← Complete research archive
Theory & no-go resultsPreregistered529 lines

R12 Source-Firewalled Separating Query Basis Theory

Decision in one sentence: supervising one query-oblivious state against a separating family of future questions can certify that the state has not aliased the tested causal quotient, but it is predictive-state representation learning, not a new reasoning primitive; the only survi…

R12_SEPARATING_QUERY_BASIS_THEORY.mdOpen original Markdown ↗

R12 Source-Firewalled Separating Query Basis Theory

Status: theory and falsifier schema only. No implementation, data build, fit, score, cluster copy, CPU publication, accelerator job, capability claim, or novelty claim is authorized by this file.

Decision in one sentence: supervising one query-oblivious state against a separating family of future questions can certify that the state has not aliased the tested causal quotient, but it is predictive-state representation learning, not a new reasoning primitive; the only surviving experimental conjecture is that source-firewalled multi-query supervision may allocate source-processing compute more efficiently and make Shohin's existing post-DRS state easier to read, update, and reuse.

1. Empirical trigger and claim boundary

The immutable post-DRS probe establishes a narrow but strong fact. On the same 40 matched carry directions used by the raw-200k control, layer-29 residual swaps move the DRS carry-token margin toward the source on 40/40 directions with mean delta-logodds +3.147395; raw-200k is 20/40 with mean +0.014188. DRS result-digit swaps are 40/40 toward-source at every tested layer 17--29. The evidence spans fit width four/six, value-OOD width four/six, and width-OOD width eight.

This proves neither autonomous state update nor reasoning. The prefixes are teacher-forced, the intervention replaces a full residual vector, and each measurement concerns one microstep. Autonomous digit-motor results further show that a high-accuracy teacher-forced reader can fail to improve the closed loop. The unresolved failure is therefore not simply absence of a local signal. It is the joint problem of:

  1. reading a task-relevant state rather than exploiting a global logit bias;
  2. committing the read state before output corruption;
  3. updating the state under a new event;
  4. consuming the updated state on later steps; and
  5. preserving unrelated language and direct-answer behavior.

The candidate in this document is a training and measurement protocol for those properties. It is not claimed to be a new state ontology, architecture, or computational class.

2. Formal object

Let H be a finite set of admissible histories, C a set of future continuations, Q a set of late queries, and A a finite answer set extended with inadmissibility symbol bottom. The task relation is deterministic here:

R(h, c, q) in A union {bottom}.

Histories have the usual residual equivalence

h == g  iff  R(h,c,q) = R(g,c,q) for every (c,q) in C x Q.

Write S(h)=[h] for the causal state. An event e induces the residual derivative

U_e(S(h)) = S(he),

and a late query observes

O_(c,q)(S(h)) = R(h,c,q).

An exact realization remains an ordinary Moore/Mealy residual transducer up to a change of coordinates. Nothing below changes that no-go theorem.

2.1 Separating query basis

A finite family

B = {(c_1,q_1), ..., (c_k,q_k)}

is separating on a subset S_0 of reachable causal states when

s != t  implies  there exists b in B with O_b(s) != O_b(t).

Its response signature is

Phi_B(s) = (O_b(s))_(b in B).

The basis need not be minimal or unique. Calling it a basis asserts only that its signatures separate the frozen state set. It does not assert linearity, independence, canonical coordinates, or a direct-product factorization. The counterexamples in R12_QUERY_KERNEL_FACTORIZATION_NO_GO.md remain binding.

2.2 Source-firewalled realization

An encoder commits to a query-oblivious state before learning which basis query will be asked:

z_h = E(h).

After commitment, the original history bytes, source-token KV, generated result tape, cached answers, and any host-computed semantic state are inaccessible. Only z_h, the late continuation c, and query q may reach the consumer:

D(z_h,c,q).

For recurrent use, the updater may receive only the prior committed state and the current event:

z_(he) = T(z_h,e).

Re-encoding the full history, reading previous generated answers, parsing an emitted state, host arithmetic, retrieval, verifier correction, retries, or source-visible attention after commitment changes the resource model and is an ineligible treatment.

3. What a separating loss can certify

Theorem 1: exact non-aliasing

Let B separate S_0. Suppose one deterministic encoder E and consumers D_b satisfy

D_b(E(h)) = O_b(S(h))

for every reachable S(h) in S_0 and every b in B. Then

E(h) = E(g)  implies  S(h) = S(g)

on S_0.

Proof. If E(h)=E(g), every deterministic consumer receives the same input, so every basis response is equal. Hence Phi_B(S(h))=Phi_B(S(g)). Separation implies S(h)=S(g). QED.

This is an injectivity certificate only on the frozen state set and query family. It gives no minimality result. The representation may contain the entire source, an arbitrary lookup key, or irrelevant information unless the firewall and resource ledger exclude those paths.

Theorem 2: approximate geometric separation

Let each true query response be represented by a distribution P_b(.|s). Let B be gamma-separating in total variation:

for every s != t, some b has TV(P_b(.|s), P_b(.|t)) >= gamma.

Assume every learned decoder Q_b(.|z) is L-Lipschitz from state norm to total variation and has uniform error

TV(Q_b(.|E(s)), P_b(.|s)) <= eta

for all s,b, with eta < gamma/2. Then every distinct pair satisfies

||E(s)-E(t)|| >= (gamma - 2 eta) / L.

Proof. For the separating b, the triangle inequality gives

gamma
 <= TV(P_b(.|s), P_b(.|t))
 <= eta + TV(Q_b(.|E(s)),Q_b(.|E(t))) + eta
 <= 2 eta + L ||E(s)-E(t)||.

Rearrange. QED.

The bound is diagnostic, not a reasoner theorem. A highly non-Lipschitz decoder or a weak empirical estimate of uniform error makes it vacuous.

Theorem 3: representation sufficiency does not imply update correctness

For any injective E on a finite state set with at least two states, there is an updater T_bad that is wrong on every non-self transition while every immediate basis decoder remains exact.

Construction. Let T_bad(z,e)=z for every input. Immediate basis queries at the encoded histories are unchanged and remain exact, but every event that changes causal state is mapped incorrectly. QED.

Therefore a query-basis fit cannot promote autonomous reasoning. Update and later consumption require separate interventions and closed-loop tests.

Corollary: query-specific readers do not prove a shared state

If each query is allowed its own encoder E_b(h), perfect answers do not imply that any one state separates the quotient. Query identity must be hidden until after a single committed z_h is frozen, and all consumers must be proven to read those exact committed bytes.

4. Equivalence and prior-art dossier

The central object is established machinery:

  • Predictive Representations of State represents state by predictions of tests and proves that a linear PSR need not exceed the minimal POMDP state count. A separating future-query signature is a deterministic finite PSR / observable quotient.
  • Bisimulation and latent-model learning, including DeepMDP, train latent states to preserve reward/transition behavior. Adding an explicit updater and future observations is state-representation learning under the same boundary.
  • Contrastive Predictive Coding trains a context representation to predict future latent content. Negative sampling or a contrastive query signature would be a CPC-family auxiliary loss, not a new primitive.
  • Multi-task future prediction, successor features, auxiliary state losses, knowledge distillation, and ordinary supervised sufficient-statistic learning all cover nearby training objectives.
  • A learned updater over the committed state is an ordinary RNN/Mealy transducer. Fixed-length execution can be unrolled into a feed-forward circuit. Any contribution must be resource-relative rather than ontological.
  • The 2026 global-workspace study provides evidence that some verbalizable residual vectors are broadcast and flexibly consumed. It motivates the intervention surface, but does not make a future-query auxiliary objective new or prove that Shohin can update the relevant state.

Novelty decision: rejected as a new reasoning primitive. The only reopenable claim is a bounded training/oracle-allocation and causal-measurement protocol under an explicit resource vector.

5. Surviving resource conjecture

Let K separating queries supervise one history prefix. A naive direct-SFT baseline that presents the source separately for every query processes roughly K copies of the source. A source-firewalled shared-state protocol can encode the source once and apply K small consumers:

naive source work       Theta(K * F_source)
shared-state work       Theta(F_source + K * F_consumer).

When F_consumer << F_source, this is a real source-processing advantage. It is not automatically a training-FLOP, example, or oracle-call advantage:

  • a packed multi-answer transformer may reuse source KV and erase the gain;
  • one oracle that returns a complete state label may dominate both methods;
  • constructing a separating basis may require more oracle calls than direct answers;
  • the shared encoder may need greater width or optimization effort;
  • an ordinary recurrent transducer may exploit exactly the same sharing.

The mandatory resource vector is

(parameters,
 retained bits and bytes,
 numeric precision,
 source bytes read before and after commitment,
 source encodings,
 training examples,
 oracle calls and oracle output bits,
 training FLOPs,
 inference FLOPs,
 model calls,
 sequential depth,
 external memory,
 external execution,
 generated-token/KV bytes visible to update,
 wall time and accelerator allocation).

Conjecture SQB-1

On a task whose reachable quotient has a compact separating query family, source-firewalled multi-query supervision can reach a fixed autonomous transition error with fewer source encodings than independently prompted direct SFT, while matching an ordinary recurrent transducer in all other resources.

This is intentionally weak. If packed direct SFT or the favorable recurrent control matches the source-encoding count and score, the resource claim is rejected. A behavioral gain without matched resources is package-level only.

6. Exact collapse tests before neural code

Any executable successor must first pass a CPU-only symbolic suite.

6.1 Residual-table reconstruction

For a finite deterministic board, enumerate all reachable states, events, continuations, and queries. Compute the exact residual table, minimize it by partition refinement, and verify:

  1. the declared basis separates every frozen quotient state;
  2. every proper subset claimed nonseparating has an explicit collision witness;
  3. appending the same event preserves declared equivalences;
  4. every claimed distinction has a concrete continuation-query witness;
  5. inadmissibility bottom is included in the signature.

6.2 Planted leakage negatives

The suite must prove that the validator rejects all of the following even when they score perfectly:

  • source-visible consumer after commitment;
  • query revealed before or during state encoding;
  • one encoder or cached answer per query;
  • full-history re-encoding on every step;
  • generated result tape or generated-token KV feeding the updater;
  • host arithmetic, parser repair, verifier correction, retry, or search;
  • answer table indexed by episode, seed, width, or prompt hash;
  • source bytes hidden in padding, dtype slack, filenames, environment, timing, RNG state, allocator state, or external files;
  • a nonseparating query family reported as separating;
  • a stale identity updater that fits only immediate queries;
  • score-dependent basis, seed, threshold, or confirmation selection.

6.3 Matched constructive controls

The same finite board must instantiate:

  1. the minimal exact residual transducer;
  2. a parameter/state-favorable ordinary RNN/Mealy realization;
  3. direct SFT with independent prompts;
  4. packed multi-answer direct SFT with source/KV reuse;
  5. neutral and shuffled future-query objectives;
  6. a source-visible upper bound;
  7. a query-before-state leakage upper bound;
  8. a favorable constant-logit calibration null at every typed readout;
  9. a favorable nuisance-only calibration null that may use preregistered operation, width, cursor, and terminal-position metadata but no committed hidden state; and
  10. retrieval and full-state-label controls with their resources disclosed.

If the proposed package reduces to any control with equal behavior and a constant/polylogarithmic resource-vector overhead, the corresponding novelty or resource claim is rejected.

7. Smallest Shohin falsifier schema

This section is not executable. A later byte-bound preregistration must freeze every row, source, hash, seed, threshold, denominator, and consumer before any fit.

7.1 Frozen parent and task

  • Parent: immutable post-DRS checkpoint, not the raw 300k base.
  • Backbone: frozen for the first falsifier.
  • Aperture: one predeclared late residual/register location.
  • Family: digitwise add/subtract transitions with exact protocol-derived oracles.
  • Fit regimes: widths four and six.
  • Development: disjoint width-four/six values and transition depths.
  • Confirmation: fresh value-OOD widths four/six and width-OOD eight/ten.
  • State is committed before a random hidden-independent query index is opened.
  • After commitment, original prefix/source history is unavailable. A recurrent update receives only the previous committed state and the current typed event.

The exact query family must be derived from the minimized finite residual table, not hand-picked after seeing model scores. It should include continuation-query witnesses that separate carry/borrow, terminality, and every other retained state distinction required by the declared task. Local digit labels alone are not a separating basis for multi-step execution.

7.2 Arms

All arms use frozen train/development/confirmation rows and receive favorable matching where exact matching is impossible.

ArmPurpose
baseFrozen DRS without new training.
sqbSource-firewalled shared state, separating queries, and recurrent update.
neutralSame query counts/tokens/compute, labels independent of causal state.
shuffledSame label marginals and budgets, state-query associations permuted within frozen strata.
direct_independentSeparate direct prompt/answer examples for each query.
direct_packedFavorable packed multi-answer SFT with shared source/KV.
rnn_mealyFavorable ordinary recurrent transducer with the same state bytes, events, heads, losses, and compute.
source_visibleUpper bound whose consumer may reread source; ineligible for a reasoning claim.
constant_biasBest grammar-gated state-independent logit calibration at every typed readout.
nuisance_calibrationFavorable grammar-gated calibration using only operation/width/cursor/terminal metadata, never the committed residual. Width extrapolation is fit on training widths before width-OOD reveal.
stale_updateCorrect reader with identity/frozen updater, proving immediate query fit is insufficient.

Treatment and rnn_mealy must share ordered data, semantic targets, loss coefficients, optimizer, update count, precision, state allocation, consumer capacity, source calls, and inference depth. If that leaves no treatment variable, SQB is only the training package and no architecture attribution is allowed.

7.3 Required interventions

  1. Query concealment: changing the unopened future query cannot alter the committed state bytes.
  2. State swap: swapping committed states between matched histories must redirect all basis answers toward the source state.
  3. Event swap: holding state fixed and swapping the next event must redirect the successor state and later answers according to the exact derivative.
  4. Updater ablation: zeroing or staling the updater must damage two-plus-step execution while leaving immediate reader accuracy largely intact.
  5. Consumer ablation: ablating the reader must damage answers without changing the committed state identity.
  6. Source deletion: the exact same scores must survive physical removal of all pre-commit source bytes available to the consumer/updater.
  7. Generated-output deletion: removing emitted tokens and generated-token KV from the causal path must not change state transitions.
  8. Counterfactual continuation: the same committed state must support fresh unseen separating continuations without re-encoding the prefix.

8. Noncompensatory decision gates

A later executable protocol may tighten these floors, but may not weaken them after any score is observed.

A. Custody and resource gates

  • Every source, row set, implementation, runner, runtime closure, checkpoint, report, and receipt hash matches before and after execution.
  • The protected base remains immutable and every candidate output is isolated.
  • Total unique parameters remain strictly below 150,000,000; with base 125,081,664, additions must not exceed 24,918,335.
  • Post-commit source bytes, generated-output bytes, host semantic operations, retries, retrieval calls, and external execution visible to the causal path are exactly zero.
  • The complete resource vector is machine-readable and replayed independently.

B. Basis and representation gates

  • The exact residual table proves that the frozen basis separates every scored quotient state; one counterexample is failure.
  • Query identity is hidden-independent until after commitment; changing it changes zero committed-state bytes.
  • Confirmation basis accuracy is at least 99% on fit widths and at least 95% on both value-OOD and width-OOD states.
  • Every matched state-swap direction follows the source signature on every required query family; aggregate means cannot compensate for a failed family.

C. Calibration and feature-use gates

  • sqb must beat the favorable constant_bias arm by at least 10.0 percentage points on each of positive carry, negative carry, value-OOD, and width-OOD exact readout, with positive difference on every seed.
  • sqb must also beat nuisance_calibration by at least 10.0 points on every confirmation stratum. Within-stratum label balance, not aggregate balance, is mandatory. A metadata-only tie rejects feature-dependent state reading.
  • Gate-off logits are bitwise identical across base, sqb, and both calibration controls.
  • sqb must beat shuffled and neutral controls on every confirmation stratum.

D. Update and autonomous gates

  • One-step successor-state signature exactness is at least 95% on every confirmation stratum.
  • Two-, four-, and eight-step exactness are reported separately. Every depth must beat stale_update; no shorter-depth gain can compensate for an eight-step failure.
  • Width-eight and width-ten complete transition traces each reach at least 50% exact and beat base by at least 15.0 points on every seed.
  • sqb beats the favorable rnn_mealy arm by at least 10.0 points on the mean of value-OOD and width-OOD complete-trace exactness to support an SQB-specific package claim. If they tie, retain the simpler ordinary transducer and reject SQB-specific advantage.
  • Autonomous execution uses no teacher-forced state/query answers and no post-failure repair.

E. Preservation and interaction gates

  • Existing frozen language, primitive, and direct-answer preservation boards do not regress beyond their preregistered confidence bounds.
  • Fresh direct interaction includes unseen arithmetic transitions, state perturbations, requests for review, and compact-state reuse. Transcripts are read manually before any summary claim.
  • A visible trace or correct final answer without an exact internal transition record cannot substitute for the autonomous state gates.

9. Kill conditions

Stop before H100 work if any of these holds:

  1. the basis is nonseparating or was chosen after score access;
  2. a packed direct or ordinary recurrent control preserves the resource vector and subsumes the claimed advantage;
  3. source, query, output, host, retry, or retrieval leakage is nonzero;
  4. constant or nuisance-only calibration matches the learned reader;
  5. immediate query fit is high but event-update or later-consumption interventions fail;
  6. width/value-OOD autonomous execution does not improve on every required stratum;
  7. the parameter ledger reaches or exceeds 150,000,000;
  8. any custody identity, receipt, runtime closure, or confirmation secrecy gate fails.

No rename, wider head, extra recurrence, larger fit, or new seed may rescue a failed exact gate without a new preregistration and independent review.

10. Final claim boundary

Passing Theorems 1--2 empirically would establish only that one committed state is sufficient to answer the frozen separating query family without tested aliasing. Passing the interventions and autonomous gates would establish a bounded learned state-update/consumption result for the declared arithmetic transducer. It would still not establish a new computational primitive, consciousness, arbitrary natural-language reasoning, broad intelligence, or an asymptotic separation.

The strongest admissible positive is:

Under a frozen source-firewalled protocol and complete resource ledger, multi-query supervision produced a causally read, updated, and reused state on the declared held-out digitwise family, and outperformed the named matched controls by the preregistered margins.

Until those gates pass, the current conclusion remains narrower: DRS contains a causally actionable local residual, while autonomous transport, update, and reuse remain unsolved.