Hindsight is cheap. Pre-registration isn't.
In every computational discipline that publishes "predictions," there's a quiet failure mode: the model gets adjusted, the test set gets curated, or the "blind" test set is built with the answers already known. The result looks impressive and is worthless.
Pre-registration with cryptographic locking is the structural defense. If you can't change the prediction after the answer arrives, you can't game the test.
How a prediction gets locked.
- Draft. Analyst (or model) produces a complete prediction: target metric, point estimate, 80% and 95% intervals, mechanism justification, list of relevant references.
- IBC review. 19-agent reasoning audit runs (see M1). Any hard flags block lock; soft flags are annotated.
- Hash. The locked prediction (JSON + Markdown) is SHA-256 hashed. The hash and a timestamp are committed to the public credibility ledger.
- Wait. The validation event arrives — a trial readout, an FDA action, an earnings release, a phase-3 result.
- Reveal. The locked prediction is published. The hash is verified to match the original lock. Any divergence between prediction and reality is reported honestly, including the magnitude.
- Learn. Wrong predictions feed back into model refinement, but the original wrong prediction is preserved on the ledger.
MS Study P1–P7 predictions, locked 2026-05-08.
2C83D42F22…8290645E (truncated)
Subject: Atlas Bio MS Study — 4 PD-pipeline blends repurposed for RRMS, MS-IMM-01 lead candidate.
Predictions locked: P1 (NF-κB modulation), P2 (cytokine profile), P3 (T-cell repertoire), P4 (BBB permeability), P5 (lesion burden), P6 (relapse rate), P7 (disability progression).
Validation event: Phase IIa trial readout (timing TBD; pre-registered against published meta-analysis benchmarks in the interim).
The full pre-registration document is available to verified clinicians and sponsors. The hash above can be independently verified against the public credibility ledger.
Honest failures we've published.
Pre-registration's whole point is that wrong predictions stay on the ledger. Two honest failures Atlas Bio has published:
- LEAP-002 over-prediction. v1.0 predicted positive primary endpoint. Trial was negative. The original wrong prediction is preserved; the v2.0 recalibration is documented separately.
- Alpha Predictor 96.6% → 91%. Self-discovered hindsight contamination caught by R02/R07. The original 96.6% claim is preserved on the ledger alongside the honest 91% re-test.
These are the most important entries on the ledger. They show the discipline works.
Want to inspect the ledger?
The full pre-registration ledger (with hashes, timestamps, validation events, and outcomes) is available to verified reviewers under NDA.
Request access →