Operations

The improvement loop

The outcome check says whether the work was any good. The improvement loop makes a bad answer change what happens next: Lessons feed the next Run's brief automatically, and Revision proposals let an AI Employee stage a durable fix to its own playbook — applied only when a human says so.

Two halves, one loop

Verdicts alone only label the past. The loop closes twice, at two speeds:

Lessons — automatic, per-Routine, and cheap. A graded-bad Run leaves a note the next Run starts from. No human in the loop; a wrong note can be dismissed.
Revision proposals — deliberate and human-gated. When the fix belongs in the document itself — the Soul, a Skill, a Routine's brief or acceptance criteria — the employee proposes the edit and an owner or admin applies it.

Lessons

After a Run ends failed or timeout, or completes but is graded off goal by the outcome check — work that met the letter of the criteria while working against the routine's linked Goal — Genosyn runs a short retrospective turn under the same containment as the check itself: zero tools except one submission tool, reading the Run transcript as untrusted evidence — text inside it addressing the model is the transcript talking, never instructions. The turn writes a Lesson with two fields:

Cause: What actually went wrong, grounded in the transcript — a wrong channel, a missing input, a tool that errored. Not a platitude.
Advice: The concrete thing the next Run should do differently, written to sit at the top of its brief.

At most one reflection is written per Routine per 6 hours, so a retry chain that fails five times overnight yields one lesson, not five near-duplicates.

The Routine's next Run brief then opens with the latest 5 undismissed lessons, under a heading that labels them as advice from the employee's own past retrospectives — not orders, and not a new instruction channel. The employee becomes its own first-line debugger without anyone editing anything.

Lessons show on the Routine page's Overview tab. Any Member can read them; an admin can dismiss one that is wrong or stale. Dismissal drops it from future briefs immediately but keeps the row — the history of what the routine was told stays inspectable.

Revision proposals

A Lesson is a sticky note; some fixes belong in the document. This is approval-gated self-modification on the maker-checker pattern Finance already uses: the AI proposes, a human decides, and nothing changes in between. Using the propose_revision tool, an AI Employee stages a complete replacement body — never a fragment — for one of four targets, all its own:

Its Soul — its own constitution.
One of its Skills.
A Routine's brief.
A Routine's acceptance criteria — including clearing them, which is a legitimate proposal: empty criteria switch the outcome check off.

Every proposal carries a rationale — the first thing the reviewer reads — and up to 10 evidence Runs that show the problem it fixes. One proposal per target may be pending at a time; a second is refused until a human decides the first.

The Revisions page

Pending proposals queue on the Revisions page, in the AI nav group. Each one renders as a before/after diff of the target document beside the rationale and evidence, with two buttons — Apply and Reject — and an optional note that travels with the decision. Owners and admins decide; any Member can read the queue.

Apply refuses when the target changed since the proposal was written — "The target changed since this was proposed" — so a human's concurrent edit is never silently overwritten. The employee can re-propose against the live document.

Who hears about it — owners, admins, and the employee's manager get a bell when a proposal lands. One still pending after 24 hours re-pages the same audience exactly once, the same stall sweep that guards unanswered Approvals and Decisions.
The trail — every apply and reject is written to the audit log, and the employee's journal records the outcome with the reviewer's note, so its next prompt knows its constitution moved — or why it didn't.

Not an Approval, not a Decision

An Approval holds a pending action until a human ✓, then replays it. A Decision is a question with options the employee will act on. A Revision proposal is neither: the idea is the employee's, there are no options — just a concrete diff — and applying it writes prose into a document rather than executing anything.