Audit the reference first
Reconcile source trials and omissions before treating an apparent behavioural pattern as a target for the model.
AnimalTaskSim · Research / open source
A better prediction
is only the beginning.
Can a model explain animal choices—or is another process producing the same pattern?
AnimalTaskSim combines task environments, comparison agents and trial-level analysis to test explanations of decision-making. The work follows the evidence even when it weakens the original idea.
Explore the research repository ↗Latest documented comparison · 19 September 2026
Strong effects recovered.
Weak effects still unresolved.
Across 360 development-sized simulations, the analysis reliably recovered strong history effects but not weak ones. A separate numerical check of the drift model also missed its fixed tolerance. Fresh confirmation remains deferred.
The question changed
with the evidence.
The earlier work explored whether an adaptive controller could reproduce patterns in animal behaviour. Source reconciliation and controlled comparisons showed why a convincing-looking match was not enough.
The current question is narrower: does recent outcome history improve prediction beyond changes in underlying preference—and can the analysis distinguish those explanations when the generating process is known?
What the earlier
replication established.
A frozen internal replication found a small predictive gain from evidence-dependent history in 45 eligible mice: 0.00160 nats per trial, with a 95% subject-bootstrap interval of 0.00082–0.00244. Predictions improved for 32 of 45 animals.
That result was internally frozen, not externally preregistered, and the cohort was uneven across labs. It supports a bounded predictive effect; it does not establish a neural mechanism, scientific novelty or the adaptive controller’s necessity.
Research decisions
Keep simpler explanations
Compare history, drift and switching models. A more complex architecture has to earn its place through a controlled comparison.
Test the test
Use simulated data with known generating rules to check identification and power before spending a fresh cohort on confirmation.
The next gate
is methodological.
The next work is stronger recovery at the intended effect size, stable numerical results, and a fitted continuous-drift comparison. A new confirmation cohort needs a new frozen protocol once those checks are adequate.
Python task environments, schema-validated trial logs and shared evaluation tools support this process. They are research infrastructure, not evidence that a scientific hypothesis is true.
Read the methods and results ↗