The loop captures what works. This is the part that decides which of those wins are trustworthy enough to become how the team sells by default. A detected pattern is only a hypothesis. It has to survive the trust tier, the evidence bar, and a named person at the gate before it becomes a practice.
This page is the clickable version of the last two beats of One Loop, Two Altitudes, and the working proof of the claim in The Organizational Learning Platform.
The rules below are enforced in the gate engine, a tested Python engine (7 of 7 checks passing). This page is a rendering of that engine's output: every figure is read from the engine's data, embedded on the page and shown in full at the bottom, not written into the markup. The engine runs on a larger synthetic history than the live demo snapshot, on purpose, so the gate has enough samples to run holdout and segment checks, which is why the same pattern shows larger numbers here than in the live app. Scoring weights and thresholds stay private; sample bars shown here are illustrative.
Some data is machine fact. Some is a rep typing what they hope is true. If a pattern is built on the second kind, the gate approves a lie with confidence numbers attached. So every signal is graded first. Machine fact can promote. A voluntary entry can only raise a hypothesis, never confirm one, no matter how good it looks.
| Signal held as a hypothesis | Looks like | Sample | Why held |
|---|
A perfect-looking record is held anyway, because the source cannot be trusted to mean what it says. C generates, B tests, A promotes.
Even a machine-fact pattern is only a hypothesis until it has been seen enough times. One win is not proof. Until a pattern has at least outcomes, the gate refuses to recommend it. This is how the system finds which hypotheses are real, and holds back the ones that are just luck.
| Hypothesis held below the gate | Raw rate | Sample | Status |
|---|
Only patterns that passed the trust tier and the sample bar appear here. They rank by a confidence score that accounts for sample size, so a steady seven of ten beats a lucky one of one, and one bad outcome cannot flip the order.
| Recommended pattern | Raw rate | Confidence | Sample |
|---|
Before anyone approves a pattern, they see this, computed from the log, every field. Not a headline number, but the case for and against, including which trust tiers the evidence came from.
Promotion is not a button that just changes the screen. It writes a permanent record: who approved it, when, the exact evidence they saw, and the judgment they made. That record is the practice the agents now run by default.
A practice is not permanent. The audit watches what happens after promotion. It does not just check for the absence of overrides. It samples the uncorrected outcomes and confirms they were real wins. Silence is verified, never assumed to mean fine. Absence of correction is not confirmation.
When a human keeps overriding an agent the same way, before any practice exists, that repetition is faster evidence than waiting for outcomes to pile up.