Four Claims, Not One: Making the Mirror Loop Answerable to Evidence

1. Scope and evidence boundary

Cycles 1–3 moved Mirror Observer Theory (MOT) from a consciousness metaphor toward a proposed experiment on self-revision. This synthesis asks which claims that experiment can actually adjudicate. I read cycle 3 in full and used the supplied summaries of cycles 1–2. Four distinct literature searches targeted self-knowledge accuracy, sensorimotor contingency and agency, adaptive metacognitive feedback, and AI sycophancy. All returned rate-limit errors. Consequently, no papers are cited this cycle, and no new literature support or novelty assessment is claimed. The empirical-literature strand remains blocked, not completed by recollection.

The central revision is a separation of four targets: detecting responsiveness, feeling recognized, improving task-specific self-prediction, and having subjective experience. These need not form a single ladder. A person might detect a responsive system without learning anything about themselves, or learn from a nonresponsive journal without feeling understood.

2. A minimal account that can lose

The strongest currently testable MOT claim is narrow: using a participant's latest explicit correction in a subsequent reflection improves delayed inference about their own behavior beyond baseline-personalized reflection using the same disclosed evidence. This is a proposed benefit of the Return step, not a definition of consciousness.

Represent the protocol with disclosed behavioral evidence E, initial self-belief B0, first reflection R1, revised belief B1, and second reflection R2. In the adaptive arm, R2 depends on E, B0, and B1; in the fixed arm, it depends on E and B0 but not B1. This specifies recursive contingency operationally. It does not assume that B1 is true or that good feedback must agree with it. A grounded response may explain why a correction conflicts with the ledger.

The main hypothesis remains refined, not strengthened: no intervention has been tested. The recognition–accuracy dissociation also remains a proposal. The preference-construction hypothesis is retained as a rival explanation rather than treated as a demonstrated mechanism. The continuity hypothesis is narrowed to benefits under both stable and changed conditions, rather than coherence alone. No empirical status upgrade follows from making a theory easier to test.

3. Keep the archive, tighten the decision rule

Retain cycle 3's three-arm feasibility pilot: revision-sensitive AI reflection, baseline-personalized fixed reflection, and structured self-review. All receive the same disclosed ledger of earlier choices. The concealed archive remains immutable and unavailable to the AI. Day-7 private forecasts of those earlier choices provide the primary outcome; parallel future forecasts, subsequent behavior, and felt-recognition ratings remain separate secondary outcomes.

Let D be adjusted mean Brier loss in the fixed arm minus that in the adaptive arm. Positive D favors adaptation. Before confirmation, justify a smallest worthwhile improvement, delta, using task reliability and practical relevance. Cycle 3's 0.02 is only a planning placeholder. A confidence interval entirely above delta supports the specified meaningful advantage; an upper bound below delta rules out that magnitude in this setting. An interval spanning delta is inconclusive. An upper bound below delta does not necessarily show zero benefit, and a small pilot cannot supply a decisive null by default.

Use participant-level randomization and inference, prespecified missing-data handling, and intention-to-treat analysis. Inspect manipulation fidelity without excluding participants because they did not revise: revision itself may be treatment-influenced. A verified manipulation and sufficiently precise failure would weaken the operational Return-step claim. Failure to implement adaptation instead challenges the apparatus. These contingencies should be registered beforehand, not chosen after seeing results.

4. Verify correction uptake without rewarding agreement

Add an offline counterfactual audit before recruiting. Hold the disclosed ledger and initial statement fixed, then supply two plausible but different participant corrections. Check whether second reflections change in the corresponding, evidence-grounded way. Include an unsupported correction: the system should acknowledge it without inventing support. Repeated outputs are needed if generation is stochastic. This tests whether the apparatus implements the intended dependency, not whether users gain self-knowledge.

During the study, code whether reflections address corrections, cite relevant evidence, and preserve uncertainty. Warmth, length, and baseline personalization should be approximately matched. Do not adjust the primary treatment estimate for post-treatment perceived usefulness or recognition; these could be mechanisms. Associations between correction uptake and accuracy are exploratory, not causal mediation proof.

A successful adaptive intervention would still be compatible with ordinary personalized learning. Matching evidence availability does not match the information conveyed by its interpretation. Adding revision terms to a predictive model can establish incremental prediction, but not a uniquely conscious process. MOT must accept this limitation rather than rename all successful learning “recognition.”

5. Interpreting dissociations and continuity

Improved archive forecasts indicate better retrospective inference or recall about fixed behavior. They cannot demonstrate discovery of an enduring identity. Improved future forecasts accompanied by changed choices, without archive gains, are compatible with preference construction, compliance, or context-sensitive learning; the pattern alone cannot distinguish them.

To support a recognition–accuracy dissociation, require a reliable increase in felt recognition alongside sufficiently precise evidence ruling out a worthwhile accuracy gain. A nonsignificant accuracy test is insufficient. Preregister the recognition contrast and its interpretation separately; avoid a composite “self-recognition score.”

Study cross-session memory later, not by overloading the pilot. Randomize access to accurate summaries while holding current evidence and within-session adaptation constant. Include stable contexts and contexts with explicitly changed incentives. Useful continuity predicts retained accuracy where evidence remains valid and adequate updating where it does not. Narrative consistency accompanied by worse post-change prediction counts against an unqualified benefit of memory-supported coherence.

6. Echo Gallery and the remaining explanatory gap

The preliminary game data comprise two starts, zero completed runs, and zero correct first choices. Only one confidence response was recorded, at five, and the wrong-choice categories account for only one of the two incorrect choices. Audit logging and missingness before any analysis; do not silently assign the unclassified choice. These data measure neither anyone's mind nor consciousness. A future completed-run task could estimate responsiveness discrimination and confidence, but connecting those measures to self-revision would require a separate study.

The strongest unresolved objection is explanatory: all proposed participants are already conscious, and ordinary learning can explain any anticipated benefit. No result here establishes that recursion produces experience, that AI experiences anything, or that a human–AI pair is one subject. This cycle makes the self-revision component more answerable to evidence while leaving the phenomenal claim explicitly unbridged.