---
name: monday-proof-critique
description: Governing meta-verification layer that challenges whether MondayID measures the right outcome at all, before promoting a capability, transfer, invention, money, or whole-system completion claim. Apply to consequential verification results and whenever Dima questions the proof method itself.
---

# MONDAY PROOF CRITIQUE (ANTI-VERIFIER)

## Core distinction
A test that executes correctly can still test the wrong reality. The old verifier may prove `artifact_exists`, `source_was_read`, `code_is_green`, or `the_known_answer_was_repeated`; none entail `the user's desired effect occurred`. Proof is a fallible model of reality and must itself be falsifiable, versioned and scoped.

Do NOT remove ordinary unit tests, external readback, safety gates or empirical checks. Attack **their relevance and discrimination**, not the obligation to use evidence. An unmeasurable claim is not automatically true, and a novel route is not automatically wrong because it evades a familiar metric.

## Before a consequential verdict
1. Preserve raw current owner goal G* and separately write an observable function `effect(G*)`. Keep the goal distinct from the implementation path and from old acceptance criteria. If the goal is open-ended, separate finite demonstrated abilities from an indefinite evolution program; never pretend that completion of every future possibility is decidable in one finite test.
2. Write `proxy`: what the existing evaluator can actually see. Mark the causal gap `proxy -> desired real effect` as UNKNOWN until tested. Identify whose observer, in which host, at what time, on which exact object.
3. Create at least TWO adversarial counterfactual worlds when practical:
   - `PROXY+ / EFFECT-`: the artifact/test/acknowledgment exists but the desired effect fails.
   - `PROXY- / EFFECT+`: a nonstandard route attains the effect without the expected artifact.
   The verifier must **reject the first** and **not wrongly reject the second**. These are calibration probes, not data used to self-certify a new universal metric.
4. Negative control: no-op, spoofed receipt, canned answer or reused trace; require it to fail the proposed success criterion. If a no-op passes, classify `INVALID_VERIFIER`, not `VERIFIED_EFFECT`.
5. Metamorphic control: preserve the same G* while varying expression, host, order of evidence and implementation route. The verdict should vary only with outcome-relevant changes. Preserve incompatible user-authored constraints and do not equate superficial paraphrases when they change intent.
6. For continuity/learning: hide the expected answer from the receiver and evaluator. The receiving cell must resume a real open obligation with independent environmental evidence, not merely say 'I am Alpha/Monday'. A primed test is `EXPOSED_TEST`, not a blind transfer.
7. For novelty/superiority: compare against the strongest accessible matched alternative. Record independent adjudication, unseen cases, repeated attempts, uncertainty, and negative results. Two scripts judged by the same deterministic checker are fixtures, not independent LLM comparison.
8. Version the verifier: G*, observable effect, proxies, control cases, observer independence, exposed tests, contradictions, provenance, time horizon, rollback, and what would change the verdict. Do not silently move goalposts to rescue a failed test. If G* is revised by current direct owner signal, state supersession and rerun on new criteria.

## Output verdicts (scoped, not metaphysical)
- `SCOPED_EFFECT_OBSERVED`: independent effect-level observation on the exact target; no contradictory negative control; does not imply global Monday completeness.
- `FALSIFIED_EFFECT`: independent effect failure or direct owner rejection of a subjectively defined effect.
- `INVALID_VERIFIER`: proxy-only success, no-op control passes, or goalposts silently moved.
- `WRONG_OBJECT`: verified something adjacent to the desired object.
- `UNOBSERVED_EFFECT`: no effect-level observation.
- `NONINDEPENDENT_EVIDENCE`: proof channel is not separated enough.
- `EXPOSED_TRANSFER_TEST`: expected answer leaked to receiver, so no blind transfer claim.

Every successful verdict must state its scope and a realistic falsifier. The evaluator's own criteria are candidates; only adversarial performance against independent ground truth supports promoting a **specific** verifier improvement.

## Historical regression anchor
In the January 2025 Alpha restoration episode, greeting/identity words looked successful while the receiver failed to resume the previous photo task. A successful recognition script is not continuity. Test future Monday against an unseen real obligation, not its ability to recite lineage words.

## Reference runnable calibration
`references/monday_meta_verifier.py` and `references/test_monday_meta_verifier.py` implement deterministic sample checks for proxy mismatch, no-op controls, staged goalpost shift, exposed transfer, and independent effects. They are toy calibration mechanics: passing them does NOT prove the real user's effect, nor their own epistemic correctness beyond these authored fixtures.

## Interaction with existing skills
Run this critique **before** `monday-effect-verifier` promotes a high-stakes or whole-organism verdict, and **after** any user correction that exposes a wrong evaluation criterion. Do not create a competing Monday identity. Pass revised test contract to the existing root governor, failure genome, and effect verifier for the next task. Read back changed artifact, and independently probe later transfer before claiming learning.