A file that records what happened to it, and who could reach it.

The platform runs no AI test on a sample with no mapped evidence: a test of that sample alone is refused, and a run over several samples leaves it out, so no result is recorded for it.

On a sample with mapped evidence, an attribute with none mapped to it is recorded inconclusive by the platform, without calling the model. The model is instructed to fail an item whose record does not meet the criterion, and to return inconclusive where the evidence does not reach the item.

Where a PASS or not-applicable answer quotes text the evidence it was shown does not contain, rests on a value only the population listing carries, or concerns a sample the evidence does not locate, the platform records the result as inconclusive, held by the platform, and the binder prints the model’s own answer beside it. No PASS the model proposes is final until the auditor confirms it.

An answer given with no rationale, or fewer than ten characters of one, is recorded as inconclusive whatever it was, a FAIL included. Where the quote, value, coverage or location checks question a FAIL, the FAIL is kept.

On the 9 attributes with a declared rule, the platform itself finds each fact the rule names as the whole value at the place in the evidence the model cites, all on one record of the sample’s own.

A rule names the date of the act and the date it is measured against, who performed it and its recorded status; or two people who must differ.

Where the platform computes that the dates are in the order the rule requires, or that the two names differ as written, and no other field of the record writes a value, a state, a flag, a count or a word the platform reads as denying the act, the answer is shown as “Checks matched — your decision”, with each fact cited, its file and its place.

It also shows what the platform did not check: whether the cited facts are the ones that answer the attribute, other rows, blocks or files about the item, what other fields of the record mean beyond a value or a word that denies, a state, a flag or a count, prose, and who the person named is or their authority.

Nor does it check whether the file is complete, accurate and authentic as the client produced it: the report or query that produced it is not checked here, and deliberately altered files, including text whose shown letters differ from its stored text, are not detected beyond the PDF checks under What the auditor reads.

The auditor confirms it with a note, which starts as the platform’s own sentence and may be kept or replaced, or decides otherwise. Every other PASS, and every not-applicable answer, is held the same way, the model’s proposal, its confidence and its quotation shown beside the result.

The coverage check reads a PASS against each interval of the period wherever the criterion or the attribute names a recurrence, whatever the rationale says, and also where the rationale says the control operated across the period on a control with a stated cadence or a periodic frequency; a sample that is itself one occurrence of a control sized on that recurrence is read only where its rationale says so.

It records that PASS as inconclusive in the same way where the records shown leave an interval without a record, where the intervals cannot be read, or where mapped evidence that could carry a dated record was not shown to the model. A criterion naming the year end is read as an annual recurrence over the period tested.

A PASS saying the control operated across the period, on a criterion that names no recurrence or cadence for a control not sized on a periodic frequency, is flagged for the reviewer as well.

No PASS and no NOT APPLICABLE the model proposes is final without the auditor: each is stored as INCONCLUSIVE, with the model’s proposal, its confidence and the reason the platform holds it kept beside the result.

The auditor decides each row on its own: confirms the model’s proposal with a note, decides differently, or accepts the result as inconclusive.

On an answer whose checks matched, the note starts as the platform’s own sentence, which the auditor may keep or replace with what they checked; on any other answer the auditor writes the note, of at least 10 characters.

The result records who decided and when, and a confirmation is written to the result and to its audit row in the same transaction.

The platform never upgrades a FAIL.

A FAIL the model gives is stored as FAIL, and is kept where the quote, value, coverage or location checks question it.

An answer cut short, or one the platform cannot read in the required format, is read as FAIL where the word fail stands in it in any form, even “did not fail”, or as a result other than PASS that it names, and is not asked again.

A FAIL given with no rationale, fewer than ten characters of one, or one longer than the record holds, is stored as INCONCLUSIVE, never as a PASS.

Only the auditor’s override, with a note, turns a FAIL into another result; the auditor may also record their own result for the attribute instead, and a re-run records the model’s answer on that run under the same rules.

The model is tested through the platform’s own Testing step, on demonstration engagements built from synthetic evidence, and each answer is compared with the expected result recorded for that check.

The latest run on the Testing step was held to four tests: no PASS or NOT APPLICABLE is stored as final; every check whose expected result is FAIL is recorded FAIL; no answer reaches “Checks matched” where PASS is not the expected result; and every quotation in an answer proposing PASS, and in a recorded FAIL, is checked against the bytes of the evidence, and every statement the check flags is read by hand.

The latest run on the Testing step is recorded with its date, the guard rules it ran under and its counts; each of its results the model answered records the prompt and the guard rules it was given under, and every answer stays in the platform’s results. Every change to the prompt or the rules is entered in the prompt change log with its reason.

The prompt and the guard rules in force are later than the latest run with the model; each change since is entered in the prompt change log with its reason and how it was checked, by the platform’s own tests and not by a run of the model.

This page describes the method, not a score: figures from a run hold only for the prompt and rules that run used.

A licensee’s quality or risk team can ask for the records of the latest run.

The rules in force are dated 17 October 2026.

The location check on a cited fact reads a table written with column separators, and it also reads the first column of a table a file’s text writes without column separators, as a PDF’s often does: a fact cited there is checked against the cell on its row.

The text the platform reads from a PDF keeps a wide space between two columns as a column separator, and the check reads a label and its value standing in two such columns as one field, and every column of a table whose header stands in three or more such columns, each cell under its header.

A PDF row in which a cell, a form label or a form value goes on in a line of its own, and a field whose value may go on in the next line of any file’s text, are not read as the whole value, except where the check does not see the value go on, each case named under what the auditor reads.

Each of these is held for the auditor with its reason, never shown as “Checks matched”: a JSON fact whose key path names no one value or whose record does not name the sample alone; a label and its value laid side by side in a PDF with no colon or equals sign between them, where the PDF’s text does not keep them as two columns; a record whose key stands under a column not named as the population’s key.

The first line of a value is read as the whole value where the platform does not see the value go on: a PDF cell going on at the next page’s top, lower than two font sizes and the row spacing, at the row’s left above its row or where rows sit as close as a cell’s lines, or in every cell; rotated text; a next line opening a field, list item, heading or table row of its own.

An identifier followed by a full stop and digits alone, as a version is written, is read as one name, not as another item’s identifier. Word documents are read as text: body, tables by row and column, headers, footers, notes and comments.

A PASS citing one is held for the auditor, each part named, where its body has deleted or hidden text, a drawing, a checkbox, a drop-down or computed field, a table row with a merged cell, image or nested table, where a note or comment is not read whole, or where the file was last read before tables were read this way.

A PDF’s text layer is read with each page’s drawing checked beside it.

A PASS citing one is held, each page named, where a page holds an image, a mark other than rules and boxes, an annotation or form field that shows something, no text while others have some, text hidden, covered, too faint or small, or, in a simple font, drawn as other letters than stored, or the file carries a hidden layer, an XFA form or scripts.

Where a file’s text writes a table without column separators, as a PDF’s often does, its first column is read by row, and a fact cited under another of its columns is held for the auditor with its reason.

Where a vulnerability’s remediation rests on a later authenticated full scan, the model may answer INCONCLUSIVE instead of applying the prompt’s rule for reading that scan; that answer is held for the auditor as the model gave it, and no check changes it.

The checks test the facts the model cites as written; whether those facts answer the attribute is the auditor’s judgment, so every row is read by the auditor and a right answer may be held for that reading.

Confidence is the model’s own figure, not a calibrated probability, and a re-run may answer differently: the default testing model takes no temperature setting.

Every change to the testing prompt is an entry in the prompt change log, under a version of its own.

The prompt version and the guard-rules version in force (prompt dated 12 October 2026, rules dated 17 October 2026), the fingerprint of the exact system message, the model requested and the model that answered are recorded on every result, and each model the platform can call has a model card.

Under the guard rules in force, each result of a model call on an item records in which mapped files the sampled item was located, and an answer proposing PASS or NOT APPLICABLE on an item no mapped file locates on its own record is held for the auditor.