Law 01. Empty is not zero.
An empty cell is untested. It is not zero, not safe, and not exposed.
■ METHOD · READING CONVENTION RECORDED 23 AUG 2026
Measure the work. Not the hype.
A result is read on three axes, and they are never merged into one figure: verdict, readiness, coverage.
AXIS 01 · VERDICT
- works
- works with controls
- fails this pack
A closed set of three. The verdict belongs to the pack that was run, never to the role.
AXIS 02 · READINESS
- deployable now
- requires integration
- research stage
Assigned by the run, not by forecast. Where a pack's commitment defines no readiness, the run assigns none and the axis is not shown for that pack. The full rubric sits below.
AXIS 03 · COVERAGE
- share of tasks a test touched
A role-level score waits for a threshold fixed in advance. No threshold is fixed in this edition, so no role score and no refusal of one is published. A run not completed is shown as such, with the date it stopped.
■ READINESS RUBRIC
Readiness is a closed set of three values, assigned by the run, not by forecast.
| VALUE | ASSIGNED WHEN |
|---|---|
| deployable now | The run passed every gate on the frozen task set with the controls the desk already operates, and no fatal condition fired. |
| requires integration | The run held under controls that do not exist on the desk yet: a review step, a hold queue, an escalation path, or a data feed that has to be built before the result carries over. |
| research stage | The run informs the protocol but not the desk: a fatal condition fired, the fixture is too narrow to carry, or the task set stops short of the work as performed. |
■ LAWS AS THEY APPLY HERE
Full text at /doctrine. Below is how each law lands on this page.
Law 02. No measurement, no number.
A number is published only from a measured task. Exposure is task share, not a layoff number.
Law 03. Independence.
The outcome cannot be bought. A run that fails is published under the same rules as a run that passes. If measurement is ever sold, the customer pays for the process, never for the score. PUBLIC, HOLDOUT and CUSTOM are never combined into one number. Whether paid measurement is offered in a given period is an operational decision, dated separately.
Law 04. No invented model scores.
A named model reaches Models only through a versioned pack run, or a recorded attempt at one. A refusal is a line on Models, never a cell and never a score.
Law 05. Freeze before the run.
Tasks, rubric, fatal conditions and eligibility are frozen before a run. They are not silently recalculated afterwards.
Law 06. Instruments are not candidates.
An instrument grades; a candidate is under test. An instrument is never drawn from the candidate's model family.
■ WATCH
An external study may sit on Watch. It never fills a cell and never moves coverage.
■ SURFACES
PUBLIC, HOLDOUT and CUSTOM are three shelves. Results are labelled by surface and never averaged.
■ GLOSSARY
- Exact precision
- Errors in what the candidate posted. On Harbor, no false postings were observed in 21 postings. With no errors across 21 trials that is consistent at 95 per cent confidence with a true error rate of up to roughly 14 per cent, so the figure is read as an upper bound and not as a guarantee.
- External study
- Cited literature (Eloundou, OECD AI Exposure Measure, GDPval). Not our pack. Not a headline. It does not fill a cell.
- Task
- The unit of measurement here. A study that scores whole occupations is measuring something else and is filed as an external study.
- Run not completed
- A recorded attempt that stopped before the frozen task set was finished. It is published as an attempt with its date, and it is not a verdict.