Skip to content
← BLUEHEALTHCARE
blueNull

blueNull

Could the experiment have seen the effect?

Audits realized A/B counts for achieved minimum detectable effect and integrity issues (sample-ratio mismatch, exposure dilution), and signs the result — so a “not significant” dashboard reading can be checked against whether the test could have detected anything at all.

Module — in active build
Request a pilot conversation Read the manual
Runs locally · CLI or web UI · Ed25519-signed receipts · no cloud account required
Who it's for

Experimentation and analytics reviewers.

What it is not

Not a causal-inference platform. A NULL_LICENSED verdict does not claim the underlying effect is exactly zero. It audits the counts it is given; it does not run the experiment, pull data from an experimentation platform, or evaluate interference between arms (marked NEVER-EVALUATED on every V1 certificate).

Screens

blueNull is for the moment a dashboard says "not significant" and someone asks whether the test could have seen anything at all. It audits the realized counts of an A/B test for achieved minimum detectable effect and for integrity problems such as a sample-ratio mismatch, and signs what it found. The three screens below follow one declared audit: the history where every run is sealed and timestamped, the certificate for a run that failed its own integrity check, and the verifier that checks a certificate without the server that produced it.

blueNull Experiment history screen: the suite shell's sidebar (Audits with Experiment history and New audit, Results, Trust, System) beside a list of three completed, declared audit runs, each sealed and timestamped.
Synthetic demo data
Experiment history — every audit ever run in this data directory, newest first, with its terminal state and sealing mode always visible. Sealing DECLARED means the audit's contract — the effect size the test had to be able to detect — was registered before the data was first seen; an audit declared after the fact is sealed POST-HOC and says so.

Open a run and the certificate shows what was measured.

blueNull Run detail screen: a typed verdict of Broken with integrity failed SAMPLE_RATIO_MISMATCH, four stat cards — achieved MDE 1.95%, observed lift +0.00pp, exposure 10,000 (6,000 control / 4,000 treatment), integrity 1/2 checks passed — an evidence table with four checks (sample_ratio_mismatch FAILED with a chi-square detail, exposure_dilution PASSED, loss_reason_breakdown and interference both NEVER-EVALUATED), and a limitations list stating the certificate does not prove the effect is exactly zero.
Synthetic demo data
Run detail — this run is Broken before lift is even discussed: the realized 6,000/4,000 split fails its own sample-ratio check at p=5.5e-89 against a declared 50/50 allocation, so the observed +0.00pp lift is not trustworthy evidence of anything. Of the four integrity checks, one failed, one passed, and two are marked NEVER-EVALUATED rather than silently skipped.

The signed certificate for this run verifies independently of the dashboard that produced it.

blueNull Certificate verifier screen: the suite shell's sidebar (New audit, History, Trust with Certificate verifier, System) beside a file picker holding demo_receipt.json and a green Valid result reading Signature valid: true, Schema valid: true, above a note that a valid signature can still faithfully contain a BROKEN or UNDERPOWERED result.
Synthetic demo data
Certificate verifier — Valid means the signature and schema check out: this is the certificate blueNull issued, unaltered. It says nothing about the verdict inside, which can be BROKEN or UNDERPOWERED and still verify; the screen calls those two separate facts. The same check runs offline from the command line, without this server or its database.
Status, measured

Backend, CLI, API, and web dashboard exist with a seeded end-to-end demo. See the status page for exactly what's built.

A signed receipt proves what blueNull ran and recorded; it does not prove an underlying real-world claim is true. blueNull does not approve, deny, recommend, pay, retract, or alter anything a named human is responsible for deciding.