RISCONAPOSTOLUS

RISCON Labs · draft 2026-11-14

Twelve health studies, read one question at a time

An independent RISCON Labs programme that studies twelve health and medical questions with public data and controlled tests. Targets, synthetic checks, public-data results and auditor sign-off are kept apart, so one auditor can follow each result from its question to its replay. A completed website is not a completed research programme.

33 of 33 screens built10 evaluation lanes blocked4 split rules locked ahead of data access
1,874published cases evaluated19-1 · 204 papers · 172 in the top five
30evaluation lanes16 manifest-locked, 10 blocked
8analyses executed6 harness audits passed · 0 external-cohort evaluations
34items needing reviewBlocked lanes, held verdicts and outside inputs

28 public sources registered · 15 manifests locked by SHA-256. Synthetic checks test code behaviour and decision rules; they are not research outcomes.

19-1 published-case evaluation →Target and result comparison →Synthetic case library →Evaluation registry →Replay commands →
Plate I. An evidence map: twelve project records, each tied by a fine line to its neighbours. Select a number to open its record. The lines group fields of study; they do not assert clinical relationships.

Start here

Short entry points for readers outside the field. Each says what it covers and keeps the source numbers and limits. This site records what studies and tests found. It is not medical advice; decisions about a person’s care belong with their treating clinician.

ProgrammeIntroduction

Twelve research projects

Each project states its question first, then its internal target, its data and the limits of what the evidence shows.

Read
ResultIntroduction

19-1 on published cases

172 of 1,874 published cases had the reported gene in the top five. The result is below the target and is kept as recorded.

Read
ComparisonIntroduction

Target and measured value

Every place where a recorded target meets a measured value, with the scope of the measurement beside it.

Read
RegistryIntroduction

Evaluation data and splits

The public datasets, reference labels and locked splits fixed for each project before any model is compared.

Read
MethodIntroduction

How one auditor checks the work

Four gates from data integrity to isolated replay, and the roles that prepare the evidence the auditor decides on.

Read
RecordIntroduction

Every published file

Specifications, replay records, audit documents and run outputs, each with its size and SHA-256.

Read

Newest documents

RecordProjectsDate
Source register and comparison with the callAll2026-11-14
Review of the code attached to 5-1 and 18-25-1 · 18-22026-11-14
Evaluation data, reference labels and splitsAll2026-11-14
Evaluation plan for seven projects5-1 · 8-1 · 18-2 · 16-2 · 16-3 · 18-1 · 18-32026-11-14
Evaluation plan for the prevention projects12-1 · 12-2 · 12-3 · 12-42026-11-14
19-1 published-case evaluation record19-12026-11-14

All research briefs · Document library · Model reviews

Target ledger

Question → target → result comparison

Each project records its internal target before evaluation. A comparison is recorded only when a measured value exists, and the scope of that measurement is written beside it. A value inside a synthetic scope is not a research outcome.

12 targets recorded9 comparisons recorded
19-1Below the targetPublished cases

Rare Disease Diagnostic Research

At least 90% of causal genes in the top five

Measured · 172 of 1,874 in the top five (9.18%)

Open the comparison and its scope →
8-1Below the targetSynthetic inputs only

Early Boost for Core Health R&D Technology

At least 95% of downstream predictive performance retained

Measured · Lowest of 4 endpoints 91.4%; 3 of 4 at or above 95% (synthetic, baseline generator; release blocked)

Open the comparison and its scope →
18-2Within the criterion in this scopeSynthetic inputs only

Medication Administration and Dispensing Safety

Zero missed contraindications or look-alike medication errors

Measured · 0 of 4,500 injected errors passed (simulated vision outputs; 62.0% of correct fills passed automatically)

Open the comparison and its scope →

The four programmes

Active3 projects

Discovery and data

Computational discovery and data infrastructure: structure-based small-molecule design, privacy-preserving synthetic multi-omics data and rare-disease diagnosis from variants and phenotypes.

Active2 projects

Cognitive and neurological health

Cognitive and neurological health: difficulty control for adaptive cognitive training and dementia risk stratification from brain MRI and speech.

Active3 projects

Patient safety

Patient safety in care settings: early warning of falls, cross-checks on dispensing and prescription rules, and counting of surgical instruments.

Active4 projects

Crisis prevention

Crisis prevention: early detection of acute mental-health crises, safety-plan support after emergency care, consent-based detection of transitions into isolation and immediate counsellor handoff for students. Specifications and evidence use abstract codes and placeholders only.

Read the research goals and validation criteria →

What is being asked now

One question from each programme. Each record carries the evidence that would change its reading.

5-1 · 2026-11-14

Structure-Based AI Small-Molecule Discovery

Compare structural evidence, experimentally measured binding values and predicted properties for therapeutic targets and previously reported compounds.

5-1 AI drug discovery
16-2 · 2026-11-14

Non-Pharmacological Dementia Research

Simulate response delay and fatigue during memory, attention and executive-function tasks and inspect responses to difficulty transitions and stop criteria.

16-2 Adaptive cognitive training
18-2 · 2026-11-14

Medication Administration and Dispensing Safety

Define reference cases for pill identification and prescription review, covering lighting, occlusion, look-alike products and unit conversion errors.

18-2 Medication safety
12-1 · 2026-11-14

Suicide Prevention Research for People with Mental Illness

Align event markers and note excerpts by time in offline research. Test for information leakage after the reference time and missed signals across subgroups using controlled examples.

12-1 Acute crisis signals

All 12 questions and targets →

Newest records in the ledger

25 ledger entries. The eight most recent come first.

Show all 25
  • 2026-11-14 · RunL-09 Wrote design specifications for 12-1 to 12-4 and ran their reference implementation on fixed synthetic inputs. A separate audit program recomputed the numbers by other methods and passed four gates.
  • 2026-11-14 · CorrectedL-10 Correction: a state-machine property test found that a session already in a terminal state could repeat its notification after a policy fault. Effects now fire only on entering a state, with a regression test and an audit rule.
  • 2026-11-14 · BuiltL-11 Rebuilt the site as IATHEON screens: one screen for each menu item, in the same nine sections, and linked the 12-1 to 12-4 records to their specifications and evidence.
  • 2026-11-14 · RegisteredL-12 Fixed the 40-gene panel and split it into 20 development and 20 held-out genes by the SHA-256 order of the gene symbols. Rules are adjusted on the development genes only.
  • 2026-11-14 · LockedL-13 Froze six input files and recorded their SHA-256: the ClinGen expert classifications and panel VCF, the annotation store, the simulated cases and their VCF archive, and the literature text windows.
  • 2026-11-14 · RunL-14 Classified 8,667 expert-classified variants and ran 480 simulated patients through the agents and the audit gate, then injected 9,493 unsupported claims and 360 ledger edits. Repeated runs give a byte-identical result file.
  • 2026-11-14 · RunL-15 Ran a language-model curator on 24 held-out variants under two instructions, and a literature reader on 96 sampled variants with a second reader from another model family. The stored answers are audited again on every evaluation.
  • 2026-11-14 · CorrectedL-16 After cross-model review, removed the argument that let a caller hand the audit its own ledger check. The gate now verifies every ledger itself and works only on the verified copy.
  • 2026-11-14 · CorrectedL-17 After cross-model review, removed the option to admit literature findings without the second reader and split checking from admission. Findings enter the ledger through one function that requires the second reader’s verdicts.
  • 2026-11-14 · CheckedL-18 Recomputed the audit record from the stored result and the frozen input hashes and checked four gates. The check command reproduces the audit file byte for byte.
  • 2026-11-14 · RunL-19 Ran the 16-2 virtual-cohort Monte Carlo: 1,000 synthetic patients per run, three seeds, 14 distribution shifts, baselines, fault injection and five protocol cases. 962 met the convergence and consecutive-error criteria in the primary run.
  • 2026-11-14 · RunL-20 Ran the 16-3 cross-cohort evaluation on synthetic multi-site cohorts: both train/test directions, three harmonisation modes, label-permutation and leakage controls, and 10 seed replications.
  • 2026-11-14 · CheckedL-21 Audited the 16-2 and 16-3 run with a separate program that does not import the solvers: 254 exact binomial bounds, the primary run recounted from per-patient rows, five verdicts re-derived from trajectories and 21 AUROC figures recomputed from per-case scores. Four gates passed.
  • 2026-11-14 · RunL-22 Wrote the 18-2 specification and ran its audit kernel on simulated vision outputs and abstract orders. Of 4,500 injected errors 0 passed automatically; 1,859 of 3,000 correct fills did. A separate audit replayed every decision. No image inference was run.
  • 2026-11-14 · CorrectedL-23 Correction: review found that the draft kernel treated a missing audit ledger as a successful commit, that the drift monitor reused a fixed calibration table, and that out-of-range vision values could pass. The kernel now fails closed, the monitor uses online p-values and inputs are range-checked; verdicts did not change.
  • 2026-11-14 · RunL-24 Wrote the 8-1 specification and ran its harness on a synthetic multi-omics world. A separate audit recomputed 860 checks and passed four gates. The private baseline generator passed the privacy gate but kept 91.4% of rank agreement, so its release was blocked. No real cohort was run.
  • 2026-11-14 · CorrectedL-25 Correction: cross-model review found that the DP-SGD aggregate added no noise for an empty batch, that the baseline generator failed on an empty data set, and that the specification tied the sampling rate to the private row count. All were fixed with tests; verdicts did not change.

Next checks

5-1 · AI drug discovery
Select records with measured binding values, specify assay conditions and lock the split by scaffold, target family and date.
Evaluation data partly establishedunlocks: Binding free-energy error of 1 kcal/mol or less
8-1 · Synthetic data
Run the same harness on the fixed TCGA-BRCA institution split with the baseline generator, then train the main generator and require it to beat the baseline on every endpoint at the same privacy budget.
Evaluation data readyunlocks: At least 95% of downstream predictive performance retained
18-2 · Medication safety
Train the vision modules on the fixed ePillID split, build per-category calibration sets from its validation split and fill the same table on its test split; categories below the minimum calibration size are reported as uncertified with no automatic pass.
Evaluation data partly establishedunlocks: Zero missed contraindications or look-alike medication errors
19-1 · Rare disease diagnosis
Read the literature for every expert-pathogenic missense variant instead of a sample, so that functional, segregation and case findings enter the ledger. Then re-measure sensitivity on genes outside the current 40, with the condition that specificity does not fall.
Evaluation data readyunlocks: At least 90% of causal genes in the top five
16-2 · Adaptive cognitive training
Import stratum-level summary statistics inside the data-use zone, re-freeze the fixture and re-run; then test a response generator whose functional form differs from the scheduler's model.
Evaluation data partly establishedunlocks: Difficulty convergence in at least 95% of simulated episodes
16-3 · Dementia diagnosis
After approved data import, run the same harness with ADNI and OASIS-3 as the two external directions and evaluate speech separately; the synthetic figures are then replaced.
Evaluation data partly establishedunlocks: External cohort AUROC of at least 0.90

All next checks in the validation roadmap →

© 2026 RISCON Co., Ltd.Privacy PolicyTerms of UseCookie SettingsContact

RISCON Labsrisconlabs.com
RISCONAPOSTOLUSIndependent Health and Medical Research