v0.9.1 release notes

v0.9.1 is an agent role refinement release. Producers, the Leader, and QC now work to a per-operation standard. Every task is classified by the kind of work it is — building, improving, fixing, measuring, explaining, researching, assessing, operating — and that classification selects three things: the definition of “done” the work is judged against, the approach guidance handed to the producer, and the bar the Leader and QC verify against.

Still Beta — the Beta calibration still applies

This release sharpens how the team is held to standard; it doesn’t move the engine’s ceilings. The honest limits on the Beta calibration page still hold. This page is the delta on top of v0.9.0.

---

One bar per kind of work, not one bar for everything

A single generic standard can’t tell whether a fix worked, whether a research task is grounded, or whether an assessment is earned. So the engine now commits the right standard for the work:

• a fix is judged on the reported problem actually being gone on a fresh check — not on surrounding work merely passing;
• a research task on its sources being real and synthesized, not fabricated or merely collected;
• an assessment on every judgment tying to specific evidence, not generic opinion;
• a build on doing what was asked, correctly and completely.

The classification is orthogonal to the artifact kind — you can fix code or data or a document — so it stays product-agnostic.

How it holds the team

• Triage → bar. The engine stamps each task’s class of work and derives, deterministically, the definition of “done” the verifier judges against.
• Producer conformity. Each producer’s brief carries a terse, product-agnostic approach card for that class of work — so a team of producers works to a consistent discipline regardless of model strength.
• Leader & QC. Both judge the deliverable against the standard its operation demands.

Safe by default

An un-classified task defaults to a strict general standard — never a loose one. There is no soft “generic” path a producer can fall into; the default is chosen by the engine, not the producer. Work that declares no operation behaves exactly as before.

---

Fully reviewed across coherence, hull, and code passes. 4374 tests pass. Builds on v0.9.0 — stability + reporting. Read the Beta calibration page before serious work. Bug reports and discussions welcome at the issues tab and discussions.