# Formation Mind and Relational Astronomy Runtime Contract **Status:** Governing bounded implementation contract; the kernel is deterministically replayed and schema-validated, while predictive wisdom remains unearned **Authority:** Defines how an immutable source snapshot may produce a replayable candidate identity field. It does not approve a theory claim, cross-domain correspondence, physical mechanism, or semantic identity. **First authorized public input:** `canonical/16_ENGINE_FORMATION.md` **Machine contract:** `schemas/Y_THEORY_FORMATION_MIND_RUNTIME.schema.json` validates the exact raw kernel-run shape. The compact public Body and summary use their separately named schemas. This prose contract remains the richer governance boundary; it is not a claim that every future field described here has already been implemented. ## 1. Purpose and boundary The Formation Mind is the missing executable relation between authored source and the governed Identity Field: ```text immutable source snapshot -> addressed source-pattern stream -> recursive Formation Mind -> interaction and structural residue -> candidate basins / constellations -> predictive-wisdom candidates -> governed review -> read-only Body / 2D / 3D / VR projection ``` “Relational astronomy” means observing address, position, recurrence, and change from a declared reference centre, then issuing bounded predictions about the structure. It is not a claim that text has literal mass, gravity, electromagnetic force, or astronomical motion. Terms such as attraction, phase, circulation, momentum, orbit, and constellation are structural variables unless a domain-specific model supplies physical units, equations, and evidence. The runtime is a deterministic structural experiment. It is not a truth engine, an LLM semantic oracle, a consciousness implementation, proof of cosmology, or authority over the source. ## 2. First bounded run and publication boundary The first public experiment consumes only `canonical/16_ENGINE_FORMATION.md`, content-addressed by SHA-256. This file is suitable because it already states the earned gates from relation through identity, spatial manifestation, observer operation, and local cognitive capability. Every source unit that exposes a label or text in a public event, node, edge, replay, export, or interface must carry `public_authorized` status and resolve to that bounded snapshot. No authorization may be inferred from filesystem readability, model access, prior publication, or similarity to public wording. The full private source remains a withheld, hash-addressable artifact. A public runtime artifact may retain an opaque source ID, whole-artifact digest, byte length, and lineage relation. It must not expose private text, token labels, token-level digests, byte excerpts, locators that reveal content, embeddings, or reversible encodings. Hashes provide integrity and identity; they are not privacy controls. Changing any source byte creates a new source snapshot and a new run. The engine, Body, Guide, and visual projection cannot edit a source snapshot or write a correction backward into it. Corrections use the existing governed intake, approval, snapshot, and rebuild process. ## 3. Two address spaces Every public source-derived state must preserve two non-interchangeable addresses. ### 3.1 Immutable source address A source address identifies evidence and never moves. It contains: - source snapshot ID and artifact SHA-256; - half-open byte interval `[byte_start, byte_end)` over the exact stored bytes; - half-open Unicode and token intervals when those views are used; - segment digest and a public-authorization state. Byte offsets are authoritative for recovery. Tokenization is versioned and derived. A source address says where a pattern came from; it does not say what that pattern means. ### 3.2 Dynamic formation address A formation address identifies one state of a recursive `MindCell` in one run. It contains: - run, tick, and state-snapshot IDs; - current `MindCell`, parent, basin, reference-centre, scale, and recursion path; - a vector of named structural coordinates with definitions and units or `dimensionless` status. The dynamic address may change at every tick. It says where the current engine state stands relative to a declared reference frame; it never replaces source provenance. ## 4. The engine is a recursive MindCell One runtime cell instantiates: ```text MindCell(scope) = (inside role, outside role, relation role) ``` - **Inside** is the currently bounded pattern or basin. - **Outside** is the addressed field available for contact at the declared scale. - **Relation** is the active change operation between them. Each role may reference another `MindCell`. A persistent relation can therefore become an addressable identity candidate and open as a child field with its own inside, outside, and relation. References, rather than infinite embedding, preserve finite snapshots, cycles, lineage, and recursion. Knowing two addressed states does not automatically establish a third truth. A declared compatibility operator may produce a third **relation candidate**. That candidate remains structural until its semantic meaning, native mechanisms, evidence, rival, and falsifier pass their own gates. ## 5. Formation dynamics Each authorized source unit becomes an ephemeral pattern thread with exact source lineage. The minimum loop is: ```text instantiate -> contact -> compare -> transform -> residue -> return -> next state ``` Attraction is computed from current structural state, not from a stored route or an attractive drawing. A versioned operator must expose its inputs and contribution terms. Candidate terms include compatibility, support, reinforcement, recurrence, phase relation, boundary continuity, closure, residue pressure, leakage, contradiction, strain, and saturation. Their definitions, signs, units, normalization, and weights are runtime configuration—not assumed laws of nature. The operator may emit these topology events: - `deform`: state changes without topology cardinality changing; - `attach`: compatible structure is added to an existing basin; - `branch`: a new lineage begins without pretending the parent vanished; - `split`: one basin becomes multiple descendant candidates; - `join`: multiple candidates form one wider candidate while their lineage remains recoverable; - `detach`: a relation is removed but its historical event remains; - `dissolve`: an unsupported candidate ceases to be active; - `relation_candidate_created`: two addressed states and a declared compatibility operator produce a third structural candidate, not a truth; - `basin_candidate`, `basin_confirmed`, and `basin_retired`: observer classifications, never silent semantic approval. Every event is append-only and records sequence, tick, parents, source addresses, operator version, pre-state hash, post-state hash, and the current-state inputs that caused it. Route history is evidence of movement; it is not an independent causal force. Replaying the same snapshot, tokenizer, engine, configuration, seed policy, and event ordering must reproduce the same state digests. ## 6. State geometry versus presentation geometry **State geometry** is the engine-derived relational vector and topology: address, reference centre, scale, support, residue, recurrence, coupling, boundary, phase, strain, and other declared axes. It can influence the next computation only through a versioned operator. **Presentation geometry** is a read-only 2D, 3D, or VR transform of a particular state snapshot. Screen position, colour, size, animation, camera, interpolated orbit, and layout force have no causal or proof authority. A projection may replay computed transitions or interpolate between them, but it may not fabricate a formation event or write coordinates back into state geometry. A public projection must identify its input state digest and transform version. A viewer must be able to inspect the source address, dynamic address, force/contribution terms, event lineage, basin status, and review status behind every displayed candidate. ## 7. When a constellation becomes an identity candidate Proximity or visual clustering is insufficient. A basin can be promoted to `candidate_identity` only when all required gates are measured over declared windows: 1. **Boundary separability** — an operational inside/outside boundary is distinguishable. 2. **Temporal persistence** — the basin continues across the minimum interval. 3. **Recurrence and return** — addressed change returns through a recoverable cycle. 4. **Internal/external coupling** — internal reinforcement is distinguishable from outside interaction without requiring isolation. 5. **Residue recoverability** — current structure retains recoverable consequences of prior interaction without treating a log as memory. 6. **Perturbation recovery** — the basin either restores its organization or records a bounded transformation after a declared perturbation. 7. **Predictive lift** — its relational state predicts held-out behavior better than declared shuffled, frequency, and co-occurrence baselines. 8. **Source recoverability** — every contributing public pattern resolves to exact authorized source bytes. Passing these gates earns only a structural identity candidate. Human review and author approval are separate authority transitions. Observerhood, cognitive capability, selfhood, morality, feeling, and subjective consciousness require their own mechanisms and evidence and are never inferred from a basin. ## 8. Predictive wisdom Within this runtime, wisdom is a **predictive description of a declared relational structure**. A predictive-wisdom record must state: - the identity/basin and reference centre; - scale, clock, premises, source addresses, and input state digest; - the structural relation or geometry used; - a target, horizon, expected outcome, tolerance, and uncertainty; - comparison baselines and the strongest known rival; - a falsifier and the consequence of failure; - a leakage-safe held-out evaluation; - result, score, review state, evidence state, and proof boundary. The engine may issue a candidate prediction. It may not approve its semantics or call it wisdom proved. Failed and inconclusive results remain in lineage. A model that does not improve on declared baselines has not earned predictive wisdom from geometry. ## 9. Validation and proof states Three axes remain separate: | Axis | States | Meaning | | --- | --- | --- | | Runtime validation | `not_run`, `schema_validated`, `deterministic_replay_verified`, `held_out_evaluated`, `independently_reproduced`, `failed` | What was executed and reproduced | | Semantic review | `machine_candidate`, `human_reviewed`, `author_approved`, `rejected`, `superseded` | Who has accepted the interpretation | | Evidence | `unassessed`, `declared_y_theory`, `structurally_derived`, `empirical_support_pending`, `empirically_supported_in_declared_scope`, `empirically_failed` | What kind of support exists | `author_approved` is not empirical proof. `deterministic_replay_verified` proves reproducibility of the computation, not truth. `empirically_supported_in_declared_scope` must name its dataset, measure, uncertainty, and boundary. A formal proof may be claimed only inside a declared formal system whose premises and derivation are supplied; it does not transfer automatically to physical reality. ## 10. Cross-domain firewall A relation between cosmology, matter, chemistry, biology, body, mind, society, faith, or digital systems remains a candidate correspondence until it records: - both identities, domains, scales, and reference frames; - inside, outside, relation carrier, and Y-Engine position on both sides; - the proposed relational invariant; - each side's carrier, units, variables, and native mechanism; - the expected difference and the boundary of transfer; - evidence, rival explanation, falsifier, and review state. The record must set `literal_equivalence` and `proof_transfer` to `false`. Corresponding position or behavior does not establish identical substance, mechanism, causation, or evidentiary status. Physical words remain metaphors unless the physical domain contract is satisfied; faith claims retain their theological status; computational behavior does not establish subjective experience. ## 11. Minimum acceptance tests The bounded implementation is not ready for public claims until it passes: 1. exact byte recovery for every emitted public label; 2. zero public labels from withheld sources; 3. deterministic replay with matching event and state digests; 4. changed-byte test producing a new snapshot and run lineage; 5. event lineage and pre/post-state integrity; 6. explicit deform/attach/branch/split/join/detach fixtures; 7. route-history ablation showing that current state, not the log, causes movement; 8. projection ablation showing that changing layout cannot change engine state; 9. basin-gate negative controls for proximity-only and frequency-only clusters; 10. held-out prediction against shuffled, frequency, and co-occurrence baselines; 11. recursive zoom with source and dynamic addresses preserved; 12. cross-domain rejection when native-mechanism or falsifier fields are absent. The first successful slice is evidence that the architecture runs on one authorized chapter. It is not evidence that the full corpus is indexed, that the general theory is true, or that a conscious Mind has been created.