An evidence-driven investigation of what biological neural systems can — and cannot — tell us about learning, prediction, memory and intelligence.
Not a single score. A chain of results, limits and changed questions.
Firing-rate and network summaries failed to improve experiment-level performance prediction under organoid-held-out evaluation. A rank association did not become calibrated predictive accuracy.
Simple burst/backbone features added short-horizon predictive information beyond population-rate lags in the tested sessions, under time-blocked, training-only feature derivation.
Files with different whole-file digests contained exactly equal targeted neural arrays in selected repeated pairs. File-level uniqueness was not biological independence.
When aligned condition labels were unavailable, the investigation moved from supervised decoding to label-free temporal prediction instead of manufacturing a variable the data did not contain.
The record is stronger because it keeps failed interpretations visible.
Aggregate summaries are not a reliable performance shortcut.The tested summaries did not beat simple baselines.
A missing label is not permission to invent one.Subject identity, timestamps and filenames are not experimental conditions.
File-level uniqueness is not biological independence.Different digests can still contain equal neural payloads.
Synthetic control is not organoid evidence.Computational trajectories do not establish intelligence or consciousness.
M2-E does not prove that more memory is better.The policy-class confound remains explicit.
Controlled computational experiments can be run with explicit tasks, seeds, metrics and provenance.
Simple temporal features can carry incremental short-horizon predictive information in bounded sessions.
Replication can expose a provenance problem instead of confirming the intended interpretation.
Contradictions and uncertainty can be preserved rather than smoothed away.
A missing label can force a more defensible, label-free research question.
A research-state model can preserve knowledge, questions, hypotheses and next moves.
The sequence matters. Each stage changed what the next stage was allowed to ask.
Public papers, tools, datasets and licenses were checked at the source. Claims were corrected before they became project assumptions.
Experiment-level performance prediction was tested on 32 experiments from 16 organoids. Simple baselines won.
The available neural objects did not carry a clean condition or performance label aligned to the windows.
The question moved to label-free next-window prediction, with explicit time blocks and no learned embeddings.
Burst/backbone features added a narrow predictive direction, survived most contiguous blocks and reproduced in a third tested session.
Equal neural payloads inside different files changed the interpretation from biological stability to source lineage and curation.
A separate synthetic laboratory tested closed-loop control, memory windows and perturbations while keeping its biological boundary explicit.
A synthetic control laboratory, published as a concept and a boundary — not as a biological substitute.
The public edition describes the protocol shape behind M1, M2, M2-D and M2-E: action, response, observation, transition memory, objective, metrics and the next action.
The private laboratory contains the operational machinery. This page contains the portable idea.
Public record. Private machinery.
Evidence before interpretation.