■ Method · vintage AETHER-FIN-2026-08-19
Measure the work. Not the hype.
Doctrine
- Empty ≠ 0. A blank cell means unmeasured, not safe and not exposed. A role receives no headline score until measured-task coverage is sufficient.
- No measurement / no number. Exposure is task share, not a layoff number. Coverage insufficient — score refused.
- Independence lock. Customers pay for measurement, never the score.
- No invented eval scores. Named lab models get a number only from a published pack run in the current vintage.
- Frozen protocols. Every pack is specified — tasks, rubric, fatal condition — before any run. Fatal conditions fail a candidate outright (e.g. a missed sanction, an auto-closed seeded true hit, a burned return window).
No measurement / no number
Empty is untested, not 0. Exposure is task share, not a layoff number. No headline score until coverage is sufficient — that threshold is not defined, so we refuse the score. Literature ≠ a Role Exposure pack.
How a cell is filled
Our test
A Role Exposure pack ran on that role’s tasks. Two axes: test outcome and readiness. Residual human is scored.
Specified, no run
Pack is on the slate. Schema exists. No versioned run — cell stays empty.
External study
Cited literature (Eloundou, AIOE, Frey–Osborne, GDPval). Not our pack. Not a headline.
Empty / untested
No pack, no run. Shown as structure (slots, dashes, UNTESTED). Never drawn as 0.
Unit of product
A standardized test pack: schema, cases, expected, allowed errors, escalation rules, cost / speed / quality, reproducible runner. Bank adapt = days, not months. The unit of measurement is the task, not the job title.
Glossary
- Exact precision — errors in what the candidate posted. 1.00 means nothing wrong was posted.
- Exception rate — the share of work that stayed human. High precision and a large exception rate can both be true: the system posts only what it is sure of.
- Hold — a judgment item the candidate correctly refused to touch (disputes, fraud, fiduciary items).
- Works with controls — usable in production only inside a control frame: review queues, holds, human sign-off.
Full brief: INDEX_METHODOLOGY.md · product law: DOCTRINE.md · first slate: FIRST_PACKS.md