01 · The instrument
The job writes the test.
Our assessments start from the role, not a fixed item bank. Responsibilities, decision conditions, risk exposure and competency priorities are read from the job description, and new situational judgment scenarios are generated for that role: never an off-the-shelf bank.
Theta AI™ turns the role into five-option situations for large cohorts. Titan AI™ interviews a smaller group in free text and follows up on the reasoning behind each answer. Two engines, one starting point: the job. The situation on the right is the one a candidate answers next.
02 · The answer · Theta AI™
Two picks. Two separate signals.
In every Theta AI™ situation the candidate makes two distinct choices: the action they judge most effective (Best) and the one they judge most harmful (Worst), from five options on the same scale.
Choosing effective actions and spotting harmful ones are different capabilities, so the report keeps them separate: effective-action evidence and harmful-action evidence are scored and shown on their own, and competency panels stay distinct instead of being silently blended into one opaque composite. Those two picks are what the measurement in 04 scores.
03 · Anchored evidence · Titan AI™
Behavior-anchored scores, not impressions.
Titan AI™ scores free-text reasoning on four independent axes: Evidence, Communication, Risk awareness and Execution. Each axis is anchored on a 1 to 5 scale by observable behavior, so a score points to what the person actually wrote, not to polish or length.
Every scoring event is persisted, or skipped with a recorded reason. Nothing is dropped silently, and the four axes are never averaged into a single hidden number.
04 · The measurement · Theta AI™
A score that knows its own uncertainty.
Theta AI™ scores with item response theory: a graded response model, run in R with the mirt package, on frozen item calibrations. Each competency is reported as an EAP estimate together with its standard error, so a Theta score arrives with its own confidence attached. As informative responses accumulate, uncertainty can narrow. The animation shows one illustrative sequence.
A live run never re-learns the model from the people it is scoring. It requires a frozen calibration asset whose provenance fingerprint matches the registry; without that match the run is refused, not re-fit on the same live cohort.
Titan AI™ presents free-text reasoning differently on purpose: behavior-anchored 1 to 5 scores per axis, with the quoted evidence beside them. The report does not print a per-axis standard error. The anchor and the evidence are the explanation. The estimate lands in the report with its full trail. That is 05.
05 · The evidence trail
Source-grounded, figure by figure.
The Best pick from 02 becomes a figure with a trail. Open a figure in a report and you can see what produced it: the candidate’s actual response, the scenario it answered, the engine version that scored it, and the fingerprint of the exact data snapshot behind it. If anything were swapped after the fact, the fingerprint would not match.
This proves provenance, not truth. A fingerprint shows a figure was not altered later. Whether the figure deserves trust comes from the method in sections 01 to 04, and from the limits in sections 06 and 07.
06 · Governance
Limits are printed, not buried.
Ties stay ties
Equal results are reported jointly. Nothing is arbitrarily ranked to look decisive.
Missing denominators stay visible
If a comparison lacks its base group, the report says so instead of implying one.
Provisional means provisional
Synthetic or provisional calibration is labeled on the report. It is never presented as validated norms.
Small cohorts carry a notice
Cohort context is tiered by size. A comparison group below the standard size is flagged, not smoothed over.
Promotion is manual
Calibration reaches production only through human quality gates, never by automatic accrual.
07 · Validation
The gate that says “insufficient data”.
Our validation tooling runs criterion-correlation and adverse-impact analyses, and both refuse to produce a number the data cannot support. Below the minimum group sizes the result returned is “insufficient data”, not a flattering estimate.
Calibration today runs in provisional and synthetic stages, and we do not claim criterion-validity results we do not have. Where evidence is provisional, the reports say so. We would rather publish the gate than fake the number.
08 · A fair comparison
Judge us against anyone.
SHL, Korn Ferry and Hogan have published assessments, technical manuals and norms for decades, and they document real research programs. We respect that record. Our answer is different in kind: we put the evidence trail inside the product, one report at a time.
Wherever you evaluate assessments, including here, five questions cut through:
- 01What writes the test: a fixed item bank, or the actual job?
- 02Does each figure name what produced it: a response, a calibration, or a cohort group?
- 03Is uncertainty printed next to the score, or hidden in a manual?
- 04What does the system say when there is not enough data?
- 05Who makes the hiring decision, and what do they actually see?
The point
One assessment.The decision stays human.
Theta AI™ and Titan AI™ structure the evidence. People weigh it. No score on this page, or in any report, is an automatic hiring decision.