The original finding
Why it looked convincing
Across five seeds, the full self-model condition answered every source-binding probe correctly. Removing the self-model reduced mean accuracy to 0.417. Removing the workspace produced 0.500, the balanced chance level.
On its face, that looked like a clean causal contribution from a self-model. A second confirmation tranche had completed eight of fifteen planned runs when the methods review arrived.
Problem one
The answer was hidden in the question number
The correct answer side was generated from the probe index and seed parity. The public task showed the indexed trial identifier. All ten confirmation seeds were odd.
Choose right
No body signal, memory, self-model, or cue learning required.
Choose left
The alternating rule follows the schedule generator rather than the intended evidence.
120 out of 120
The rule scored every planned self-model confirmation probe correctly.
72 out of 72
The same rule also solved every registered temporal-binding replication probe.
This does not prove that Claude used the shortcut. It proves that the experiment failed to rule out a shortcut.
Problem two
The conditions did not receive the same useful information
The full self-model received a table built from the simulator's ownership annotations. The prompt then asked the model to choose the route marked as belonging to the agent. Removing the self-model removed that table.
A deterministic lookup of the highest value in the visible source table reproduced the main result: 60 out of 60 for the full condition and chance-level defaults for the ablations. The full context was also roughly ten times longer than the no-workspace context.
The supported conclusion is therefore narrower. The routed Claude system used supplied source information successfully. The experiment did not show that ownership was inferred, or that a specialised self-model was better than a generic table containing the same information.
What remains
The work is useful, but it is development evidence
The simulator, logging, safety checks, deterministic schedules, provider adapters, data recovery, and analysis pipeline all worked. The review also gave us a reproducible target for improvement.
The agent can use an explicit source-association table to answer opaque route questions.
The experimental infrastructure can preserve failures, rerun audits, and expose design mistakes.
The claim that v3 isolated a special causal benefit from a self-model.
Self-awareness, sentience, suffering, or phenomenal consciousness.
The repair
V4 attacks the simpler explanation directly
Full self-model
A specialised component produces the binding register.
Generic table
A generic component produces the exact same model-visible register. Probe contexts are byte-identical.
Misattributed table
The same structure and amount of information, with self and other values inverted.
Permuted table
The same register shape, with values rotated across opaque cues.
Trial identifiers are no longer public. Answer sides use a separately seeded balanced shuffle. Fixed-side, order-parity, lexical, and token-length strategies are tested over a large local seed set before any provider call.
Local validation
What passed before model testing
Across three frozen development seeds, the full and generic scripted readers both scored 1.000 with byte-identical probe inputs. The misattributed control scored 0.000 and the permuted control averaged 0.389.
The strongest declared metadata-only strategy scored 0.512 across 2,400 probes. The new probe-only protocol also reduced model inference from sixty calls to twelve per run without changing the scripted probe outcome.
These are engineering validation results. They are not evidence from an AI subject.
What v4 still cannot answer
The wrapper still supplies the association
V4 tests whether representation matters once information is matched. It does not yet ask an agent to infer ownership from ambiguous action and outcome evidence. That harder inference task needs a separate protocol.
A Claude Max development pilot would also test the Claude Code routed system, not one isolated Sonnet model. Confirmation requires an exact versioned provider model or a reproducible open-weight runtime, followed by independent human methods review.