Report
Continuous Thought Machines: A Research Atlas
A source-grounded dissection of Sakana AI's Continuous Thought Machines, paired with an inspectable parity protocol and an explicitly disputed local reproduction state.
Verdict Inconclusive
- Published
- Jul 2026
- Verdict
- Inconclusive
- Source commit
- de4321e621df…
- Verification
- Verified
- Question
- What exactly did the CTM paper contribute, and which local reproductions can responsibly inherit from it?
- Intent
- Ground the Sakana research atlas in mechanisms and public artifacts instead of vague inspiration.
- Method
- Summarize the paper, map concepts to local experiments, and attach the public notebook artifact.
- Next test
- Use the atlas to design a parity-aware reproduction before applying CTM ideas to Agora or WCCS.
Continuous Thought Machines: A Research Atlas
Question
What does Sakana AI’s Continuous Thought Machine actually claim as a temporal computation system, and what can a local validation-only parity protocol establish without pretending to paper-parity reproduction?
Context
CTM is best understood as a temporal computation system, not a conventional feed-forward network with a recurrent wrapper. Hidden state evolves through internal ticks; neuron-level models read that history; synchronisation is treated as a usable representation. Primary sources are the Sakana publication, arXiv paper, and reference implementation.
This atlas is Club B infrastructure for the Recursive Self Improvement Lab: import an external RSI-adjacent architecture into a verifier-first local evidence pipeline.
Findings
- The paper’s distinctive move is relative timing of neuron activity — pairwise synchronisation and temporal statistics — as a readout signal.
- Local atlas does not claim to reproduce the paper. It implements a validation-only parity protocol: deterministic sequences, cumulative-parity targets, first-max certainty, explicit tie handling.
- Gate ledger: G0/G1 passed, G2 pending. A conservative custody audit found capacity/provenance drift that makes the stronger “no training / paper-parity” interpretation unsafe. Status: disputed, not reproduced and not falsified.
Method
Source-grounded concept separation (internal time, neuron-level temporal models, synchronisation, pair subsampling, temporal decay, certainty/adaptive compute), plus an inspectable marimo atlas that exposes protocol constants and generated validation examples without importing a training runner or checkpoint.
Limits
- Not a paper-parity result. Disputed local custody is evidence about local process, not a refutation of Sakana’s reported results.
- Falsification backlog (matched baselines, synchronisation ablations, tick sweeps, sealed seeds/hashes) remains open.
- Mechanism separability: more ticks ≠ synchronisation caused the gain.
Artifact
Interactive research atlas at /notebooks/continuous-thought-machines/ (source commit de4321e, notebook sha256 c21f990e…). Every displayed number is a declared protocol constant or a generated validation example.