POWM / OCTAVE — Eight-observer predictive memory architecture Design review, 30 September 2026 Decision Build a causal probabilistic world model with eight communicating temporal slots, persistent memory and explicit uncertainty. Eight is the experimental hypothesis supplied by the project, not an established law of brains or reality. The system must be able to falsify the hypothesis. This package implements a small recurrent research baseline and correctness fixes; it does not implement the entire scaled design below or demonstrate frontier performance. What the current implementation establishes 53,200 learned parameters. Eight observer identities and relative circular anchors; learned shared phase routing; gated fast and slow recurrent states; trainable communication and global feedback; three shared-backbone prediction heads; prioritized training replay. No generating period or physical phase is passed to the student. Verified exact 32-bit forecasts on all eight contexts of the fixed 11110000 pattern after 400 updates. Numerical gradient error below 0.000006. Streaming chunks reproduce full-window outputs. Hidden sensor NaNs and caller-supplied differences cannot affect inference. Atomic checkpoints retain parameters, optimizer, internal RNG and replay; the next identical update is bitwise equal after restore. External teacher RNG and live streaming state remain caller responsibilities. Core limitation A fixed 128-bit threshold history can be identical for two different continuous oscillators whose futures disagree (a witness exists in the preceding project). A universal exact-future guarantee is therefore impossible for that observation contract. Use calibrated distributions over futures, additional observations and selective decisions. Finite exam success is a narrower claim than universal mastery. Proposed scaled architecture — not yet implemented 1. Causal sensory encoder. Encode value, visibility, elapsed time and observed differences. Missing values have no content embedding. Apply all normalization using training data only. Sensor outages are represented explicitly. At streaming boundaries carry timestamps, previous visibility and recurrent state. Neither future targets nor teacher latent variables enter this interface. 2. Eight temporal observers. Each observer holds a fast local state, a slow state and a phase belief. Circular relative positions provide a soft relational bias, never a supplied true phase. A learned oscillator state represents uncertain speed and direction; allow a nonperiodic residual so the model can escape the cycle assumption. Route observations to all slots with a small minimum attention floor, plus a normalized learned allocation. The floor and its value require ablation: it prevents starvation but could dilute specialization. 3. Communication. Use gated sparse attention among neighboring and selected remote observers, followed by a global workspace bottleneck. Shared messages carry source identity, age and reliability. Update slots simultaneously from the previous tick to avoid observer ordering artifacts. Two rounds per tick is an initial setting, not a requirement. Track overlap, message utilization and whether removing one observer changes decisions. 4. Hierarchical belief dynamics. Infer q(z_t | observations through t); predict p(z_t | z_(t-1), action, dt). Train an observation decoder and latent transition with reconstruction, predictive and regularized posterior/prior losses. Use a mixture distribution or particles when multiple hidden worlds match the same observations. A single deterministic phase estimate would erase legitimate ambiguity. Roll out transitions step by step; evaluate drift separately at horizons 1, 8, 32, 128 and 512. 5. Episodic memory. Separate online inference memory from labeled training replay. Store past observation-derived keys, causal summaries, timestamps and model version. Retrieve only entries available before the prediction cutoff. Never write target futures into inference memory. Gate retrieved values against present evidence; cap capacity; test stale and contradictory memories. For training replay, prioritize prediction error while retaining uniform sampling to avoid feedback loops. 6. Questions and actions. Action candidates are observe longer, expose direction sensor, request a higher-resolution sample, or predict. Score estimated entropy reduction minus observation cost using simulated posterior updates; validate these estimates against realized information gain. Begin with an auditable policy before replacing it with a learned policy. No claim of active sensing is warranted from the current heuristic. 7. Prediction and calibration. Separate sensor forecasts, hidden-state estimates, confidence and action scores. Train with proper scoring rules. Calibrate on a reserved validation split using independent-backbone models or suitable posterior approximations; shared heads alone do not establish epistemic uncertainty. Report log loss, Brier score, interval coverage, risk versus coverage, and exact-sequence accuracy. Permit abstention; disclose its coverage alongside its accuracy. Initial scaled configuration — experimental budget, not an optimum Eight slots, 256 fast and 256 slow channels each; 512-channel workspace; four communication heads; 128-dimensional episodic keys; 4,096 causal memory entries; 32 stochastic latent channels per slot; three independently trained replicas. Train with 256 observed ticks and truncated backpropagation chunks of 64; detach state between chunks while preserving forward state. Benchmark full-window gradients on shorter cases to measure truncation bias. Use mixed precision only after float32 parity checks. Measure the actual parameter count, memory and throughput from a concrete implementation before publishing them. Do not infer scaling laws from this NumPy prototype. Training contract Level 0: finite fixed pulses, exhaustively enumerate contexts. Level 1: randomized continuous pulses, disjoint frequency bands and untouched seeds; evaluate ambiguous histories probabilistically. Level 2: multiple sensor channels, mask random channels, infer hidden quantities and test whether added supervision helps Level 1 when sensors are hidden again. Later: irregular sample times, phase resets, frequency changes, noise, mixtures and previously unseen motifs. Change one family at a time before combining them. Teacher may adapt training sampling from training errors and a separate development set. Freeze final tests, never adapt to their per-example errors, and never select checkpoints on final test scores. Record dataset hashes, seeds, update budgets, sample counts and teacher interventions. Keep learned targets separate from teacher-only generator metadata. Objectives Use masked likelihood for visible reconstruction and future observations, latent prediction loss, posterior/prior consistency, and optional auxiliary hidden-sensor supervision. Stop gradients through target embeddings in predictive representation learning. Tune loss weights on the development set. Avoid forcing uniform slot activity or orthogonal states unless ablations show benefit; diversity penalties can create meaningless differences. Check that all claimed mechanisms receive gradients and cause measurable behavioral changes. Required comparisons before claims Compare 1, 4, 8 and 16 observers with approximately matched parameter count AND separately matched compute. Include a plain recurrent model, the earlier recurrence/copy model, a simple transformer and a state-space baseline. Run no communication, no slow memory, no workspace feedback, no routing and no episodic retrieval ablations. Use at least five paired seeds for the initial study. Report each seed, mean and uncertainty; bootstrap at independent-world level, not individual correlated bits. Give learning curves against both observed samples and elapsed compute. Eight-phase claim gate Measure observer routing versus true phase only after training, for analysis. Use circular alignment metrics with a permutation-invariant matching and test stability across seeds, changing frequencies and phase jumps. Teacher phase must never be an input. Compare learned routing with random anchors and learned unconstrained slot identities. If eight observers do not outperform matched alternatives, publish that result and retain eight only as an interface choice. Level 2 transfer gate Clone a common Level 1 checkpoint including optimizer. Compare equal-budget Level 1 continuation against Level 2 training with fixed Level 1 replay. Hide Level 2 sensors at exam time. Report current-state inference separately from future forecasting. Earlier project experiments found negative future transfer; the richer architecture must reverse that result in paired tests before claiming that Level 2 helps Level 1. Operational acceptance Finite losses and gradients under 128+ ticks; bounded activations over long streams; input schema rejection tests; hidden-value invariance; no target leakage; deterministic resume with external RNG explicitly included by the trainer; replay cutoff audit; interrupted-save recovery; inference state reset rules. Evaluate latency and peak memory as observer count and history grow. Numerical correctness is necessary but not evidence of generalization. Prior art and scientific boundaries Recurrent processing, multiple timescales and complementary memory motivate design choices; they do not make this a biological brain model. Related ideas precede this project: learned external memory (Graves et al., Nature 2016), intrinsic cortical timescales (Murray et al., Nature Neuroscience 2014), predictive world models (Hafner et al., Nature 2025), and structured state-space sequence models (Dao & Gu, ICML 2024). The testable contribution is the combination and evaluation of eight relative-phase observers, not the invention of recurrence, memory, attention or world models. Primary references: https://www.nature.com/articles/nature20101 https://www.nature.com/articles/nn.3862 https://www.nature.com/articles/s41586-025-08744-2 https://proceedings.mlr.press/v235/dao24a.html Reproduction Requires Python 3 and NumPy. From this folder: OPENBLAS_NUM_THREADS=1 python run.py OPENBLAS_NUM_THREADS=1 python verify.py OPENBLAS_NUM_THREADS=1 python review_checks.py run.py performs the fixed-pattern smoke training only. review_checks.py checks hidden NaNs, delta invariance, checkpoint continuation and all eight pulse contexts. results.json, verification.json and review_checks.json are measured outputs. Architecture.txt is the earlier baseline description; this document supersedes its input and checkpoint details. Public claim that the evidence supports today An executable eight-observer recurrent prototype learns an exhaustively checked basic periodic pattern, with verified gradients, causal sensor masking, streaming continuity and reproducible optimizer continuation. Broader mastery, observer superiority, calibrated uncertainty and frontier capability remain hypotheses requiring the experiments above.