An assessment platform in 14 days

Fourteen days, and most of it is not the questionnaire

What sixty-one tables buys, beyond a form: multi-organisation tenancy where a cross-tenant read returns nothing rather than depending on a filter someone remembered to write; a prepaid wallet with a ledger; assignment to individuals and to groups; a runner that autosaves and resumes; a scoring engine with no database access at all, so it can be re-run on stored responses years later; norm handling that records which reference population produced each standard score; and a report rendered server-side as a PDF with the fonts embedded, because Persian needs contextual letter shaping and a renderer that does not do it produces disconnected, reversed letters that are legible to nobody.

The rule the whole system is built around

Every number in a report must be explainable and reproducible years later.

That sentence is written once in the architecture decision record and it forces most of the rest.

It is why the scoring engine is pure — it takes a test version and a set of responses and returns raw scores, touching nothing else. It is why a result is immutable and bound to the version it was scored under, so a stored report still renders under its own version’s bands after the instrument has been revised. It is why every standard score records where it came from: an exact table row, an interpolated one, or a mean and standard deviation. “The norm table probably had that row in 2026” is not an answer to a candidate’s lawyer.

"Someone is going to challenge this result in three years. Can you show me exactly how the number was produced?"

The audit that found the key was wrong

The strongest piece of work in this project is also the least flattering, which is usually how it goes.

The scoring key — which item counts toward which facet, and in which direction — was audited by reading what every one of the 240 Persian items actually says and comparing that to the direction it was keyed. Thirty-six items were keyed against the direction of their own text.

The worked example from the audit says it better than a summary can: a respondent answering strongly agree to “I am a very disciplined person” was scoring zero on that item for self-discipline instead of the maximum.

Two further items turned out to be translation defects that negate the item they render. They are keyed to the original meaning, and the cost of that — a respondent answering the Persian sentence is scored against its opposite on those two items — is written down rather than smoothed over, because fixing the wording changes a published instrument and detaches those items from the reference data collected on them.

The regression test that now guards this asserts the reverse-keyed items by number, not by any tidy rule. That detail is the whole lesson: the two earlier keys both passed the checks they were given, and were wrong anyway, because the checks were satisfiable by a pattern that looked orderly.

Where the model is allowed to go

The boundary is stated to users on the product’s own landing page rather than buried, on the reasoning that a buyer who discovers it later assumes it was hidden.

1. The engine scores. It applies the version's recorded key, handles reverse-keyed items, produces the domain and facet scores, and determines each band.
2. The model explains. It receives computed scores — never answers, never a name, never anything about the employer — and writes a plain-language reading of them.
3. The model changes nothing. It cannot alter a score, invent a trait, give a diagnosis, or make a hiring decision.
4. The system checks the output. Chart references the version does not have are dropped, importance values are clamped, malformed output degrades to prose rather than to an error, and clinically loaded terms the model used get their definitions appended server-side.

That last step exists because of a real incident. The prompt asked the model not to write its own definitions of clinically loaded facet names, and a live run produced exactly those words with no definition anywhere and no statement that neither is a diagnosis. A prompt is a request. So the definitions were moved into the payload as data, and a post-hoc scanner appends the glossary whether the model cooperated or not.

The gap we will not paper over

The system has no response-quality screening. The team’s own planning document states it in one sentence: a test answered “agree” 240 times currently produces a confident, wrong profile.

The indicators that would catch it — longest straight run, response variability, distinct options used, and a validity flag shown above everything else on the result page — are designed in detail and are not in the database. The only control actually running is a minimum-answered ratio.

This is the most important thing to know about the platform in its current state, so it is on this page twice: here, and in the limitations at the top.