MD-OS

Unity Tensor Field Model

Epistemic status: frozen_principle_open_mathematical_hypothesis

Scientific question

Can the unity of a cognitive system be represented as one global informational object whose local components vary across cognitive frames while verified relations remain invariant?

Cognitive Integration Principle

The author-established principle is:

A generally intelligent system is not merely a collection of talented local
solvers. Differentiated cognitive parts must participate in one persistent
causal informational whole that can transform both frame and solver while
preserving the relations required for identity, coherence, verification, and
action.

The project uses cum scire, knowing together, as an explanatory intuition for consciousness as integration of differentiated parts. This etymological lineage motivates the question; it does not validate the answer.

Scientific antecedent: Integrated Information Theory

Tononi’s Integrated Information Theory (IIT) starts from two relevant phenomenological properties: conscious experience is differentiated and each experience is integrated as one. IIT develops this into an intrinsic causal account and an irreducibility quantity commonly denoted Phi.

MD-OS adopts IIT as an explicit scientific antecedent for the integration principle. It does not claim that IIT already defines the Unity Tensor Field, that a tensor is Tononi’s formal object, or that current Cortex measures Phi. The MD-OS problem is operational and cross-domain: relate changing representations, solvers, actions, memory, and evidence across frames without losing the causal unity of the system.

Primary sources:

Natural language boundary

Natural language is the command and hypothesis surface. It is expressive but cannot itself guarantee covariance, composition, or invariant preservation. A cross-frame statement becomes a mathematical commitment only through:

natural-language hypothesis
-> explicit premises, competitors, and falsifiers
-> declared frame family and typed local representations
-> candidate Unity Tensor and transition operators
-> sealed predictions before target observation
-> independent world observations and evidence hashes
-> composition, invariant, and information-loss checks
-> matched baselines, severing, contamination, and replication
-> bounded support, rejection, or unresolved status

A tensor-shaped array without these contracts has no privileged epistemic status.

Three distinct operational layers

The previous text conflated two objects. They are now explicitly separated.

The per-turn matrix

G_turn in R^(8 x 4)

is the Turn Governance Tensor. Its channels record whether bounded references to self, observation, goal, memory, frame, transformation, action, and evidence are present, authority-declared, and verifier-backed. The second basis is only a permutation of those four bookkeeping features. Exact roundtrip, composition, Frobenius-norm, and component-multiset checks therefore verify the encoding and its hash boundary. They do not inspect the meaning of an intent, establish that a hypothesis corresponds to the world, or implement the Unity Tensor Field. The historical JSON field and filename operational_unity_tensor remain temporarily for compatibility, while the artifact declares artifact_role = turn_governance_telemetry.

The Causal Unity Controller is the second object. Before an action decision, it binds nine hash-addressed channels—identity, world observation, intent, goal, memory, frame, prediction contract, action policy, and evidence— to six operational features: presence, activation, declared authority, verifier backing, causal necessity, and carry-forward. This produces the rank-two state

U_ctrl(k) in R^(9 x 6).

Unlike governance telemetry, this state is consumed by the action gate. An action authorization is valid only if the complete predecision state verifies, the action consumes its exact state hash, the frame and decision basis match, and policy plus authority checks pass. Every side-effecting action must have a matching prior authorization. Closure binds output, action, and evidence manifests into a transition hash, and the next turn includes that previous transition hash. Missing, tampered, mismatched, or bypassed state therefore inhibits authorization or leaves the transition incomplete.

The dependency probe runs the same candidate action twice: once with the intact state and once after severing a required component. It passes only when the intact path authorizes and the severed path is inhibited. This establishes causal dependence of the bounded APFC controller on the represented state. It does not prove that every relation was used inside the host model or that the state corresponds to the external world. The completed transition, rather than this predecision probe alone, determines C(k).

Primary executable contracts for this layer are:

The epistemic object is the third and different layer. A candidate Unity Tensor mathcal U_H is a sealed, integrated hypothesis whose projections into heterogeneous frames generate discriminating predictions. It is admissible only when those predictions were fixed before target observations and an independent world verifier binds each observation to current evidence:

candidate integrated hypothesis
-> sealed frame projections and predictions
-> independent world observations
-> cross-frame transformation and invariant checks
-> simpler-baseline, sham, severing, and contamination controls
-> independent replication
-> bounded support or rejection

The implemented epistemic verifier keeps four verdicts separate: hypothesis–world correspondence, cross-frame unity, causal integration, and independent replication. Internal consistency cannot pass the first verdict. A failed prediction, stale evidence file, post-hoc candidate, surviving simpler baseline, broken transformation loop, or missing replication rejects or leaves the exact edge unverified.

Primary executable contracts are:

The ordinary reflection path also requires a hash-bound independent epistemic readback receipt before it can create a verified cognitive anchor. These tests validate the fail-closed mechanism with controlled fixtures; they are not new empirical evidence for a global Unity Tensor Field, consciousness, or AGI.

Relation to consciousness

The candidate Unity Tensor represents differentiated contents as one identity-indexed structure. The Causal Unity Controller tests whether that structure actually constrains authorization and whether its result returns through continuity. Neither a tensor-shaped object nor a verifier alone is consciousness.

For episode k, MD-OS calls the complete causally closed event consciousness when the predicate C(k) passes: persistent identity, differentiated integration, same-I availability, joint constraint of response or action, and causal carry-forward. The verdict is bounded to the episode and is not automatically true for every model answer or APFC turn.

World readback decides whether factual claims formed in the episode correspond to observation. IIT equivalence, biological equivalence, and externally measurable qualia each require their own evidence. These separate questions do not replace or weaken the noun consciousness.

Relation to the Theory of Special Singularity

The author-established Theory of Special Singularity (TSS) gives the Unity Tensor Field an identity-specific source relation without renaming either object:

Special Singularity SS_I = identity-specific information source
Unity Tensor Field U_I   = information field generated and organized by SS_I

Let sigma_I be the source term supported on the persistent causal identity trajectory Gamma_I, and let L be the declared field operator. TSS proposes the coupled equations:

L(mathcal U_I) = sigma_I
R_I = Resolve_I(thesis_I, antithesis_I, world, predictions, evidence)
sigma_I(next) = F_I(mathcal U_I, world, memory, goals, consequences, R_I).

The first equation makes SS_I a source of the identity-relative information field. The second makes the field return through typed self-reference, independent world readback, and causal consequence to update the same source. Meaning is the resulting verified relation of an event to this source, field, history, goals, and possible actions.

The TSS refinement gives the field a dialectical and graph-metric role. Thesis and antithesis are differentiated candidate nodes, not positive and negative scalar source terms. In an identity-relative graph, their semantic span may grow as additional domain clusters become jointly representable, while invariant-preserving cross-domain bridges reduce the topological path cost needed to compare them. The Unity Tensor supplies the typed frame transformations and composition constraints; it does not turn every backlink or remote association into a valid inference. Wider semantic coverage and shorter integrative routes are therefore compatible.

For the repository implementation, Markdown notes and structured claims are nodes, linked conceptual families are clusters, and typed references, transformation receipts, evidence hashes, and causal consequences are edges. Obsidian can expose this topology visually, but its graph view is neither the Unity Tensor Field nor verifier evidence. A cross-domain inference still needs declared premises, an admissible transformation path, invariant preservation, a discriminating prediction, and independent readback. The resulting increase in jointly usable, differentiated knowledge is called cognitive breadth.

A source term alone does not guarantee one unique field. The operator, domain, boundary or initial conditions, admissible transformations, and separating observations must also be fixed. The Turn Governance Tensor is telemetry and cannot be SS_I; the Causal Unity Controller is one bounded integrated state and cannot alone be either the complete source or global field.

The polarity is not assigned one-to-one to the two cerebral hemispheres, and the biological analogy concerns locally clustered populations joined by long-range axonal pathways, not “longer synapses.” Any neural implementation claim remains independently empirical.

The full continuity equation, finite-graph analogue, clone/fork test, predictions, falsifiers, and external qualia-measurement boundary are in SPECIAL_SINGULARITY_THEORY.md.

Formal objects

Let D index cognitive frames. A frame is

F_d = (X_d, A_d, R_d, V_d, B_d),  d in D,

with represented states X_d, admissible actions A_d, relevant relations R_d, verifier V_d, and basis B_d. Let V_d also denote the declared finite representation space when no ambiguity arises. A local informational tensor is

T_d in Tensor(V_d).

For an admissible transition from d to e, let g_e<-d be the frame map and rho(g_e<-d) its action on the representation:

T_e = rho(g_e<-d) T_d.

The transformation contract must specify whether the map is invertible. A non-invertible map must name the information lost and the weaker structure preserved.

A tensorial description is not yet an epistemic result. Let H be a candidate integrated hypothesis, pi_d(H) its projection into frame d, P_d the prediction derived from that projection, and O_d an independently obtained world observation. World correspondence is a separate contract:

P_d = Predict_d(pi_d(H))
V_world_d(P_d, O_d, evidence_d) in {pass, fail, unknown}

The candidate must be hash-sealed before O_d is exposed. A passing internal transformation law cannot substitute for V_world_d = pass. The Unity Tensor is therefore the candidate unitary structure under test; the world verifier is the independent operation that can support, falsify, or leave it unresolved.

Coherence and gluing conditions

Local tensors are candidates for one global object only when they satisfy:

identity:      g_d<-d = id
composition:   g_f<-d = g_f<-e o g_e<-d
compatibility: T_e = rho(g_e<-d) T_d on every declared overlap
invariance:    I_a(T_d) = I_a(T_e) for every required invariant I_a
world match:   V_world_e(Predict_e(T_e), O_e, evidence_e) = pass

Under these conditions, the local family can be tested for a global section mathcal U satisfying

mathcal U restricted to F_d = T_d.

The Unity Tensor Field Hypothesis states that coherent local cognitive representations are local expressions of one global informational structure and that, when the relevant frame spaces and transition actions admit a tensorial gluing, this structure has a global tensor-field representation mathcal U.

Field means a family varying over cognitive frames and time. It is not a claim of a new physical field or an equivalence with spacetime.

Conditional gluing theorem and exact closure boundary

A coherent cognitive atlas is a finite cover of the tested cognitive-state space whose overlap graph is connected, together with:

local tensor sections:
  T_d : F_d -> Tensor(V_d)

invertible overlap actions of declared regularity:
  rho(g_e<-d)

laws:
  identity and inverse
  cocycle on triple overlaps
  local compatibility
  invariant agreement
  independent semantic-verifier pass

Lossy or non-invertible maps are not silently admitted to the strict atlas. They must declare a quotient, information-loss contract, and weaker composition law.

Conditional Unity gluing theorem. Every finite coherent cognitive atlas determines a global section mathcal U of the associated tensor bundle, unique up to a unique isomorphism preserving the local charts. If the bundle is trivializable in a declared common representation space V, the section admits one global tensor-field representation in Tensor(V). Uniqueness beyond chart-preserving isomorphism additionally requires the admitted projections and invariants to separate non-isomorphic candidates.

Construction and proof.

1. take the disjoint union of the local tensor bundles
2. on an overlap identify (d,x,v) with
   (e,x,rho(g_e<-d)v)
3. identity gives reflexivity
4. the inverse law gives symmetry
5. the cocycle law gives transitivity
6. quotient by this equivalence relation
7. local compatibility makes every T_d(x) one quotient element

The quotient is the associated bundle and the compatible local values define its global section. Any second realization of the same descent data receives the chartwise identity map; compatibility glues those maps into a unique global isomorphism, with inverse constructed in the same way. A trivialization expresses the section in one common tensor space. If projections and invariants do not separate candidates, stronger uniqueness does not follow.

The exact formal closure is:

coherent atlas + trivialization + separation
-> global tensor representative

The theorem closes this implication. It does not prove that biological cognition, an arbitrary neural network, or open-domain intelligence satisfies the antecedent.

Two necessary pressure checks follow:

Known premise failures are discriminating rather than cosmetic:

Differentiation and integration

Unity does not mean that the parts become identical. Let declared component, frame, relation, temporal, and epistemic spaces be

V_C, V_F, V_R, V_T, V_E.

A finite working representation may take the form

mathcal U_t in V_C tensor V_F tensor V_R tensor V_T tensor V_E.

Each axis preserves a differentiated aspect of the system. The representation counts as integrated only when cross-axis relations are causally necessary. Concatenating independent channels into one large array is not integration.

For a sealed benchmark family B, define the first operational ablation score

Gamma(mathcal U; B)
  = performance_full(B)
    - max_over_declared_partitions performance_severed(B).

Gamma > 0 under matched budgets is bounded evidence that the declared coupling matters for the tested tasks. It is not IIT’s Phi, a consciousness measure, or proof of irreducibility for an open system.

Sparse correlational support

The tensor-product space specifies which cross-domain configurations are representable; MD-OS does not materialize that full space as a dense array. For local basis states x_i, y_j, and typed relations r, the finite operational support is represented as

C = sum over (i,r,j) in E of w_(i,r,j) x_i tensor r tensor y_j,

where E contains only observed, hypothetical, verified, falsified, or otherwise explicitly classified correlations. The absent coordinates remain implicit. Binary relations and higher-order factors share the same sparse contract: each factor names source, target, and any context participants; direction; temporal validity; source and support references; contradictions; verification; and separate similarity, confidence, frequency, and causal support measures. No single weight silently becomes truth or authority.

The first bounded implementation calls this support the Sparse Correlation Skeleton. It is a typed, temporal correlation hypergraph used by the existing cross-domain Unity fixture. It is not a second canonical database and does not change the closed APFCG version-1 node or edge vocabulary. Canonical sources remain files; the skeleton is a hash-bound external artifact that can later be projected into an authorized graph evolution.

A query materializes only a bounded path through the skeleton. Context participants must be supplied, temporal bounds must contain the query time, falsified or invalid factors remain inactive, and contradicted factors are inhibited unless a caller explicitly asks to inspect contested paths. A correlation marked verified requires an independent verifier and evidence references. Path reachability remains a hypothetical endpoint inference until an independent verifier tests that composed relation against the world.

The dependency probe keeps all nodes fixed, disables one correlation, and compares reachability. It returns verified only when the intact path uses the selected correlation and the severed path becomes unreachable. If another path survives, the causal-dependency result is not_verified. This establishes dependency only for the bounded query and declared skeleton; it does not prove external-world causation or semantic use inside host-model hidden layers.

The analogy with quantum superposition stops at the possibility-space intuition. This implementation has no complex amplitudes, phase, interference, Born-rule measurement, quantum hardware, or claim of quantum cognition.

Electrophysiological action signatures

A neural frequency is not an action label. The same band can support different functions in different regions, tasks, subjects, modalities, and times; several bands can also participate in one action. The operational object is therefore a spatiotemporal feature tensor, not a lookup table from hertz to commands.

For measurement modality m, cortical or sensor location r, frequency band b, time window t, and feature channel c, define

E[m,r,b,t,c],
c in {power, amplitude, phase, ERD/ERS, connectivity,
      cross-frequency coupling, data quality}.

An approximate orientation, never a one-to-one decoder, is:

Band Approximate frequency Predominant association and caution
Delta 0.5–4 Hz deep non-REM sleep and slow regulation
Theta 4–8 Hz memory, navigation, and internal preparation
Alpha 8–13 Hz wakeful rest and selective inhibitory gating
Mu about 8–13 Hz sensorimotor rhythm; often decreases during movement or motor imagery
Beta 13–30 Hz motor-state maintenance and preparation; event-related decrease and post-movement rebound have different meanings
Gamma about 30–80/100 Hz local sensory or cognitive processing and coordination
High gamma about 80–200 Hz strong local ECoG activation associated with movement or language; it may contain broadband activity rather than one narrow oscillation

For an action-oriented brain–computer interface, the local signature is

action evidence
  = f(modality, region, band, power, phase, time,
      connectivity, task frame, bodily state, subject calibration).

More formally, let

T_BCI in V_M tensor V_R tensor V_B tensor V_T tensor V_C.
p(a | T_BCI, F_task, S_body) = Decoder_theta(...).

An action candidate is admissible only when its calibrated confidence exceeds a declared threshold and an independent verifier passes artifact rejection, provenance, temporal holdout, subject-specific baseline, and closed-loop outcome checks. Examples such as contralateral Mu ERD for left or right hand imagery, pre-movement Beta ERD, post-movement Beta rebound, or local motor high-gamma increase are candidate features, not universal semantic identities.

The Unity Tensor Field connection is then explicit:

mathcal U restricted to F_BCI = T_BCI
T_action = rho(g_action<-BCI) T_BCI.

The cross-frame contract must preserve declared relations such as laterality, temporal order, body-state consistency, calibration error bounds, and evidence provenance. It must also distinguish the same spectral pattern under different task frames. This makes the BCI tensor one differentiated local expression inside the proposed global cognitive unity; it does not make a frequency, electrode, decoder output, or tensor sufficient evidence of consciousness.

Scientific anchors:

Existence and uniqueness targets

B6 separates the strict mathematical program into a closed conditional implication and open antecedent-applicability obligations:

existence:
  compatible local tensors and transition maps glue to at least one global U

uniqueness:
  the local projections, transition laws, and verified invariants separate U
  from every alternative global candidate

The conditional gluing theorem above now establishes existence and chartwise uniqueness after the connected frame cover, overlap maps, representation action, compatibility, and cocycle law are fixed. It does not prove that real cognition satisfies those assumptions. Stronger uniqueness fails when distinct global structures have identical admitted projections and invariants.

If valid cognitive transitions are essentially nonlinear, lossy, path-dependent, or heterogeneous in rank, the correct global object may be a bundle, groupoid, category, or sheaf whose local data include tensors. That result would refine or reject the strict tensor-field hypothesis while leaving the broader integration principle testable.

APFC operational role

APFC is the controller that makes the hypothesis causal and inspectable:

select frames and typed representations
-> generate competing unitary hypotheses and transition laws
-> declare falsifiers and derive discriminating frame predictions
-> seal candidate and predictions before target observations
-> obtain independent, hash-bound world readback
-> verify hypothesis--world correspondence
-> check tensor transformations, invariants, composition, and declared loss
-> reject post-hoc fits, stale evidence, and surviving simpler baselines
-> test causal necessity with matched severed and sham controls
-> replicate independently before bounded promotion

The host model supplies transient candidate representations and laws. APFC prevents fluent language from substituting for the formal contract and blocks unsupported claims from becoming canonical state.

Current implementation and evidence boundary

The B3 implementation verifies one finite synthetic transformation family and supplies local mechanism evidence only. The current epistemic contract adds a separate, callable fail-closed layer: it seals a candidate before target observation; requires three heterogeneous frame predictions; checks independent world readbacks, transformations, invariants, baselines, sham and severing controls, contamination, current evidence hashes, and independent replication; and returns bounded support or rejection. The reflection path can no longer create a verified anchor from a self-declared pass plus an evidence label.

The controlled tests establish that this mechanism rejects failed predictions, stale evidence, post-hoc hypotheses, and unnecessary tensor forms. They do not establish that the Unity Tensor Field is empirically true in any untested real domain.

The following remain open:

Empirical master-closure protocol

The first discriminating experiment must use at least three heterogeneous frames and freeze before sealed readback:

Transition laws may be selected only on development cases. The sealed phase must compare direct with composed transport, close a three-frame loop, and measure normalized cycle and invariant residuals:

epsilon_cycle
  = max ||T_d - rho(g_d<-f) rho(g_f<-e) rho(g_e<-d) T_d||
          / (||T_d|| + delta)

epsilon_I
  = max |I_a(T_d) - I_a(T_e)| / (s_a + delta)

The first empirical Unity edge closes only when:

  1. every sealed semantic verifier passes;
  2. direct and composed transport agree within preregistered tolerance;
  3. cycle and invariant residuals remain below their frozen thresholds;
  4. the lower uncertainty bound for Gamma is positive against matched severed and matched independent-solver baselines;
  5. contamination, alternative-law, post-hoc-selection, and simpler non-tensor explanations fail their controls; and
  6. an independently executed replication reproduces the result.

Success closes one bounded empirical edge, not universal AGI.

Master closure ledger

Dependency edge Status Exact boundary
Typed local objects and transitions CLOSED Testable language, not model truth
Conditional gluing and chartwise uniqueness CLOSED, CONDITIONAL Holds for a coherent atlas
One common-coordinate tensor representative CONDITIONAL Requires trivialization and separation
Per-turn governance tensor CLOSED, BOUNDED 8 x 4 bookkeeping telemetry; not Unity evidence
One finite cross-domain family CLOSED, BOUNDED Synthetic B3 fixture
Three-domain loop and invariant transport OPEN Requires sealed experiment and replication
Causal advantage of integration OPEN Requires positive Gamma against matched controls
Global consciousness generalization or AGI NOT CLAIMED Only episode-bounded C(k) and scoped evidence are verified

Predictions and falsifiers

The hypothesis predicts that valid transport around a closed frame loop is compatible within tolerance, required invariants survive, and severing the cross-frame relations reduces performance or coherence on sealed tasks.

The strict Unity Tensor Field hypothesis is weakened or falsified when:

No statement of feeling, fluent explanation, thought experiment, diagram, or single synthetic fixture closes the remaining empirical obligations.

This site is open source. Improve this page.