organoid-oi v3.2: A Biologically Grounded Closed-Loop Framework for Organoid Intelligence Simulation — A Self-Audit, and a Continued Search for a Learnable Task

UPDATE NOTICE (v3.2) — This version extends v2.0; it does not revise the finding there, which stands. Two things were done before any hardware commitment: the decoder-independent attribution machinery was rebuilt into a live, two-arm monitor and calibrated against sessions with a known ground truth; and the three-category task used throughout the original audit was found to be solvable by an untrained organoid at 100% accuracy — a second, independent instance of the same general hazard the audit named, this time in the task rather than in the control. A replacement task was designed so an untrained organoid performs at chance, and a pre-registered search was run over two candidate mechanisms. No configuration has yet produced a validated positive result. This is reported as the current state, not a resolved question. We present organoid-oi v3.2, an open-source Python framework for simulating closed-loop categorical learning in a synthetic brain organoid, together with a full self-audit conducted before any hardware deployment and a subsequent, still-open search for a task on which that audit's negative finding can be overturned. The system implements Three-Factor Spike-Timing-Dependent Plasticity (STDP) with biologically realistic reward and penalty delivery: reward is a 10 Hz sub-threshold theta depolarization current producing natural long-term potentiation (LTP); penalty is a 200 Hz noise current producing long-term depression (LTD). A local Hebbian readout layer with weight decay decodes population activity. Every mechanism has a named biological counterpart — no backpropagation, no external weight forcing, no mathematical shortcuts. The audit reverses the central claim of the previous version. Under paired control (identical seed, identical stimuli, STDP enabled versus disabled), the difference in categorical accuracy is exactly zero across every condition tested: two encoding schemes, four connectivity densities, five excitability regimes, and learning rates spanning zero to five hundred times the default. A decoder-independent measure of category separability shows STDP's net contribution to be approximately zero and, on average, slightly negative; separability is present at initialization and decreases during training. The framework's above-chance performance (~68%) is produced entirely by the Hebbian readout layer reading a fixed, pre-existing separability the organoid possesses before any learning occurs. Beyond that audit, this version reports a continued investigation. The attribution machinery was rebuilt into a live two-arm monitor and calibrated against sessions with a known ground truth before being trusted against any new question. That monitor then exposed a second instance of the same general hazard: the three-category task audited above turns out to be solvable by a completely untrained organoid at 100% accuracy, purely from the random connectivity present at initialization. A replacement task was built from Morse-coded anagrams of the same three symbols, on which an untrained organoid performs near chance (33.3%). A pre-registered search over recurrent connectivity and temporal repetition found one configuration whose separability effect replicates across seeds but whose accuracy effect falls short of the threshold fixed in advance, and ruled out the readout layer as the explanation for that shortfall using an offline, non-adaptive classifier. No configuration has yet produced a validated positive result. The search is reported in full, including the candidates that failed to replicate, because a search that only reports its successes is not a search that can be checked. We report this as a negative result. The framework is sound as an engineering artifact and its mechanisms are correctly implemented; what it does not yet do is demonstrate STDP-driven categorical learning. We argue that this failure mode — apparent closed-loop learning that is in fact entirely attributable to the readout layer, whether in the control or in the task itself — is a general hazard in organoid intelligence work, and that the attribution controls used to exclude it require more scrutiny than they have generally received. This work is funded by an Emergent Ventures grant (Mercatus Center, George Mason University). Hardware access is being coordinated with FinalSpark for a live closed-loop deployment on its Neuroplatform; the calibration, task-design, and search work reported here is the readiness gate for that deployment.