read-only research replay · no trading controls · no live data

AlphaMorph // REBUS Experiment 001

What happens when you temporarily destabilize an artificial brain's established beliefs?

JavaScript is off: static summary

Experiment REBUS-001, status COMPLETED, window role HELD_OUT, data RECORDED.

Headline verdict: NO_MEANINGFUL_DIFFERENCE

One recorded stream and one candidate circuit (c0); results do not generalise beyond them.
c0 enters very rarely: internal dynamics and decision divergence are the primary outcomes; performance is descriptive and usually INSUFFICIENT_EVIDENCE.
Seeds only randomise the RANDOM comparator (5 seeds): the REBUS-vs-RANDOM comparison is coarse. The REBUS condition itself is deterministic.
Shadow-weights (counterfactual learning) are a research construct that never feeds back into the decision actually taken; their effect is shown only through 'shadow decisions'.
Counterfactual outcomes come from fixed-notional paper positions opened at abstentions (selection-biased); they are not trade records.
Decisions of different conditions are matched on (token, decision obs, apply obs, sensory flag); decisions that exist in one condition only are reported as unmatched, not imputed.
Tokens are the bootstrap cluster; decisions of one token are not independent. Observations of the same stream are not independent of the market regime.
Fixed hand-chosen registered maxima and intensity levels; no claim that any level is 'optimal'.

Loading recorded experiment…