A credential you can verify.
The whole platform rests on one thing: that a Kafa score is trustworthy. If companies do not trust the result, the platform is worthless. So trust is built into every layer.
Depth over speed.
We ask the follow-ups. We don't accept shallow answers, even when they're well-phrased.
Candour is a strength.
We reward candidates who name the edge of their competence, and mark down bluffing.
A sealed dossier.
Every evaluation is signed and verifiable — companies can confirm its integrity at any time.
A score is only worth what stands behind it. Four mechanisms shape what every Kafa evaluation is — and what it refuses to be.
Rubric-anchored
Every question carries its own rubric — what good looks like, what shallow looks like, what's outside scope. The score is the rubric applied, not a free-form impression.
Dimension coverage
Every session probes all four dimensions — technical depth, problem solving, communication, consistency. A candidate can't slip through by being strong in only one.
Per-session variation
A candidate retaking won't see the identical question set. Questions are drawn adaptively with a seeded shuffle — coverage preserved, sequence different.
Adaptive follow-ups
When an answer is thin, Kafa probes — like an examiner. A vague reply doesn't pass; depth is asked for.
Integrity is a confidence signal, never a verdict. A flagged session still passes if competency was demonstrated. The integrity panel exists so a human moderator can read the session in context — not so the platform can punish a candidate automatically.
Paste detection
Captured per answer. Large pasted blocks signal external assistance; the signal is recorded and rendered for moderator review.
Focus-loss tracking
How long the candidate spent on other tabs, and when. Context, not accusation.
Typing rhythm
The cadence of how an answer was written. Anomalous patterns are noted — long stretches with no rhythm, sudden bursts.
AI-text authorship read
An LLM reads each answer for the hallmarks of generated text. Reads as the candidate's own writing, AI-assisted in parts, or predominantly AI-generated — context for moderators.
The evaluation belongs to the candidate. Nothing public or shareable is enabled by default; everything that reaches a recruiter does so through the candidate's own explicit, scoped, revocable choice.
Discoverability is opt-in
A candidate appears in company search only if they explicitly enable it. Default is private.
Public sharing is opt-in
The public verify page is enabled by the candidate, instantly revocable. The credential exists either way; the public surface is the candidate's choice.
Scoped recruiter access
A candidate grants a specific company access to their interview at a chosen scope — read only, or read + recording. Each grant is revocable; revocation is immediate.
Private interview storage
Recordings live on a private disk. Never publicly exposed. Never on the public verify page or any PDF export. Accessible only through channels the candidate explicitly opened.
A credential is only as good as the trail behind it. Each evaluation carries its evidence; each AI decision carries its log; each session can be re-read by a human.
Token-keyed verification
Each passed evaluation issues a public token. The verify page resolves it — name, dimensions, narrative — and nothing else. The integrity panel and the interview layer are structurally absent from the public view.
Time-gated expiry
Credentials carry an issuance date and an expiration set by the track's validity window. Expired credentials read plainly as expired on the public page — transparency, not silent failure.
Every AI call is logged
Model, prompt, response, token counts, duration. Available to admin moderation. The platform's own evidence trail.
AI-assisted, not AI-automated
Admin can review, retry, or moderate any session. The dossier narrative will gain a candidate-side review checkpoint in a future release; today, admin moderation is the human check in the loop.