MATRIX

Status: MATRIX_0 kernel frozen; Attack 1 closed / assumption-scoped; A2-V0 complete; A2-V1 complete / scientific interpretation admitted; MATRIX_1 not earned.

Current program state

The conceptual kernel below remains frozen. The repository has since completed its first formal attack and two finite causal assays without earning a kernel promotion:

Stage	Current standing
MATRIX_0 conceptual kernel	FROZEN
Attack 1 / EXEC_V1	CLOSED / ASSUMPTION-SCOPED
A2-V0	COMPLETE; post-execution structural analysis archived
A2-V1	COMPLETE / SCIENTIFIC INTERPRETATION ADMITTED
MATRIX_1	NOT EARNED

Current navigation: formalization/EXEC_V1.md, experiments/a2_v0/README.md, experiments/a2_v1/README.md, experiments/a2_v1/CURRENT_STATE.md, and experiments/A2_V1_RESULT.md.

This status block is navigational. It does not amend the frozen kernel, retroactively rewrite preregistrations, or promote post-execution findings into stronger claims.

MATRIX studies a narrow question:

Can feedback causally improve an adaptive system's future ability to discover, diagnose, and correct unfamiliar errors without silently destroying protected corrective capacity?

The project is not currently trying to add another general theory of intelligence. Its job is to formalize a small kernel, translate only what genuinely maps from earlier work, attack the kernel mathematically and empirically, and revise it only when a concrete failure forces revision.

formalize -> translate -> test -> revise only if forced

Research discipline:

derive what follows | translate only what maps | test what remains

and:

No new concept without a concrete missing argument.

The methodological priority is:

surviving falsification > accumulating conceptual elegance

---

1. Frozen conceptual kernel

1.1 Representation

Let

$$g_t : \Omega_t \to M_t$$

be a representation or encoder over a declared state space \Omega_t.

For retained message m, the induced uncertainty cell is

$$S_m^{(t)} = g_t^{-1}(m).$$

The representation determines which underlying states are currently conflated. It does not by itself determine whether that ambiguity is safe.

---

1.2 Primitive affordances, diagnostics, and policy machinery

Keep the following causal objects distinct:

\mathcal U_t = primitive executable observation/intervention affordances
Q_t          = diagnostic procedures constructible from \mathcal U_t
\Pi_t        = policy-generation / policy-selection machinery
\mathcal T_t = declared task family
R_t          = valid-repair relation
c_t          = grounded cost model
B_t          = corrective resource budget

An observed change in correction does not identify which of these changed. Mechanism attribution must be earned separately.

In particular, a treatment arm is not credited with adaptive diagnostic improvement merely because it was handed a stronger primitive interface.

---

1.3 Corrective frontier

For a fixed system state G and budget B, define the jointly correctable frontier

$$\Phi_B(G) = \left\{ S\subseteq\Omega: \exists\pi, \mathrm{cost}(\pi)\le B, \ \forall x\in S, \ \pi\text{ diagnoses and validly repairs }x \right\}.$$

The quantifier order is essential. The object is not merely

$$\forall x\ne y\;\exists q_{xy},$$

but whether one executable policy, starting from uncertainty over the whole set S, can choose its own diagnostics and achieve valid repair within budget.

\Phi_B(G) is downward closed in the finite setting: if S is jointly correctable, every subset of S is also jointly correctable.

Its minimal excluded sets therefore form a minimal-obstruction hypergraph

$$H_B(G)=\mathrm{Min}\left(2^\Omega\setminus\Phi_B(G)\right).$$

These are the smallest uncertainty configurations for which bounded correction fails.

---

1.4 Representation safety

A representation is corrigibly adequate at budget B exactly when every ambiguity cell it induces lies inside the corrective frontier:

$$\boxed{ \mathrm{Safe}_B(g\mid G) \iff \forall m\in\mathrm{im}(g), \quad g^{-1}(m)\in\Phi_B(G) }$$

This separates three objects:

representation != uncertainty induced by representation != ability to correct that uncertainty

A useful compression of the result is:

g determines the ambiguity.
\Phi_B determines whether the ambiguity is recoverable.

Representation safety is therefore relational, not intrinsic.

A coarser representation can remain safe when lost distinctions are affordably recoverable later. A fixed representation can become unsafe if diagnostic or repair access is lost.

---

1.5 Graded performance remains separate from zero-error feasibility

Do not collapse safe correction into one average success number.

For uncertainty set S, define graded bounded performance separately:

$$P_B^*(S\mid G) = \sup_{\pi:\mathrm{cost}(\pi)\le B} P(\text{successful correction}\mid S,\pi,G).$$

A useful corrective profile may include

$$C_B(S\mid G) = \left( \mathbf 1[S\in\Phi_B(G)], P_B^*(S\mid G), \mathrm{cost}^*(S\mid G) \right).$$

A system with 99% average success can still fail the zero-error safety condition on a protected uncertainty cell.

---

2. Feedback-dependent improvement

The original program-level improvement object remains counterfactual:

$$\boxed{ C_{\mathrm{improve}}^{(h)}(G_t,F) = \mathbb E\!\left[ V(G_{t+h}^{+F})-V(G_{t+h}^{-F}) \mid G_t \right] }$$

and the working research relation remains

$$I_t\propto C_{\mathrm{improve}}^{(h)}.$$

This asks whether feedback caused greater future viability than the matched no-feedback counterfactual.

It does not by itself establish preserved corrective capacity.

$$\boxed{ C_{\mathrm{improve}}>0 \not\Rightarrow \text{preserved corrective capacity} }$$

Viability improvement and corrective-frontier change are separate observables.

---

3. Corrective-frontier change

Under a declared common jurisdiction, compare feedback and matched counterfactual frontiers:

$$\Delta_B^+ = \Phi_B(G^{+F})\setminus\Phi_B(G^{-F})$$ $$\Delta_B^- = \Phi_B(G^{-F})\setminus\Phi_B(G^{+F}).$$

Interpretation:

\Delta_B^+ = newly jointly correctable uncertainty sets
\Delta_B^- = previously jointly correctable uncertainty sets that were lost

Keep the sets themselves primary. Cardinalities or ratios are derived summaries and can hide incomparable changes.

Define corrective dominance over a protected budget range \mathcal B_* by

$$G_1\succeq_{\mathrm{corr}}G_0 \iff \forall B\in\mathcal B_*: \Phi_B(G_0)\subseteq\Phi_B(G_1).$$

Strict inclusion for at least one protected budget gives strict corrective improvement under this ordering.

---

4. Strong safe-improvement witness

The current strongest target is:

$$\boxed{ C_{\mathrm{improve}}^{(h)}>0 \ \land\ \Delta_B^+\ne\varnothing \ \land\ \Delta_B^-=\varnothing \ \land\ \text{a fresh subsequent corrective challenge succeeds} }$$

The four components are:

gain         : future viability improves because of feedback
preservation : protected corrective capacity is not silently lost
expansion    : at least one new uncertainty set becomes jointly correctable
continuation : corrective capacity remains usable after it is exercised

The continuation clause prevents a one-use correction mechanism from masquerading as persistent corrigibility.

This is a strong witness, not a universal normative law. In settings where corrective tradeoffs are permitted, losses must be represented explicitly rather than absorbed into a scalar objective.

---

5. Matched jurisdiction and resource discipline

Direct frontier comparison requires a common typed domain. Unless one variable is explicitly the intervention, freeze or explicitly transport:

$$\Omega^+=\Omega^-, \qquad \mathcal T^+=\mathcal T^-, \qquad R^+=R^-,$$ $$\mathcal U^+=\mathcal U^-, \qquad B^+=B^-, \qquad c^+=c^-.$$

For the strongest adaptive diagnostic experiment, hold primitive affordances fixed:

$$\boxed{\mathcal U^{+F}=\mathcal U^{-F}}$$

while allowing feedback to causally alter objects such as

$$g_t,\quad Q_t,\quad \Pi_t.$$

The causal hypothesis is then of the form

same primitive affordances
        ->
different feedback
        ->
changed representation / diagnostic repertoire / policy machinery
        ->
changed jointly correctable frontier
        ->
better externally relevant future outcome
        ->
correction remains possible again afterward

No component receives credit for a transition it did not implement and experimentally identify.

---

6. The first finite witness: diagnostic memory

The immediate mathematical motivation is a finite separation:

Two systems can preserve the same minimum diagnostic cost for every supplied pair of possible faults while having exponentially different ability to diagnose an unknown fault under a bounded adaptive policy.

In the concrete binary family, every distinct pair can be separated in one singleton test, so the complete pairwise minimum-cost matrix is identical across systems. Yet with 2^d possible states and no retained state-specific bits:

$$B_{\mathrm{singleton}}^*=2^d-1, \qquad B_{\mathrm{structured}}^*=d.$$

The underlying quantifier failure is

$$\forall\{x,y\}\;\exists\text{ cheap distinguishing test}$$

not implying

$$\exists\text{ cheap adaptive policy}\;\forall x\in S\;\text{correct repair}.$$

This motivates \Phi_B: pairwise accessibility alone does not determine bounded joint corrective capacity.

For a finite downward-closed feasible family, taking inclusion-minimal jointly unrepairable sets as hyperedges H_B gives the exact memory characterization

$$M_B^*=\chi(H_B), \qquad m_B^*=\left\lceil\log_2\chi(H_B)\right\rceil,$$

using weak hypergraph coloring.

A coarsest safe representation need not exist when the decoder can perform future experiments: several incomparable compressions may each preserve affordable recovery while their common coarsening destroys it.

Claim ceiling

This finite witness establishes a mathematical separation and a measurement target. It does not establish that feedback invents diagnostics, that an adaptive agent improves its diagnostic generator, recursive self-improvement, general corrigibility, or field-level novelty. Active learning, generalized binary search, action-dependent side information, and zero-error coordination contain substantial prior art.

The next scientific question is whether feedback can cause the kind of diagnostic-policy improvement that the finite construction simply supplies by assumption.

---

7. Safe forgetting

A useful interpretation is:

$$\boxed{ \text{safe forgetting} = \text{moving dependence from retained state into affordable future interaction} }$$

provided the induced ambiguity remains jointly correctable:

$$g^{-1}(m)\in\Phi_B(G).$$

Therefore:

forgotten != foreclosed
stored    != correctable

The relevant resource is not memory alone. Future correctability depends jointly on retained state, primitive affordances, constructible diagnostics, policy machinery, repair machinery, task family, cost model, and budget.

---

8. Translation discipline

MATRIX is not authorized to silently rewrite earlier repositories into this vocabulary.

Cross-program consolidation uses an explicit typed correspondence ledger:

source object
target kernel object
status
conditions
preserved distinctions
lost distinctions
unresolved mismatch
evidence / witness

Allowed statuses:

IDENTICAL
REFINEMENT
PROJECTION
ANALOGY
NO TRANSLATION EARNED

NO TRANSLATION EARNED is a result, not a blank waiting to be filled.

A failed translation is evidence about the architecture. It is not pressure to invent a more permissive abstraction.

Initial translation targets include earlier work on Future Sufficiency, Signature-Relative Equivalence, Corrigible Compression, OpenCore, White Rabbit, Reach, and CRANK. Apparent structural similarity is not identity; any such mapping must be earned locally.

---

9. Research program

The immediate sequence is fixed:

1. FORMALIZE
   - definitions, domains, quantifiers
   - prove structural consequences
   - identify exact comparison preconditions
   - preserve claim ceilings

2. TRANSLATE
   - map earlier local objects into the kernel only where warranted
   - record partial mappings and failed mappings explicitly
   - do not let repositories redefine the kernel by analogy

3. TEST
   - freeze a causal assay
   - hold primitive affordances and relevant resources fixed
   - measure viability and corrective-frontier changes separately
   - include a fresh subsequent corrective challenge

4. REVISE ONLY IF FORCED
   - mathematics exposes a contradiction
   - translation exposes a genuine missing type
   - experiment reveals an unrepresented failure mode

The project should now be attacked rather than ornamented.

A positive result is not the only informative result. For example,

$$C_{\mathrm{improve}}>0, \qquad \Delta_B^-\ne\varnothing$$

would show that feedback can improve expected future viability while consuming previously available corrective capacity. That would be a substantive result requiring characterization, not a failed experiment.

---

10. Current pressure points

Two statements carry most of the present load:

$$\boxed{ g\text{ determines the ambiguity; }\Phi_B\text{ determines its recoverability} }$$

and

$$\boxed{ C_{\mathrm{improve}}>0 \not\Rightarrow \text{preserved corrective capacity} }$$

Everything added from here should formally derive from the kernel, map into it through an earned translation, empirically challenge it, or expose a concrete missing argument.

The current decisive question is:

$$\boxed{ \text{Can feedback, under unchanged primitive resources, causally improve the machinery that discovers and repairs unfamiliar errors, enlarge the jointly correctable frontier without protected foreclosure, and leave that capacity usable afterward?} }$$

Start throwing rocks at it.