Evidence posture
Every product here that makes a scientific claim rests, to some degree, on provisional science. That is ordinary for work at this stage, and it is written in each product's own documents. This page is those documents, in one table, with the threshold that moves each line stated before the data arrive rather than after.
30 claims across 18 products. A foundation that would publish its nulls should publish its not yet.
Rests on published work, cited, with the product implementing it as published. The strongest rung a product can occupy before it has its own data.
A defensible construction — computed from published parameters, or a reasonable starting value — that this Foundation has not itself validated. It is used, it is labelled, and it is not a finding.
A study is designed or running that will move this claim. The pre-registered threshold is stated before the data arrive, not after.
No study exists yet. The claim is held open and the product does not rest on it.
| Claim | Status | What it rests on | What moves it |
|---|---|---|---|
| The OMOP-shaped longitudinal record | Literature-anchored | The OMOP common data model, implemented as specified. This is plumbing, not a finding, and is listed so the ladder is complete rather than selective. | — |
| Claim | Status | What it rests on | What moves it |
|---|---|---|---|
| None. Foundation Operations is an internal console. | Literature-anchored | Listed for completeness. | — |
| Claim | Status | What it rests on | What moves it |
|---|---|---|---|
| None. Hearth coordinates a family; it makes no scientific claim. | Literature-anchored | Listed so the ladder covers every product rather than only the ones with something to admit. | — |
| Claim | Status | What it rests on | What moves it |
|---|---|---|---|
| CPIC pharmacogenomic dosing cards | Literature-anchored | CPIC guideline tables, implemented deterministically. The engine organises and flags; it never auto-doses. | — |
| Triage flags on between-visit check-ins | Literature-anchored | Published thresholds, cited in the product's own literature/ directory. | — |
| Claim | Status | What it rests on | What moves it |
|---|---|---|---|
| None. Repose is tooling. | Literature-anchored | Listed for completeness. | — |
| Claim | Status | What it rests on | What moves it |
|---|---|---|---|
| The actigraphy-derived biomarker | Provisional | 2026-09-23 source pass, then re-weighted on clinical sign-off (pipeline v1.1.0). The ADI’s weights had cited six Cohen’s d values from "Cortese 2006, JAACAP 45(7):787-800" — a citation with the wrong year, volume, issue and pages for a paper that reports z statistics, not d. Against the real meta-analyses only sleep onset latency (actigraphy, z=3.44) and true sleep time (z=2.85) are well supported; WASO and REM latency, which had been carrying 0.25 of the vector between them, are NOT significant in Cortese 2009, Cortese 2006 or Sadeh 2006. The vector now leads on SOL and TST (0.25 each), keeps a 0.05 trace on WASO for Gruber 2000 instability, and zero-weights REM latency — which this pipeline does not measure anyway, since it does no sleep staging. HRV is unchanged at 0.20 and is labelled as our own autonomic hypothesis rather than given a citation it never had. Shipped as a version bump, not an edit: adi_scores and baselines carry pipeline_version, the SDTM export publishes it as FAMETHOD, and the ICC endpoint compares versions — so the replicability measure will quantify what the change did. Both tables were empty in production, so no score moved. | The dual-wear study, plus the question no weight vector answers: under stimulant treatment TST falls and SOL rises while WASO and movement also fall (Boonstra 2007 PMID 17520787; Surman & Walsh 2022 PMID 35246824). v1.1.0 weights the rising pair 0.50 and the falling pair 0.20, so that cancellation is reduced but not removed. The ADI measures deviation from a personal baseline and still says nothing about whether the deviation is an improvement. |
| Claim | Status | What it rests on | What moves it |
|---|---|---|---|
| Subtype classification weights | Provisional | Trained on a synthetic cohort whose own manifest forbids deployment. The model refuses to serve when manifest.dataSource is synthetic. | A validation study producing a v2 trained on real data, gated on the manifest. |
| Claim | Status | What it rests on | What moves it |
|---|---|---|---|
| The Daily Report Card as an intervention | Literature-anchored | Two meta-analyses. Pyle & Fabiano (2017, Exceptional Children 83(4)): 14 papers, 40 single-subject cases, average effect sizes 0.59–0.94 on the PND/PAND/PEM/IRD/Tau-U estimators, fail-safe N of 68 studies against the smallest. Iznardo et al. (2020, J Atten Disord 24(12):1623–1636): 7 group designs, 272 participants, mean age 7.9. See relay-p1/literature/daily-report-card.md. | — |
| The specific behaviours in our goal library | Provisional | A starter set in the Fabiano/Pelham tradition, not a verbatim reproduction of a published manual. The intervention being evidence-based does not make an arbitrary list of target behaviours evidence-based — the literature is explicit that goals are chosen collaboratively and operationalized per child. | A clinician review against the source manual, treating the library as a starting point a team edits rather than a menu shipped as-is. |
| PK reference curves for the medication-coverage map | Provisional | Every parameter has now been checked against the FDA-approved prescribing information and corrected where it disagreed: Jornay PM's half-life (5.0→5.9 h) and lag (8→10 h), the amphetamine half-life (11.5→9.5 h — the adult figure was in use in a pediatric product), MPH IR (2.5→2.8 h) and guanfacine (17→18 h). See relay-p1/literature/pk-parameters.md. | A pharmacist review of the curves as clinical content, and a source for the coverage threshold — the '50% of peak' figure the rebound rule uses has no citation and is the most important unsourced number in the product. |
| The OROS (Concerta) curve's shape | Provisional | The label describes a BIPHASIC profile with an initial maximum at about 1 hour, then a gradual rise over 5–9 hours. The model has one peak and a 2.5-hour absorption lag, so it shows no drug before 2.5 h and understates morning coverage — the window a teacher's first rating falls in. | A two-component curve (immediate-release overcoat plus ascending-release core) built on published release fractions. Until then the curve is labelled as morning-underestimating rather than tuned until it looks right. |
| HEDIS ADD episode timing | Literature-anchored | Checked against NCQA's published specification. It found a real off-by-thirty-days: the continuation window runs 270 days from the END of the 30-day initiation phase, not from the index prescription, so the clock closed on day 270 instead of day 300 and reported compliant children as gaps. Corrected. See relay-p1/literature/hedis-add-measure.md. | — |
| That our HEDIS clock produces a HEDIS RATE | Provisional | It does not. The measure's denominator requires ages 6–12 at the index prescription and at least 210 days on medication, and the initiation visit must be with a practitioner holding prescribing authority. None of the three is modelled. A per-child follow-up clock is a good clinical reminder and is not a rate. | Applying all three denominator conditions in whatever export claims to be a HEDIS rate, and a billing-team review of the two-checkpoint model against 'any two visits in the window'. |
| Deterministic CDS insights (rebound window, amphetamine-class irritability) | Provisional | Two rules built strictly to the design's guardrails — no dosage numbers, non-exclusive option lists — over the PK curves above. | The coverage-threshold citation above, and outcome data from a pilot. |
| Claim | Status | What it rests on | What moves it |
|---|---|---|---|
| The Collaborative Life Skills programme it delivers | Provisional | 2026-09-23 source pass: the primary trial is real and strong — Pfiffner et al. 2016, JAACAP 55(9):762-770, PMID 27566117, NCT01686724, cluster RCT across 23 schools, significant post-treatment gains on both parent AND teacher ratings. But the follow-up into the next school year (PMID 29588050) found sustained advantage on parent, NOT teacher, ratings, and the sibling CLAS trial (PMID 24865871, N=199) found the same. Trellis is a teacher-facing instrument whose artefacts live with one teacher for one year, which is exactly where the effect evaporates. The "ES -1.09" figure is in no abstract, and CLS-S and CLS-FUERTE are pilots, not replications. | A pilot powered on SUSTAINED teacher-rated outcomes across the cross-year teacher transition — the gap the 2018 authors name ("booster treatments for improving outcomes with new teachers across school years") and the one thing a platform can do that a 12-week manual cannot. |
| Claim | Status | What it rests on | What moves it |
|---|---|---|---|
| The instruments it administers (PSQI, ASRS and others) | Provisional | 2026-09-23 source pass: the instruments are real and appropriate, but the live constants were not all sourced. config.symptomScales.adhd_rs5 = { max: 54, clinical_threshold_adult: 28, mcid: 7 } is applied by assessmentService.js to real submitted scores, and cited PMID 29570365 — "Nitrate Transport, Signaling, and Use Efficiency", Annual Review of Plant Biology. The real normative paper (DuPaul et al. 2016, PMID 26011476) covers children aged 5-17 and contains no adult norms and no MCID. Only max: 54 is sourced, and it is structural. The adherence benchmark cited a Chest 2016 paper that could not be verified to exist. | An adult threshold and MCID taken from the ADHD-RS-5 for Adults normative sample, or measured in Vesper’s own data. |
| Claim | Status | What it rests on | What moves it |
|---|---|---|---|
| NLP extraction of eligibility criteria | Under study | Built and unvalidated. 2026-09-23 source pass: the per-concept-type F1 table behind config.minConceptF1 was attributed to Gehrmann et al. 2018 (PMID 29447188), which is a document-level phenotyping benchmark on MIMIC-III ICU discharge summaries and contains no such table; the "psychiatric domain penalty" modifying it has no source either. The Perlis citation was a wrong DOI and an assembled title, and the real paper (PMID 21682950) is a better anchor: NLP on 127,504 patients’ psychiatric notes, AUC 0.85-0.88 versus 0.54-0.55 for billing data alone. The placebo-response section’s supplementary citation could not be verified to exist. | An NLP validation against human-extracted criteria. |
| Claim | Status | What it rests on | What moves it |
|---|---|---|---|
| The chronotherapy protocol | Provisional | 2026-09-23 source pass: the protocol is grounded in a real phase-response curve — Burgess, Revell, Molina & Eastman (2010), JCEM 95(7):3325–3331 — but the product was scheduling 0.5 mg melatonin at DLMO minus 6 h, while that paper states: "Maximum advances occurred when 0.5 mg melatonin was taken in the afternoon, 2-4 h before the DLMO", and "the optimal administration time for advances and delays is later for the lower dose". Corrected to 2–4 h. The literature note had been written without reading the paper; the code and its tests implemented the wrong figure faithfully. Separately, the DLMO-from-sleep-onset offset of 120 minutes, the 23:17 mean DLMO and the MEQ/MSF correlations attributed to Van Veen 2010 (PMID 20163790) are not in that paper. | The offset and chronotype corrections are verified against a source, or measured in Aubade’s own cohort. |
| That the protocol improves insomnia severity in this population | Under study | No validation cohort yet. | A 100-200 user open-label cohort reporting week-4 ISI change against a pre-registered threshold. |
| Claim | Status | What it rests on | What moves it |
|---|---|---|---|
| The core behavioural signal the nudges rest on | Under study | Un-de-risked. | A signal study. Kill criterion: a false-positive rate above the pre-registered bound. |
| Claim | Status | What it rests on | What moves it |
|---|---|---|---|
| Feasibility of the research instrument in the field | Under study | Built and unexercised. | A pre-registered feasibility study with one partner lab. |
| Claim | Status | What it rests on | What moves it |
|---|---|---|---|
| Speech-reception threshold and scale scores | Not started | Placeholder clinical content: ECLiPS items under a licence notice, synthesised stimuli, unapproved flat headphone curves and unvalidated thresholds. No real SRT or scale score can be produced today. 2026-09-23 source pass: the venture’s anchor trial is real, prospectively registered and accurately described: Shiels, Tomlin, Payne & Rance (2026), npj Science of Learning, PMID 42069773, ANZCTR ACTRN12623000512628. Two qualifications: the "~2.6x controls" reading-fluency figure is not in the abstract and is unverified, and the 30-week extension found NO significant differences in auditory processing or cognitive skills. The same group published a null result the year before on the same outcome (soundfield amplification, n=84, PMID 40841369) — which supports the personal-remote-microphone distinction Attune is built on, but only if cited rather than omitted. None of it validates Attune’s own screening instrument, which currently ships unlicensed placeholder items. | Licensed items, approved headphone curves and validated thresholds — each of which is a partnership, not a sprint. |
| Claim | Status | What it rests on | What moves it |
|---|---|---|---|
| That a browser can measure reaction time to a correctable <10 ms residual | Under study | Two published studies measured this with external hardware. Bridges et al. (2020, PeerJ 8:e9414, Black Box Toolkit): response-time LAG 8.22–67.01 ms, which is constant error and correctable; response-time PRECISION 0.39–8.44 ms depending on platform, which is not. Anwyl-Irvine et al. (2021, Behav Res Methods 53:1407–1425, robot actuator, realistic set-ups plus 200,000+ participants' device data): 'modern web platforms provide reasonable accuracy and precision for… manual response time', and no platform is best everywhere. See focal-p3/literature/browser-timing-2020.md. | The POC narrows to what the literature does not cover and the derisking plan already named as the kill condition: the jitter distribution on the devices our families actually own, in their homes, including the long tail. Bridges et al. warn their own figures are best-case and real-world timing is 'considerably worse'. Device matrix drawn from Anwyl-Irvine's published participant data rather than invented. |
| The in-browser calibration we ship | Provisional | It measures how accurately the device schedules a timer — setTimeout error against performance.now(). Neither the display presentation path nor the input-event delivery path is observable from inside a browser, which is why both published studies needed external hardware. A passing number is a necessary condition, not a demonstration of end-to-end accuracy. The screen now says so. | An end-to-end measurement against a hardware timing reference, per the derisking plan's Workstream A. |
| Normative tables for the triage signal | Provisional | Hardcoded starting values. The comparator is read and documented — Conners CPT-II (n=1,920, ages 6–55, ten age bands) in focal-p3/literature/conners-cpt2-normative.md, which also records that our age bands do not match its and that the mismatch has to be justified in any SaMD submission. What is missing is a cohort, not a reading. | A pre-registered cohort pipeline replacing them. Until then the triage signal is not a clinical finding. |
| Pupillometry and eye-tracking features | Not started | 2026-09-23 source pass: RETRACTED. The pivot path’s source — "Bijlenga et al. (2009), Neuropsychologia 47(13), 2940-2945" — does not exist: the DOI resolves to an unrelated object-priming study, no paper occupies that journal/year/page, and no paper of that title exists. Its invented numbers became config.pupillometry’s clinical cut-off, a service, and passing unit tests. Nothing was ever wired into the assessment flow, so no participant was classified by them. The real literature does not support the premise either: Bast 2020 (PMID 31853987) finds ADHD’s signature is a cue-evoked LATENCY effect with the amplitude effect belonging to ASD, and Burley 2021 (PMID 32102703) finds dilation INCREASED, not reduced. | A threshold derived from a real study — most likely Focal’s own, since no published cut-off exists for a 20 Hz consumer webcam under uncontrolled ambient light, which is a harder measurement problem than the millisecond reaction-time gate this path was meant to sidestep. |
| Claim | Status | What it rests on | What moves it |
|---|---|---|---|
| Outcome prediction from the assembled feature set | Not started | 2026-09-23 source pass: both efficacy citations behind matchEngine’s weights were wrong. "Bousman 2021" does not exist (its PMID is a fluid-dynamics paper); the real work is Bousman 2019, PMID 30520364, OR 1.71 (95% CI 1.17-2.48) — and it is superseded by Bousman 2022, PMID 36111494, 13 trials and 4,767 patients, RR 1.41. Both are depression and antidepressants, not ADHD. Newcorn 2008 is real (PMID 18281409, not ...410) and supports keeping stimulant and non-stimulant as separate ranked paths, but the 2x2 responder matrix built on it is impossible for a trial where only the methylphenidate arm crossed over. | The planned retrospective back-test (N>=500) — and a re-weighting decision that starts from RR 1.41 with a depression-to-ADHD caveat, which argues for less than the current 0.15, not more. |
| Claim | Status | What it rests on | What moves it |
|---|---|---|---|
| Phone-camera screening for developmental coordination disorder | Under study | No scoring model exists yet. 2026-09-23 source pass: DESIGN.md quoted the DCDQ at "pooled sensitivity ~0.82" with no source; the meta-analysis (Park & Kim 2024, PMID 38744072, 27 studies) gives sensitivity 0.70, specificity 0.77, AUC 0.80. The questionnaire’s own pooled AUC of 0.80 is ABOVE this venture’s Phase 0 kill threshold of 0.72, so a model that clears the gate can still add nothing to a free paper form. | Phase 0, reported as combined-versus-DCDQ-alone AUC in the same sample rather than against a fixed 0.72. |
| DCDQ items | Not started | Licence placeholders. | A licence. |
Generated from shared/contracts/seeds/evidence-ladder.json, version
1.2.1, last updated 2026-09-23. A product cannot ship
without an entry: the contract tests fail if the registry holds a product this
page does not, and fail again if any claim that is not literature-anchored
omits the threshold that would move it.