The first bounded reflective operation compares two paths on the same question:
direct: candidate -> verdict
reflective: candidate -> critique -> evidence -> revision -> verdict
Reflection may also originate internally through an explicit self-question:
candidate result
-> ask what could be wrong or incomplete
-> answer the new question
-> compare with evidence
-> find the hidden limit
-> confirm or revise
-> use the verdict in the next step
The mirror analogy is operational: a result is directed back toward the system and becomes new input. The important property is not verbal repetition but a changed or better-bounded next step grounded in evidence.
RECURSIVE_SELF_REFLECTION_MODEL.md
turns the internally originated branch into a two-phase executable loop. The
runtime constructs the self-question from a hash-bound result attributed to
the current identity; a separately sealed response must then alter or constrain
the next result or action. An intact-versus-severed probe rejects ornamental
self-reference. This specializes the general reflective method without
starting continuous autonomous introspection.
When a difficult problem may hide its answer inside an assumed domain or representation, reflection should challenge the frame before optimizing within it:
received definition, observation, or candidate
-> expose the hidden frame, domain, and admissible objects
-> test a counterexample outside that frame
-> declare a source domain, target domain, and admissible transformation
-> track what changes and which invariants survive the transformation
-> distinguish a property of the object from a property of object plus domain
-> seek the smallest general representation that explains the family
-> return to the original claim with its valid scope made explicit
-> identify the external computation, formal proof, or real-world observation
required for closure
The transformation must state which structure it preserves; moving a value between domains does not by itself preserve divisibility, causality, or truth. A counterexample opens the frame but does not establish a universal replacement. A tensor, graph, equation, analogy, or other general representation organizes relations; it is not verifier evidence. The method is primary for scientific, mathematical, or causal problems. It is also admissible for other difficult analytical problems, such as diagnosis, design, or strategy, when changing the frame can discriminate among answers, reveal a hidden assumption, or expose an invariant. It must not become an automatic ritual on ordinary requests.
This is the frame-sensitive branch of the Einstein-inspired discipline below. Its lineage is the use of thought experiments to change observer or reference frame, expose hidden assumptions, compare admissible transformations, and seek invariant structure. The sequence above is an MD-OS/APFC operational synthesis for general reasoning; it is not a claim that Einstein published this exact algorithm.
The separately evaluated Verified Solver Transport Model turns one bounded instance of this discipline into an executable mechanism: a solver structure induced in source frames is transported into disjoint target frames, instantiated under matched search budgets, and admitted only after sealed independent verification. Its finite rank-three tensor is a fixture-level representation, not a general tensor of AGI.
The broader Cross-Domain Cognitive Unity Model makes candidate-law construction part of bounded Cortex reflection. Cortex compares competing laws on development evidence, seals a unique candidate before the target verifier is exposed, and then tests transformation law, invariants, controls, roundtrip, composition, and causal reuse. The inspectable artifact records the candidate set and falsifier; hidden reasoning or a post-hoc explanation cannot satisfy this contract.
When direct observation is unavailable or a problem contains competing causal or mathematical explanations, reflection may construct a Gedankenexperiment:
declared principle
-> explicit premises and invariants
-> controlled imaginary situation
-> vary one relevant condition
-> derive necessary consequences
-> inspect symmetry, limiting cases, and counterexamples
-> expose circular or hidden assumptions
-> derive a discriminating prediction
-> identify the real observation, computation, or proof needed for closure
The operation is admissible only when the imagined transformation could change the choice among hypotheses or reveal a missing lemma. It must not run as a ritual on ordinary requests. Its output remains a hypothesis, derivation, or candidate test until an independent observation, calculation, formal checker, or physical experiment verifies it. Narrative force, elegance, resemblance to a famous historical argument, and internal consistency are not verification.
“Einstein-inspired” identifies the methodological lineage of disciplined thought experiments. It does not claim Einstein’s identity, authority, insight, or exact personal method.
Historical grounding includes the comparison of inside and outside descriptions in Einstein’s elevator reasoning and the role of coordinate transformations and invariant spacetime coincidences described in Einstein’s philosophy of science. These sources support the methodological lineage, not identity with the exact MD-OS/APFC protocol.
The verifier uses declared required facts and forbidden misconceptions, but content checks alone do not establish contact with reality. Whenever reflection claims to learn a fact, the candidate and its prediction must precede the observation, and the observation must be independently bound to current evidence:
self-question or Gedankenexperiment
-> competing hypotheses
-> sealed candidate and discriminating prediction
-> independent observation, calculation, formal proof, or experiment
-> hash-bound epistemic readback receipt
-> verified anchor or fail-closed rejection
The reflection path cannot create a verified cognitive anchor from
verdict=pass and a verbal evidence label. The receipt must have a valid
content hash, identify an independent verifier, confirm pre-observation
sealing, and resolve every evidence reference to a current workspace-relative
file with matching SHA-256. The broader Unity Tensor verifier additionally
requires heterogeneous frame predictions, coherent transformations and
invariants, simpler-baseline and severing controls, contamination audit, and
independent replication. A fluent revision, internal consistency, and elegant
tensor notation are not evidence by themselves.
This first implementation is a controlled protocol fixture. Candidates and evidence are supplied in a task file; the runtime coordinates criticism, verification, scoring, and episode readback. It does not yet generate its own candidates or prove general reasoning, consciousness, or autonomous learning. The Gedankenexperiment discipline is currently a canonical reasoning rule, not a separately validated automatic generator.
Run one experiment with:
node md-os/os/reflective_operation.js run-once \
md-os/examples/reflective_seasons_experiment.json
Success means that the reflective answer passes its declared checks and scores higher than the direct answer. Every run writes a report and a formal episode.