate, a fictional company.
Against the controls you already report on
9 control themes, and 4 of them carry a stated limit.
The mapping itself lives on the platform page. What matters here is the second number: a compliance matrix with no gaps is one you will have to defend, and this one names the places OpsAI evidences partially or not at all.
Control themes
9
across 5 instruments
Evidenced end to end
5
the record answers directly
With a stated limit
4
named, not omitted
Not evidenced at all
1
still in the table
- ISO/IEC 42001
- NIST AI RMF
- EU AI Act
- SOC 2
- DPDP Act
The limits, stated where a professional will look for them
- Inventory of AI systemspartial
- Completeness cannot be proven, only evidenced. Discovery reads the systems you already run, and it names what each source is blind to.
- Personal data handlingpartial
- Whether consent was validly obtained upstream is not something a control plane can observe. OpsAI evidences what was read and where it went, not the lawful basis for holding it.
- Model provenance and documentationpartial
- Training-data provenance and the accuracy of a provider’s model card are the provider’s to substantiate. OpsAI records what you were told and when, not whether it was true.
- Workforce competence and trainingnone
- This is an HR record. It appears here because a mapping that silently omitted it would be implying coverage OpsAI does not have.
One theme is not evidenced at all and stays in the table, because a matrix that quietly dropped its inconvenient rows would imply coverage that does not exist. The full mapping.
Where to start
Take the control your last assessment argued about, and ask what would have settled it.
Usually a timestamp, a named person and a decision — three things that either existed at the time or did not. Everything on this page is downstream of writing them down as they happen.