v0.4.0 release notes

v0.4.0 makes the team learn from its own failures. When the same kind of defect keeps coming back, the correction stops being re-derived every run at a real token cost — Modulatio codifies it into durable skill guidance that producers load next time. The smart model’s repeated fix becomes the rising floor for every cheap producer.

This is the QC-thesis reaching its natural conclusion. v0.2.0 made Quality Control (QC) a fixer — the smarter seat reviews cheap work and patches only the errors. v0.4.0 makes those fixes compound: a correction earned once is paid for once, then it’s in the library for good.

The team learns from its own failures

At the end of a run, the Leader reads the team’s recent QC fail verdicts — the defects QC has already caught and rejected — and judges whether any problem recurred enough to be worth remembering: roughly three or more instances of the same kind of defect, across tasks. A single mistake is not a lesson; a pattern is.

When it finds one, it codifies the lesson into the skill library — either improving an existing skill (it appends a “Learned” section and bumps the skill’s version) or creating a new single-purpose one. The guidance is written as a general rule a producer follows to avoid the defect, not a recounting of the specific instances, so it applies to any future work of that kind.

Recurrence is the model’s judgment over the log, not a mechanical counter or a special tag the engine has to emit. That’s deliberate: the seat that reads the failures is the one that decides what they mean. The loop is on by default and runs once per kickoff, at the very end, after the work is done.

Your skill library is now git-backed

Every codification is versioned and committed to your skill library, which is now a git repository. A lesson the team earned at a real token cost is never lost, and it is always revertible — if a learned rule overreaches, the run’s Product Quality Report points you at it, and git puts the old version back.

The git layer is best-effort and stays out of the way: it never touches your global git config, and on any system without git it simply does nothing rather than failing a run. Learning is a feature, never a fragility.

Why nothing re-reviews what the team learned

A natural question: if QC reviews everything else, why doesn’t it review the skill the Leader just drafted?

Because QC already voted. The lesson is distilled from the very fail-verdicts QC kept issuing — its repeated rejections are the signal that this needs fixing. Re-checking the drafted skill would just count the same QC opinion twice. And the Leader is the strongest seat at the table; having a weaker QC gate its judgment would invert the capability floor — the smartest model marking its homework against a less capable one.

It’s the same shape as QC-as-fixer, where the Leader doesn’t re-check the fix QC authored. The engine binds the parts that should be mechanical — the codification is versioned, git-committed, and revertible — while leaving the judgment to the model best placed to make it. And runtime QC still reviews every artifact the new skill later influences, so a bad lesson can’t hide: it shows up as worse output, downstream, where QC catches it.

If a learning loop ever stalls silently — a bad key, a network blip — it leaves a breadcrumb in the activity log (skill_codification_skipped) so the silence is diagnosable, without the loop ever being able to break a run. You can turn the whole thing off with MODULATIO_SKILL_CODIFICATION=0.

What it means for you

Your team gets quietly better at the work you actually give it. The corrections your strong models make stop evaporating at the end of each run and become standing guidance your cheap models load the next time — so over a project’s life, the same mistakes stop recurring and the cost of quality falls. It’s the QC thesis closing its own loop: pay for a fix once, keep it forever.

It was proven on a live run — a real Leader read a team’s repeated failures on a multi-company research brief (one company’s section kept shipping as a placeholder), recognized the pattern, and improved the research skill with a general rule about never leaving a section unfinished. No prompting, no bookkeeping; the team taught itself.

For the full engine contract and the honest ceilings, the Beta calibration page remains the source of truth. The CHANGELOG has the precise commit-level delta.