BP-0050 · Eencellig organisme / Proto-AI

This first standalone dossier preserves a historical proposal for a computational gridworld agent. The agent would regulate internal resources, populate explicit Bio-Frames from a small learned visual module, select actions through an active-inference model, edit symbolic knowledge after feedback, and coordinate through threshold messages. These are proposed components and targets. This dossier supplies documentary evidence, without implementing or running that agent.

The planned destination is the WR-owned private repository Flynn/bp-0050. Repository, commit and platform IDs are unassigned. Rights are unknown, known source and review answers are exposed, and publication has not been authorized. Historical scientific assessment, this dossier's independent review and external root adoption have separate statuses.

Read the materials

Material	Role
Full original	All 50,890 original bytes, retaining the literal filename's spelling.
Full accepted assessment	All 16 BP0050-C claims, rationales, statuses, evidence scopes and 22 literal excerpts.
Historical catalog	Unchanged accepted catalog record with its original path conventions.
Historical review guide	New guide to the earlier two-source review and its exact two-field metadata closure.
Selected evidence	Exact historical selections, their upstream identities, line ranges and path bases.
Source provenance	Copy identities, reading coverage and reference-only evidence roles.
Machine record	Claim, comparison, analytical and question routes with explicit local bases.
Rights	Unknown licensing and distinct original, assessment, catalog and derivative-guide roles.
Placement	Eleven unique prospective repository members and their roles.
Manifest	Ten non-self member identities; the external author handoff binds the manifest.

Follow the computational model

Start with C01–C04 and C10–C11 in the full assessment. Homeostasis needs state measurements, viability limits, action costs and a fallback/recovery contract. Preserving the agent's resources does not specify acceptable external effects. Bio-Frames and rule firings provide inspectable records, but a valid deduction can use an incorrect CNN classification or an observational association mistaken for a cause. The example at original line 25 omits the operands in its threshold conditions; this dossier preserves that omission.

C08 separates variational posterior fitting from policy selection using expected free energy, preferences and information value. C09 records the missing observation likelihood, hidden states, transitions, variational family, horizon and compatible updates. Physical resource energy, prediction error, variational free energy and RL reward cannot be interchanged by their names. These are distinctions in the accepted assessment, not a newly supplied integration algorithm.

C12–C13 distinguish update targets. In the historically checked RLSF paper, symbolic certificates supply rewards for PPO updates of a neural language model. Proto-AI instead proposes changing rule priorities, generating rules and creating frame associations. That proposal still needs a search space, update rule, delayed-reward credit assignment, conflict handling, revision and forgetting. Recording one labeled event is different from generalizing its effect to unseen or noisy situations.

For perception, count the labeled visual examples, attribute and relation supervision, object tracking and uncertainty handling needed to populate the frames. The source's initial lookup table grants different state access from a trained CNN. Its later proposal to gather more observations acknowledges errors; recovery is not guaranteed when the model itself is wrong. A symbolic trace covers only the recorded stages, not independent truth of every perception or causal premise.

C07 preserves both the quorum proposal and the later admission of deadlocks and resource hoarding. Signal propagation, decay, delays, coupling, thresholds and interactions remain unspecified. A second stress signal adds feedback but supplies no global stability or task-completion proof.

The source proposes survival beyond 1,000 steps, 100% auditability and 95% recognition, plus memory and GPU estimates. None is an observation of this proposed system. Original lines 175–176 leave the transparency target and DEQ expression blank. Trace coverage, correctness, resilience under a defined shock distribution, held-out generalization and measured compute are distinct evaluations. Hardware preferences and the numerical budget remain historical source assertions.

Compare under the same information budget

The following four routes are from the accepted assessment's /competent_simpler_baselines array. The blueprint does not supply their results.

Assessment route	Comparator and matching conditions	What its proposed comparison would measure
/competent_simpler_baselines/0	Finite-state or model-predictive control with the same state estimates, resource dynamics and action budget.	Survival, resource cost, hazards, recovery and decision latency under matched shocks.
/competent_simpler_baselines/1	A tabular Bayesian estimator or contextual bandit with confidence and revision, plus an explicit rule table.	Held-out decisions, noise, reversals, cumulative harm and all labeled interactions.
/competent_simpler_baselines/2	The same trained CNN feeding a simpler controller; report an oracle-state controller separately.	Attribute accuracy and calibration, downstream control failures, pretraining and labeled frames.
/competent_simpler_baselines/3	A barrier/handshake protocol or state-machine coordinator with equivalent communication access.	Completion, deadlocks, messages, resources, delays and agent failure.

One optional reading question is preserved exactly from /research_questions/0: Which explicit probabilistic model makes homeostatic reward and expected-free-energy policies compatible? It remains unexecuted. Conditional reasoning about survival, traces, updates and metrics belongs to the accepted assessment; no new empirical or analytical result is claimed here.

Keep evidence versions and scopes distinct

The historical dynamics review dated 20 October 2026 covers Proto-AI and BP-0192: 32 claims in total, with no actionable findings. Proto-AI contributes 16 claims. The six-source lane and twelve-source parent have wider scopes. The 29,121-byte initial Proto-AI assessment differs from the accepted 29,218-byte file only in review_status and disposition. Reversing those exact values reproduces the initial bytes; technical content is unchanged. Historical independent acceptance does not stand in for current dossier review or root adoption.

The previously read primary records are selective support for component claims. Chemotaxis feedback and repeated Physarum exposure are bounded historical observations, not validation of one-trial symbolic learning. The Physarum transfer paper was assessor supplied; it is different from the Physarum article in the blueprint bibliography. The original assessor read selected mirrored PDF content. The earlier independent reviewer used retained text after a live PDF request failed, checked the PDF identity, and did not independently re-extract or replicate statistics. Active-inference equation images and most bibliography targets remain unaudited. The Deep AIF entry is abstract-only and was not independently reopened because no promoted claim depends on it.

The casefolded value queries bp-0050, bp0050-c, eencellig, proto-ai and proto‑ai return no hits in the five pinned registers. No standalone blueprint PDF is identified within that bounded query. The external Physarum PDF/text has its own historical evidence role. Primary PDFs, text extractions, other-source records and whole shared registers remain reference-only; none is bundled transitively. No new primary reading, PDF scan, biological experiment, controller execution, neural training or hardware benchmark was performed.