Neuropsych scores your results against published norms, works out what the pattern can and cannot support, and only then lets a language model write the prose. The model cannot change a number, raise a conclusion, or cite a source it was not given. This page shows each step and where the numbers come from.
Age-band table for ages 25 to 34 from the engine's symbol coding norms (Smith, 1982; Wechsler, 2008). Band width uses the task's assumed reliability of .80.
Eight steps. Six run as ordinary code and always give the same answer for the same inputs. Two use a language model, and validators check everything it writes.
Each raw score is compared with the published mean and standard deviation for your age band. That gives a z-score, a percentile and a standard score. Timed tests where lower is better are flipped, so a higher standard score always means a better result.
Nine tests do not have adult norms yet. Their results are reported with the accuracy you reached and the chance level, and they are never compared with other people.
Sessions that are empty, degenerate or physically impossible are excluded before anything is interpreted. A report needs one usable normed test in each of five core areas: processing speed, working memory, attention, executive function and reasoning. If a slot is empty, the engine stops and lists the tests still needed. The attention and mathematics analyses add their own requirements before they are evaluated.
Every normed score gets a descriptor on the seven-level AACN scale, a 90 percent confidence band based on a conservative reliability for that task, and a base rate: how often healthy adults obtain the same number of low scores in a battery this size. With four or more normed results, reliable strengths and weaknesses are identified against your own average. No composite scores are invented.
Your questionnaire answers and test results are mapped to the DSM-5-TR criteria for attention (ADHD) and for mathematics (dyscalculia). Each criterion is marked met, partially met, not met, unknown, or requires a clinician, with the evidence listed. The strongest level the engine can reach is "consistent with". The report never says you have a condition, and the section stays marked provisional until a licensed clinician reviews it.
Recommendations are selected from a reviewed library, each with a badge for the strength of its evidence. Safety items are pinned first. Every citation comes from a registry of 69 reviewed references: DSM-5-TR, the AACN consensus statements on labels and validity, the base-rate literature, NICE guidance, and meta-analyses for ADHD and dyscalculia.
The model receives computed facts only: your plain-language positions, the findings, the level for each concern and the plan. It does not see raw responses, confidence intervals or anything it could recompute. It writes named slots: a summary, one paragraph per domain, the pattern, an impression per concern, and a short note on why each "start here" item fits you.
Validators check every slot. Numbers must match the computed values. A condition may be named only where its level allows. Medication, doses and prescribers are off limits. A slot that fails gets one correction, and if it fails again, deterministic text written from the same findings takes its place.
A second, separate review pass reads the accepted slots against the facts. It flags results described as more extreme than their position, numbers attached to the wrong test, wording that goes beyond the level, and plan items presented as proven when their evidence is not strong. It can only replace a slot with its deterministic fallback. It cannot add anything.
A fixed template produces the HTML and the PDF: your results, the findings, the plan, then Appendix A with every score and its confidence band, Appendix B with the numbered references, and Appendix C on how to read the report. The report carries a participant ID rather than a name, and a clinician sign-off block is always printed.
Every comparison in a report uses an age-band table taken from a published study or manual. This is the current list, test by test, with the sample each table was built on.
| Test | Sources behind the table | Sample |
|---|---|---|
| Symbol coding | Wechsler, WAIS-IV (2008); Smith, SDMT (1982); Peña-Casanova et al., NEURONORMA (2009) | US adults, N = 2,200, ages 16 to 90; Spanish adults, N = 356 |
| Simple reaction time | Der & Deary (2006) | UK adults, N = 7,130, ages 18 to 80+ |
| Digit span | Wechsler, WAIS-IV (2008) and WAIS-5 (2024); Woods et al. (2011) | US adults, N = 2,200, ages 16 to 90 |
| Corsi blocks | Kessels et al. (2000); Monaco et al. (2013) | Meta-analysis, N = 739; Italian adults, N = 353, ages 20 to 89 |
| N-back | Jaeggi et al. (2010); Eisenberg et al., Self-Regulation Ontology (2019) | Healthy adults, N = 584, ages 18 to 50; 522 adults |
| Operation span | Unsworth et al. (2005) | US undergraduates, N = 256 |
| Continuous performance | Riccio et al. (2002) | Review of CPT research |
| Go/No-Go | Bezdjian et al. (2009); Eisenberg et al. (2019) | Review of published norms; 522 adults |
| Flanker | Zelazo et al., NIH Toolbox validation (2014); Eisenberg et al. (2019) | US adults, N = 268, ages 20 to 85; 522 adults |
| Stop signal | Verbruggen et al. (2019); Eisenberg et al. (2019); Lee & Hsieh (2017) | Consensus guide; 522 adults; 72 adults, ages 40 to 77 |
| Stroop | Strauss, Sherman & Spreen (2006); Scarpina & Tagini (2017); Eisenberg et al. (2019) | Compiled normative studies; 522 adults |
| Card sorting (WCST) | Heaton et al. (1993); Strauss, Sherman & Spreen (2006) | US adults, N = 899, ages 6.5 to 89 |
| Trail making | Tombaugh (2004); Strauss, Sherman & Spreen (2006); Giovagnoli et al. (1996) | Canadian adults, N = 911, ages 18 to 89; Italian adults, N = 287 |
| Tower of London | Culbertson & Zillmer, TOL-DX (2001); Shallice (1982) | US adults, N = 1,036, ages 7 to 80 |
| Matrix reasoning | Wechsler, WAIS-IV (2008); Raven (2000); Condon & Revelle, ICAR (2014) | US adults, N = 2,200; SAPA online sample, N = 96,958 |
| Similarities | Wechsler, WAIS-IV (2008) | US adults, N = 2,200, ages 16 to 90 |
| Verbal analogies | Sternberg (1977); Condon & Revelle, ICAR (2014) | SAPA online sample, N = 96,958, ages 14 to 90 |
| Word fluency | Tombaugh, Kozak & Rees (1999); Peña-Casanova et al. (2009); Casals-Coll et al. (2013) | US and Canadian adults, N = 1,300, ages 16 to 95; Spanish adults, N = 356 |
| Dot comparison | Guillaume & Van Rinsveld (2018); Halberda et al. (2012); Norris et al. (2015) | Meta-analysis of 47 samples; Panamath, N = 10,548, ages 11 to 85 |
| Number line | Williams et al. (2020) | College adults, N = 227 |
Magnitude comparison, arithmetic verification, dot estimation, fact retrieval, counting fluency, place value, number dictation, vocabulary and pattern completion. Published adult norms that fit these tasks are not available. The report shows what you got right and what chance would produce, and it never turns those results into a percentile, a descriptor or a finding.
Browser versions of these tasks are compared with published norms from similar, not identical, tests. Every report says so next to its score table. Where a computerized norm exists for the same paradigm, it is the one used.
Reports in Spanish go through the same eight steps, with Spanish prose slots, validators and references.
Animal fluency in Spanish is scored against NEURONORMA (Casals-Coll et al., 2013): scaled scores by age with an adjustment for years of education, built on Spanish adults. Word counts in one language do not transfer to another, so this is the one table that differs.
Every other test uses the same published tables as English. Most of them are nonverbal or timed, so the comparison holds. The three English verbal tests (vocabulary, similarities and verbal analogies) are not offered in Spanish.
These are properties of the code, not promises. Each one is enforced by a validator that rejects the text and falls back to deterministic wording.
Percentiles, standard scores, confidence bands, base rates and descriptors are computed before the model runs. Any number in its prose that does not match a computed value is rejected.
The level for each concern comes from the criteria mapping. Wording stronger than that level is rejected, and so is any form of "you have", "diagnosed with", "rule out" or "confirm".
Citations are limited to the reviewed reference registry, each with a stable ID. An unknown ID is removed from the text.
Medication, doses and prescribers never appear in model-written text. The one sentence on professional options is written by the engine, not the model.
Any sentence that reports what you said ("you mentioned poor sleep") has to match a recorded questionnaire answer.
A failed slot gets one correction round and then deterministic text. A provider outage produces the fully deterministic report, or a retry, never a broken one.
What the engine computes, on which assumptions, and which model writes the prose.
z = (raw − mean) / SD for your age band, percentile = Φ(z), standard score = 100 + 15z. Tests where a lower raw score is better are reversed first. NEURONORMA tables use their published scaled scores (M = 10, SD = 3) with the study's education adjustment, and the percentile is derived from the adjusted scaled score.
AACN uniform labels (Guilmette et al., 2020), seven levels: exceptionally low below the 2nd percentile, below average to the 8th, low average to the 24th, average to the 75th, high average to the 90th, above average to the 97th, exceptionally high from the 98th.
90 percent, from the standard error of measurement, 15 × √(1 − r), with a conservative reliability per task (default .70). The band is printed next to every normed score.
The expected number of low scores in a battery of this size follows Binder, Iverson and Brooks (2009), using a one-factor model with an assumed mean correlation of .30 between tasks. Relative strengths and weaknesses are ipsative: a result must differ from your own mean by more than its critical value and by at least one standard deviation.
Online testing has no stand-alone validity test, so the engine combines three deterministic signals: sessions excluded as unusable in step 02, engagement flags on tasks where engaged adults are near ceiling, and the conditions you reported for your sessions. A genuinely low but valid score is never lowered or removed.
A language model from Anthropic writes the prose slots, and a separate review pass on the same model checks them, in English and Spanish. Every model call is optional: the deterministic report exists before the first one is made.
Reports cite a registry of 69 reviewed references, rendered as numbered entries in Appendix B. Every citation in a report comes from that registry. Behind it sits an index of 1,300+ research papers, split into 55,530 searchable passages.
Results are saved to your account when a test finishes. Generating a report sends your computed scores and questionnaire answers to Anthropic's API, which does not use API data to train its models by default. Raw trial-by-trial responses are not part of that request. Details are in the privacy policy.
Every report prints a sign-off block with the reviewer's name, credentials, date and notes. Until a clinician signs, the diagnostic considerations carry a provisional banner.
Limitation. This is a screening tool, not a diagnostic instrument. Normative comparisons give context but cannot account for medication, fatigue, the testing environment or motivation. Some age ranges are better represented in the published tables than others, and the report notes this where it applies. Talk to a professional about your results.