Action state
EXTEND · HOLD · RETRACT
What the governed process is permitted to do next. HOLD is a deliberate unresolved posture.
Machine implementation
Cognitive Basin / BasinLab
What if memory changes the shape of the next available action rather than merely adding another line to a log?
Boundary: this is a governed architecture for persistent state and replay. It does not establish consciousness.
Migration notice: This page is now framed as an applied Consequential Formation architecture: retained formation, contradiction scars, recovery routes, and history-sensitive reasoning accessibility.
- Question
- Can a capable machine acquire an individual formed history that causally changes how it reasons next?
- Current approach
- Bounded Basin structures plus the MFM v0.1 candidate: an auditable Formation Ledger and indexed executable susceptibilities that recruit a sparse coalition.
- Status
- Cognitive Basin has bounded software evidence; MFM and the Formative Field runtime remain specification / immediate build-test targets, not an achieved intelligence.
- Evidence
- Local fixtures and public claim records are linked through Documents; CF v0.4 RC1 sharpens the causal and matched-baseline tests.
- Open problem
- Demonstrate reversal learning, recovery formation, single-susceptibility transfer, and ablation beyond transcript, RAG, RETE, procedural-memory, and temporal-KG baselines.
A topological view of machine state
Repeated routes can deepen. Contradictions can leave scars. Recovery can create a new path without pretending the contradiction never happened. A HOLD region keeps unresolved work visible until evidence changes.
Two namespaces, kept separate
Evidence posture
SUPPORTED · UNRESOLVED · CONTRADICTED
How the current record bears on a proposition. These labels do not name actions.
Natural Math uses EXTEND / SENSE / RESTRICT for a different purpose. SENSE is a local operation; HOLD is not another name for it.
What BasinLab makes auditable
Persistent state
Changes survive beyond a single request.
Governance
Rules constrain which local action is available.
Contradiction and recovery
Corrections supersede without erasing prior evidence.
Replay
A reviewer can inspect how the present basin was shaped.
Current immediate engineering target
Minimum Formative Machine v0.1
Can a capable machine acquire an individual formed history that causally changes how it reasons next?
CAPABILITY != HOST POLICY != ACQUIRED FORMATION. The host may already provide language, coding, mathematics, simulation, tools, and spontaneous habits such as checking or decomposition. The experiment must profile that bare-host policy before attributing behavior to the developing system's own formation.
Theta_ledger = audit, provenance, contradiction, transferTheta_exec = indexed susceptibilities, gates, masks, tolerancespresent condition + Theta_exec -> sparse recruited coalitionG(typed witnessed consequence) = governed formation updateThe old assumption that a passive graph should be searched and compiled on every query is demoted. A small COMPOSE? office may turn a recruited coalition into coherent host input, but it must be independently ablated and deleted if raw coalition delivery performs equally well.
Load-bearing test: change, transplant, or ablate one executable susceptibility and measure the predicted later change in recruitment, reasoning, or admissibility under matched host, prompt, context, and random-seed controls. If nothing changes, the Basin is decoration or retrieval rather than consequential formation.
SPECIFICATION / BUILD-TEST TARGET MFM and the Formative Field are not achieved intelligence and are not claimed to be alive, conscious, or autonomous.
Candidate runtime
Only what matches the present condition should wake.
Hard admissibility first
BLOCKED and ADMITTED routes should not be assigned decorative probabilities. Weighting begins only when genuine competition remains.
Sparse recruitment
Indexed matching should make wake cost depend on the condition and active coalition, not on rescoring the full retained history.
Novelty remains visible
ABSENT or thin recruitment may trigger a bounded global probe and seed new susceptibility. UNKNOWN preserves insufficient evidence.
Prior art owns much of the skeleton. RAG, RETE and production systems, blackboard architectures, spreading activation, content-addressable memory, pub-sub, and event buses remain required comparison classes.
Preferred first causal sequence
Negative transfer
Clone the same host and formation, induce a useful route in one history, then reverse the environment and predict precisely how that retained formation should initially hurt.
Recovery and transplant
Allow recovery formation, transplant one recovery susceptibility into a naive twin without replaying the transcript, and demand its narrow predicted advantage.
Ablation and baselines
Ablate the transplanted susceptibility and demand disappearance of the effect. Compare against context, RAG, RETE, procedural memory, temporal knowledge graphs, and the same bare host at equal budget.
Candidate formation writes remain typed: positive association, scar, open question, contradiction, recovery route, pruning, and bounded experimental meta-write. This is a testable engineering hypothesis, not a universal learning law.
Developmental-search advantage
Machine-speed iteration improves the experiment, not the evidence status.
Digital systems can be cloned exactly, given parallel histories, checkpointed, restored, perturbed one variable at a time, ablated, and transplanted with full provenance. The candidate loop is CLONE → PERTURB → EXPERIENCE → MEASURE → ABLATE → SELECT → TRANSFER.
Allowed claim: machine-speed parallel causal iteration can compress search when outcome evaluation and credit assignment remain discriminating. Generating histories is cheap; assigning causal credit is not. No fixed evolutionary-span or wall-clock prediction is earned.
The governing question is: Which developmental law makes the same capable machine become better?
Project lineage
Persistent Observer remains a current project record comparing learnable novelty, weighting, Basin updates, and host-bounded action. It is not the framework's front door and does not make MFM an implemented intelligence.
Build surface
The public platform repository contains the implementation surface. Architectural language here remains subordinate to runnable behavior and recorded tests.