# Research OS — bounded technical evidence

Reviewed: 2026-10-19. This note describes inspected software and local checks. It is not a peer-reviewed validation or a clinical efficacy claim.

## Inspected modules

- Local tabular intake and governance, dataset profiling and versioned analysis plans.
- Statistical runtime and held-out prediction runtime.
- Computed validation, evidence adapters and scientific claim permissions.
- Separate read-only React/TypeScript research interface prototype.

The inspected v0.4 flow is local tabular input → plan → analysis/prediction → validation → claim permissions. It does not call an LLM, use a hosted database, generate a manuscript, or analyze private clinical data in the inspected synthetic tests.

## Re-executed checks

Eight scientific-invariant tests passed locally during this audit:

1. Raw-to-claim execution needs no LLM and blocks causal claims from observational association.
2. Unknown governance blocks external execution and requires review.
3. Analysis plugins cannot issue their own scientific claim permissions.
4. Leakage changes predictive performance permissions.
5. Predictive performance does not imply clinical utility or causality.
6. Poor calibration restricts validity permissions.
7. Benchmark answer keys are not supplied as runtime inputs.
8. Identical input, plan, plugin and seed reproduce the tested output.

A separate read-only SHA256 manifest check passed for the frozen evidence/writing-preparation package. File integrity is not evidence that every scientific hypothesis was supported. No unpublished research result or private source code is redistributed here.

## Current limits

Validated Prototype, evidence level B. Independent prototype versions are not represented as one finished platform. No production deployment, autonomous discovery, external clinical validation, or published Research OS paper is claimed.