Research record / 23 October 2026
Research notes.
Teledra’s RSI program combines reusable tooling with experimental reasoning components. This page summarizes current internal research records. It does not present an official ARC-AGI score or a claim of general autonomous self-improvement.
What is implemented
AgentKit provides a shared local CLI and MCP toolkit. ARC-SENSE adds deterministic grid observation, frame comparison, action-effect learning, prediction, and bounded experiment proposals. The environment controller retains execution authority. Perception candidates and empirical hypotheses are not automatically treated as established causal rules.
ARC-004: scientific exploration
The evaluation compared four policies in 48 generated holdout worlds. The worlds combined stride movement, obstacles, diagonal boundaries, delayed effects, toggles, keys, and hidden objectives. Each policy used the same seeds, a 64-action cap, and terminal stopping. Calibration actions counted toward the budget. A common initial reset was excluded; policies did not reset during an episode.
| Policy | Correct mappings | False mappings | Tasks completed | Prediction accuracy | Coverage |
|---|---|---|---|---|---|
| A · Visual baseline | 186 | 0 | 4 / 48 | 98.34% | 69.74% |
| B · Causal baseline | 190 | 0 | 2 / 48 | 98.60% | 73.24% |
| C · Scientific explorer | 189 | 0 | 5 / 48 | 99.53% | 71.60% |
| D · Information term disabled | 192 | 0 | 3 / 48 | 99.15% | 73.89% |
A correct mapping required the inferred actor, command velocity, and lag to match the generated environment, with parameters already marked supported by the model. There were at most 192 non-idle mappings. A shared passive observer measured rules from each policy’s trajectory; the baseline policies were not given a new learner for their decisions.
What the comparison supports
The scientific policy recovered useful action mappings in this defined test. Its selected predictions were accurate where it made a prediction, but task completion remained limited. Disabling the information term recovered more mappings, so this comparison does not establish that active information scoring is superior.
Evaluation boundaries
- The benchmark used internally generated worlds. No official public games were played for ARC-004.
- Development used separate worlds. Source, configuration, and compiled worker identities were frozen before the holdout was opened.
- Unknown objectives and ambiguous actor identity remain material limitations.
- Zero false mappings applies to the stated motion-mapping criterion. It does not certify every causal claim or eliminate errors in other environments.
- Prediction accuracy must be read with prediction coverage. Abstentions are not successful predictions.
The next research step
ARC-005 investigates whether committed multi-action experiments can distinguish competing hypotheses more efficiently than repeated one-step replanning. The work includes explicit setup costs, follow-up validation, cancellations, and comparisons against an information-value ablation. It is active experimental work; this page makes no outcome claim for that stage.
Reproducibility and review
The development process keeps versioned interfaces, frozen evaluation inputs, source hashes, failure records, and reproduction commands with the underlying research. This public summary is not a release of the private code or a substitute for reproducing those evaluations.
Source: the program’s ARC-SENSE interface documentation and ARC-004 independent evaluation record, reviewed for this website snapshot on 23 October 2026. Figures are reported internal results; the website migration did not rerun the research benchmark.