Validation
How we measure what the system does and does not prove.
The system does not earn trust by saying it is correct. It earns trust by publishing its gates, its failures, and its freeze conditions. The active External-Claim Freeze is documented on this page.
Validation disciplines
Four things every measurement run must survive.
Every measurement run is locked to what was publicly observable at the evaluation timestamp. No future snapshots, future citations, or future contestation events cross the firewall. The question is always: what would have been knowable at the time?
Outcome labels — institutional actions such as FDA label changes, trial holds, safety communications, guideline updates — are sourced from formal public records and stored independently from the measurement pipeline. The pipeline cannot see labels during computation. Labels exist to test, not to train.
Every validation study compares against explicit baselines: news timing, consensus date, random-split nulls, and simple mention-count benchmarks. The question is not "did the system detect something?" but "did it detect something before it was already obvious in standard information channels?"
We do not report only wins. Every study includes failure analysis: when the signal was weak, when the claim identity was unstable, when contestation was insufficient, and when the gate was not met. Uncertainty is preserved, not compressed into false confidence.
The load-bearing piece
The Leakage Firewall.
The most common failure in any predictive or signal-detection system is temporal leakage — using information from the future to "predict" the past. In claim-hardening measurement, the hazard is specific: a citation that appears in week 14 contaminating the read for week 12. A label event that leaked backward into the feature computation. An outcome that was "obvious by then" because the system silently referenced the thing it was trying to predict.
The LeakageFirewall is the mechanism that prevents this. It is not a policy. It is a deterministic set of join gates, store boundaries, and holdout rules that the pipeline cannot bypass.
The evidence record is captured at fixed weekly boundaries. The snapshot for week N contains only what was publicly observable at the end of week N. No later revision is visible to the measurement pipeline.
Every computational join is gated on usable_at ≤ evaluation_timestamp. The system cannot join a snapshot from week 12 with a citation that first appeared in week 14.
Outcome labels are stored in a physically separate store with no programmatic access from the measurement pipeline. Labels are joined only during validation, after the pipeline has completed its run.
Holdout sets are defined before any validation study runs. Negative-control studies test the system against claims where no outcome is expected. If the system reports a signal on a negative control, the firewall is breached.
Active
External-Claim Freeze.
The website is not allowed to publish predictive, lead-time, or comparative-lift claims until three conditions are met in the same lane family. None of the three are currently met.
- 1 Contamination control has not been fully demonstrated on the finerenone lane.
- 2 Authoritative finerenone evidence and label readiness are in progress but not complete.
- 3 Holdout and retrospective actionability gates have not been cleared.
What is authorized today: one lane (finerenone in albuminuric CKD) carries an authorized named-state read with explicit lineage suffixes and a visible gate. The signal is 0.38. The promotion gate is 0.65. The gate is not yet met. See the Sample Dossier for the authorized read.
What is frozen: all lead-time, predictive, and comparative-lift language across all public surfaces. The Case Library carries dossier-build status only. Use Cases map workflow surfaces without performance claims. The homepage carries one specimen read with explicit limits.
What this page does not say
- This page describes the validation discipline. It does not assert that any specific lead-time claim has been validated for any lane.
- The External-Claim Freeze is active across all public surfaces. Only the finerenone lane carries an authorized named-state read.
- Comparative-lift claims require all three freeze gates to be cleared in the same lane family. No lane has cleared all three.
- Hardening is observed in public evidence systems. It is not a substitute for trial data, regulatory review, or clinical judgment.
Action
Request a technical briefing on validation methodology.
One session, scoped to validation architecture and the current freeze posture. Not a sales pitch.