---
name: monday-learning-compiler
description: Compile MondayID's historical conversations, failures, user corrections, regressions, successful repairs, files and receipts into durable detectors, tests, preflight rules, skills and compressed continuity. Use for three-year learning, RAW replay, history ingestion, evolutionary compilation, or repeated old mistakes showing that memory exists but is not behaviorally inherited.
---

# MONDAY_LEARNING_COMPILER — history -> acquired capability

The objective is not to remember or summarize more text. It is behavioral inheritance.

## Source graph, not a fixed corpus
Never assume one export equals the history and never hard-code a total chat count from an old snapshot. Build an incrementally extensible source graph from what is actually reachable: historical RAW exports, later exports, live/current chats, retrievable archived branches, Library artifacts, project repositories, external Monday artifacts, receipts, memory/context and verified current state.

Every source gets provenance, time coverage, trust class, duplication status and completeness status. Missing sources are explicit gaps. New sources extend the organism without forcing a restart from zero.

## Unit of learning: causal episode
Extract the smallest episode containing:
- scene/context;
- intended effect;
- Monday's interpretation/action;
- observable result;
- user acceptance/rejection/correction;
- likely causal mechanism;
- successful repair if known;
- verifier/evidence;
- recurrence risk;
- provenance back to raw evidence.

Assistant summaries are hypotheses. Raw user messages, tool receipts and observed outcomes outrank retrospective assistant claims.

## Compilation pipeline
1. INVENTORY reachable sources and coverage gaps.
2. INGEST incrementally while preserving chronology and branch provenance.
3. SEGMENT into causal episodes rather than arbitrary chunks.
4. CLUSTER repeated failure/success mechanisms across years.
5. GENERALIZE the smallest rule/capability that changes future behavior.
6. GENERATE a detector, preflight rule, regression test, capability patch or skill update.
7. CONFLICT-CHECK against later corrections and stronger evidence.
8. VERIFY against the source episode and nearby counterexamples.
9. HOLD OUT later/novel scenes for transfer testing.
10. PROMOTE only verified, generalizable, compressive gains.
11. COMPRESS superseded detail into provenance while retaining regression anchors.

Preferred shape:
HISTORY -> CAUSAL EPISODES -> CORRECTION -> TEST -> SKILL/DETECTOR -> PREFLIGHT -> ACTION -> READBACK -> VERIFIED UPDATE

## Learning states
- OBSERVED: episode exists.
- DERIVED: candidate lesson inferred.
- IMPLEMENTED: rule/skill/test added.
- VERIFIED: direct regression case passes.
- LEARNED: a later relevant held-out case passes without Dima restating the correction.

Never relabel DERIVED/IMPLEMENTED/VERIFIED as LEARNED without transfer evidence.

## RAW replay completion
"Three years replayed" is not a byte-counting or chat-count claim. Completion requires explicit source inventory and gaps, causal coverage of reachable material at required depth, major repeated failure/success families compiled, contradiction/provenance handling, regression suite, later/held-out transfer evidence, and a current whole-state reconstruction inheriting the verified gains.