Skip to content

Trust, measured

BioSignal does not ask you to trust it. This page publishes what BioSignal holds, how often its answers are right, how often they are wrong, and what it has not yet earned the right to claim. Every number below is generated by an audit that runs against the live repository. None of them is typed by a person.

Measured 18 July 2026

The library

What BioSignal currently holds. Counted from the live knowledge graph.

Knowledge objects

246

Signal Records

Interventions. These carry calibrated verdicts.

61

Foundations

Clinical Monographs. These carry calibrated verdicts.

14

Conditions

83

Biomarkers

35

Comparisons

36

Body Systems

12

Peptide Families

5

Retrieval quality

How often Ask BioSignal is right — measured against 596 real consumer health questions, written the way people actually type them, across 22 clinical specialties. The corpus was not chosen to be easy, and it is allowed to fail.

Questions evaluated

596

Correct answers

Answered, and with the right object.

447

Retrieval precision

Of the questions Ask answers, the share it answers correctly.

99.1%

Retrieval coverage

Of all questions asked, the share BioSignal can answer at all. The rest is the editorial backlog.

75.0%

Misroutes

Answered confidently — with the wrong object. The only unacceptable outcome.

4

False declines

BioSignal holds the answer and failed to find it. A retrieval bug, not a content gap.

0

Honest declines

BioSignal has not published on the subject and said so. Correct behaviour — and the list of what to write next.

145

The four things that can happen when you ask a question

Correct answer
BioSignal holds an object that answers the question, and returned it.
Honest decline
BioSignal has not published on the subject, and says so rather than guessing. This is working as designed. A system that always answers has taught you that its confidence is free.
False decline
BioSignal holds the answer and failed to retrieve it. A bug in retrieval, not a gap in the library. It costs you a session and it costs us nothing but embarrassment.
Misroute
BioSignal answered confidently with the wrong object. This is the only unacceptable outcome, because you cannot detect it. A decline costs a session; a misroute costs the platform. We publish the count.

Editorial integrity

The standards BioSignal holds itself to, and whether it is currently meeting them. Each row is an audit that runs on every commit.

Editorial System

v1.0 (frozen)

Product & Design Standard

v1.0

Signal Records meeting the publishing standard

Seven mandatory sections. A record missing any of them is not publishable.

61 / 61

Records with verified references

Every reference is checked against its source. We do not fabricate citations — and we do not pretend to have them. This number is low, and it is published anyway.

32 / 61

Ask regression audit

Every previously-fixed query still routes correctly.

141 / 141

Round-trip integrity

Every object in the graph can be retrieved by asking about itself.

229 / 229

Forbidden-route audit

Routes that must never fire — cancer, chest pain, acute symptoms — asserted on every commit.

Passing · 448 routes

Broken graph edges

Links between objects that resolve to nothing.

0

Last editorial review

July 2026

Objects updated this month

Not measured. BioSignal has no publication log yet, and counting code commits would measure something else. We would rather show nothing than show a proxy and call it this.

Not measured

External accountability

What BioSignal has not yet earned. These rows will read “Not yet” until they are true, and they will not say “coming soon.”

Named clinical reviewers

Not yet

Editorial board

Not yet

Independent clinical audit

Not yet

BioSignal content is written to a physician-grade editorial standard — a five-phase pipeline, a published evidence-rating framework, and citations verified to source. But no external clinician has yet signed off on the library. Until one has, this page says so. The absence of a claim is itself a claim, and we would rather make it out loud.

How these measurements work

The audits run against the live repository. They compile the real retrieval engine from source on every run — there is deliberately no way to run them against a cached build, because a check that reads a stale copy of your code is not checking your code. It is checking your memory of it.

Every number on this page is generated, not entered. A separate audit fails the build if this page ever contains a hand-typed statistic. A trust metric a person types is a marketing claim; a trust metric an audit emits is a measurement.

The coverage corpus is allowed to fail, and it is not a gate. It is a set of 596 real-language questions BioSignal did not choose to be easy, kept deliberately outside the build gates. The moment a measurement becomes a target, people start making the number go up instead of making the product better — and the cheapest way to raise a coverage score is to delete the hard questions.

Passing every regression test and answering real questions are different things. BioSignal passes 141 of 141 of its regression suite — and every query in that suite is one it has already fixed. That is why it is not the number on this page. The corpus is, because the corpus is the one that can still surprise us.

Check us