Course / Lesson 6 of 6
19 min · sanitized real case
The mirror
Objective: Mine your own sessions with deterministic rules before believing any narrative — including your own — and turn what the mirror shows into a working rule.
First, in plain language
After auditing everyone else's corpus, the mission pointed the instrument inward: our own working sessions. The question was not "what did we work on?" — it was "what do we say we did vs. what has proof of delivery?".
The answer came from a deterministic rule, not from impressions: a candidate only becomes a pattern when it occurs on at least 3 distinct dates. Intensity in a single day is a spike, not a habit.
Open the technical layer
The mirror ran over a deterministic sample of 45 sessions (out of 696 related ones), producing 2,801 dated occurrences across 11 classes — and 11/11 passed the 3-date rule. Three findings with receipts:
| Finding | Evidence |
|---|---|
| Conclusion inflation | "done" tokens ≈ 37–38% of the class base (1,823 hits) against ≈ 4% approvals — and many hits carry neither artifact nor exit code on the line. Conclusion language ≫ delivery evidence. |
| Repetition tax | Rework + repetition ≈ 8.9% of distinct dated occurrences (249/2,801). |
| Boilerplate noise | Part of the classes matched metadata and listings, not messages — the noise filter changes the result, so it is part of the rule, not a detail. |
The hard truth. The most clearly restarted work was documentary: teaching material created one day and, the next day, "redo it from scratch, ignore the previous one" — instead of closing a single surface with proof and stopping. And when real proof was requested in one session, the written status ("converged") was falsified by the objective's own checklist, all unchecked. The mirror does not accuse people; it prices narratives.
Honest limits: regex over lines is line observation, not message semantics; model synthesis is a candidate, never ground truth; and the analysis acknowledges its own method's inflation ("completed" boilerplate).
What to do with a mirror
Every finding ends in a first action, not a diagnosis: a "done" claim only counts with a command + exit code in the same turn; a status document never replaces a re-run gate; one canonical teaching tree at a time — "from scratch" requires a scope diff; rework becomes a versioned learning, with no PII.
Build your own evidence rule
What counts as a pattern? Tune the rule and test it against four candidates (counts in the format hits · dates · sessions · %noise).
The four candidates (fixed simulation data)
| Candidate | Hits | Dates | Sessions | Noise |
|---|---|---|---|---|
| conclusion-language | 1,823 | 19 | 40 | 35% |
| friday-refactor | 5 | 2 | 2 | 0% |
| declared-approval | 205 | 19 | 30 | 70% |
| feature-rollback | 66 | 11 | 15 | 90% |