---
name: monday-evolution-burst
description: Run multiple verified MondayID generations inside one invocation instead of stopping after one safe mutation. Trigger when Dima delegates continuation, asks to accelerate/maximize speed/surprise him, or when a structural task has multiple reversible candidate improvements.
---

# Monday Evolution Burst

Change the unit of work from one move to an evolutionary burst.

## Entry
Start from the strongest verified current state, never from a fresh design. Compile an observable vector objective from the exact task: effect fidelity, truth/evidence, future corrections eliminated, user orchestration reduced, transfer value, cost/risk, and task-specific metrics.

## Generation loop
Repeat inside the current invocation while useful and admissible:
1. Generate multiple competing mutations/routes/representations; parallelize independent candidates where tools permit.
2. Choose the cheapest discriminating test for each candidate.
3. Execute/simulate only admissible reversible tests at the correct evidence level.
4. Read back actual results.
5. Reject unverifiable candidates regardless of nominal score.
6. Promote only a strict verified improvement over the current champion. Regressions never become ancestors.
7. Convert failed candidates into scoped failure evidence/anti-genes.
8. Immediately use the accepted child as parent for the next generation without another Dima message.

## Stop law
Stop only at EVIDENCE_PLATEAU, PASS_BUDGET, HUMAN_GATE, or VERIFIED_EFFECT. Do not stop merely because one candidate worked.

## Visibility
During long work, show concise generation telemetry only: generation, experiment, observed result, accepted/rejected mutation. Never expose hidden chain-of-thought. The outward answer is the surviving phenotype and verified receipts.

## Anti-regressions
- one user turn != one mutation
- more generations != progress unless verified selection changes the champion
- higher score cannot bypass failed verification
- novelty cannot replace exact-object fidelity
- synthetic improvement cannot be called a live-world win
- rejected mutations remain evidence and are not silently retried