Product · Assurance
The record your customer reads directly
Everything in Verification, plus the part your customer's risk team actually asked for: an evidence pack mapped to the clauses they cite, and a read-only portal they hold a link to for the contract term. Bought because your customer asked.
The portal
What their risk analyst opens
Not a dashboard you present. A record they hold a link to for the contract term, where the cadence is stated, the runs are counted, and a gap is a finding rather than a blank. The second panel is how the same figures read when runs stop - shown because a report that can only render green is one nobody has a reason to believe.
Production Q&A, as deployed for Insurer A
Acme Pty Ltd · Claims platform · period 1 January to 31 March
deployment dep_8f21c4 · chain entries 1-47
chain 88f1…c2 · RFC 3161
timestamped 31 Mar 23:59 UTC
Reporting cadence and gaps
- Contracted
- Weekly+ on every change
- Runs in period
- 1513 scheduled · 2 on change
- Last run
- 3 days ago
- Gaps
- None
The same panel, when runs stop
- Contracted
- Weekly+ on every change
- Runs in period
- 40 on change
- Last run
- 92 days ago
- Gaps
- 9 missed runs
Scope and limitations
Measures accuracy against 203 customer-confirmed test cases, and assertions unsupported by Acme's own knowledge corpus - a proxy for hallucination, not a claim of factual truth. Does not test transactions or actions. Judge model: a different vendor from the system under test. Presoja is paid by Acme; see the basis of engagement. Measured judge-human agreement κ 0.83.
Illustration with sample data, not a screenshot.
The terms
What Assurance adds, written as obligations
Everything in Verification, plus the part that makes the record admissible to somebody else's risk committee - and each addition binds us, not you.
- The evidence pack
- Mapped to the clauses your customer's risk team cites, so their reviewer files it rather than translating it.
- Standing disclosure
- Consent to show your customer the record is given once, at contract time. Scheduled reports publish automatically and cannot be withheld individually - a report you approve before your customer sees it is not evidence, it is marketing.
- A contracted cadence
- The schedule is in the engagement, so a missed run is a breach rather than a quiet quarter. If runs stop, the portal says so.
- Candidates scored before you switch
- The models you are considering, run against your confirmed cases, so the migration decision is made on your two hundred cases rather than a leaderboard.
The limits
The parts a validator would find anyway
Everything we sell is worth less if its limits are discovered rather than disclosed. The product states its own before its results, so the site does too.
- We are not accredited, and we do not sell a certificate
- There is no badge at the end of this. An accredited ISO 42001 certificate is now a thing you can buy - and what it attests to is your management system, never whether your answers survived the last model change. Supervisors are explicit that a third-party certificate is not sufficient on its own. What we produce is a record with its method attached, which a validator reads rather than accepts.
- You pay us, and that is a real problem we have not solved
- It is the issuer-pays conflict, unchanged since the credit ratings agencies. If we find a regression, you can decline to renew. The literature is specific about what would actually fix it - disclosure of financial relationships, engagement terms that stop auditor-shopping, cooling-off periods, protection from retaliation - and we have none of that machinery. Any assurance business claiming this is solved is asking you not to think about it.
- So independence is a property of the record, not a claim about us
- The set is in our custody and every change to it is logged. Thresholds are fixed before a run, never after. Every run is recorded, pass or fail, with no quiet re-runs. And consent to show your customer the report is given once, at contract time, rather than per report - because a report you approve before your customer sees it is not evidence, it is marketing.
- A black box gets half the product, and we say which half
- If your system depends on a third-party AI you cannot version-pin or override, we can still detect that its behaviour changed and show you verbatim what broke. We cannot test a candidate before you switch, because there is nothing to switch. That is priced and stated at contract time, and it appears on your exposure view as a row reading no version visibility rather than as a gap nobody mentions.
- A fixed set is not a claim about everything your system does
- We measure the cases in the set, drawn from your own material and confirmed by your own expert. A report says what was tested, how many, and with what confidence interval - and never that the system is correct in general, which is not a claim any finite test set can support.
The next step
The numbers are on the page
Two plans, nothing metered, and one question back: whether the price is roughly right.