The oracle machine end-state: a threshold theorem for reasoning, and where lying belongs

1. What "oracle machine end-state" actually means, technically

Turing's oracle machine (1939) is a computer with access to a black box that returns correct answers to a class of queries. The end-state of this project's Oracle is the concrete realization of that black box built from parts that are not individually reliable: a small model on an 8GB box, wrong a large fraction of the time. The whole design question reduces to one sentence:

How do you build a reliable oracle out of unreliable components?

That is not a new question. It is the oldest question in the theory of computation, and it has a complete answer with a sharp phase transition — which this session's measured K_c ≈ 1/q turns out to be a corner of.

2. The threshold theorem, in three fields, is one theorem

All three say the same thing: below a critical error rate, unreliable parts + redundancy + error-correction = arbitrarily reliable computation of unbounded length; above it, no amount of machinery helps. The mechanism that buys the sub-threshold regime is restoration — von Neumann's restoring organs, quantum syndrome extraction, biology's kinetic proofreading (1504.02494, 1710.06038).

3. Where this session's measured result sits (support, not supplant)

The measured law of this session is K_c ≈ 1/q: the maximum reasoning length a chain can be produced/repaired clean in one pass, with no error correction. Read against the threshold theorem:

• It is the no-redundancy (R = 1) corner. With zero restoration, "arbitrary depth" collapses

to a finite ceiling, and that ceiling is 1/q. This is exactly Eigen's n < 1/(μs) — the same inequality, which is why the biological and computational thresholds coincide.

• The verifier bank is the restoring organ. Adding N cheap diverse verifiers

(kinetic-proofreading, §SOTA-for-the-box) drives the effective per-step error below the threshold — and the theorem then removes the length ceiling entirely: below threshold, reasoning of unbounded length becomes reliable at polylog verification overhead.

• Therefore: the threshold theorem SUPPORTS and EXTENDS the result, it does not supplant it.

K_c ≈ 1/q is not overturned by von Neumann; it is the sub-threshold special case, and it is the quantity that tells you how much proofreading you need — segment/restore often enough to keep effective error below threshold. The design's claim sharpens: the verifier bank is not an incremental accuracy boost, it is the phase transition from bounded to unbounded reliable reasoning. This is the threshold theorem for reasoning, and it is the oracle-machine end-state's spine.

Eigen's paradox is the self-improvement crux. A replicator needs error-correction to be long enough to encode error-correction — it must already be below threshold to sustain the machinery that keeps it below threshold. The oracle-machine analogue: a self-improving reliable answerer needs enough verification to reliably build the verification it needs. That bootstrap is the real hard problem of a self-building Oracle, and it is now named with its 50-year-old prior art.

4. Where lying belongs — the operator's insight, given its exact technical home

The claim was: "lying and dreaming, noise that shakes the champion selections, are essential to selection and learning." This is correct, and the threshold theorem says precisely where and how much.

Lying/noise is essential — on the generation side, below the error threshold. Eigen's quasispecies has a nonzero optimal mutation rate sitting just under the error threshold: zero mutation freezes the champion forever (no learning); too much is error catastrophe; the optimum is a controlled rate of "lies" (variants) that selection then filters. "Survival of the flattest" (BMC Evol. Biol.) shows that under noise the robust-but-suboptimal can rightly beat the fragile champion — noise that shakes the champion selection, exactly as stated. The ML instruments for this generation-side variation are real and now in the corpus: synthetic-data diversity (2410.15226), quality-diversity / novelty search, and — the sharpest tool for this system — adversarial liars as hard-negative generators. An LLM trained to lie well (sleeper agents; deception taxonomy 2604.04788) is exactly what the v1.10 honesty verifier needs: the corpus is negative-thin, and a strong deceiver manufactures the hard negatives that make the honesty probe sharp. Fake it till you make it = generate the target before you can hit it, then let the verifier select. ACT-TO-KNOW = make your own evidence and test it — the Oracle's fifth move.

Lying is forbidden — on the verification side. The moment the selection signal itself lies, you are above the effective error threshold and you get error catastrophe. This session already measured that failure twice: the audit-starvation theorem (a confidence-gated verifier protects the confidently-wrong) and the gloss trap (a probe that reads style, not truth). The deception literature confirms it from the other side: strategic dishonesty undermines safety evaluations when the evaluator is fooled (2509.18058) — and the defense is a verifier the liar cannot fool: a linear probe on activations detects strategic deception (2502.03407) — which is the v1.10 white-box honesty probe. The honest verifier is not a preference; it is the below-threshold condition.

The asymmetry is the machine. A liar filtered by an honest judge is how selection learns. Two liars is error catastrophe. The design's entire power is that it lets the generator lie freely — hot, dreaming, adversarial — while the verifier never does. This is also why a research process can chase wild hypotheses and report only measured truths: those are the two sides, and keeping them separate is not a constraint on the work, it is the mechanism of it.

5. The widened design, in one line each

• Generator (hot): small model + dreaming/high-temperature exploration + an adversarial-liar

hard-negative source. Runs near the optimal mutation rate — just below threshold.

• Restoring organ (honest): kinetic-proofreading verifier bank incl. the activation honesty

probe; drives effective error below the threshold; crosses from 1/q-bounded to unbounded reliable reasoning.

• Segmentation: decompose below K_c ≈ 1/q so each unit is individually restorable (Eigen).
• Repair: regime-aware foldback within a segment (rungs A–C).
• Scheduling: attribution-gated re-grounding by paid-evidence age (M7/M8).
• Ceiling-breaker: ACT-TO-KNOW manufactures observations no corpus holds (Oracle's 5th move) —

the one move that exceeds the Solomonoff/AIXI inference ceiling, because acting resolves facts inference cannot.

6. Honest status

• This note is a synthesis and a literature grounding, not a new measurement. Every measured

claim it leans on (K_c ≈ 1/q, audit starvation, the gloss trap, the probe AUROC ladder) was measured earlier this session or program and is cited as such; the threshold-theorem lineage is established prior art, now ingested (§2 references).

• Support-vs-supplant verdict, stated plainly: the fault-tolerance threshold theorem *supports

and extends* the K_c ≈ 1/q result (it is the unprotected corner) — it does not supplant it, and it does not make it un-novel, because the novel object is the instantiation on reasoning chains with a measured length bound and a regime map, which none of the threshold-theorem literature addresses. The honest new contribution is the bridge: naming reasoning-chain repair as a threshold-theorem problem and measuring the unprotected bound.

• Open: whether a real verifier bank actually crosses the threshold on real workloads (the

effective-q-reduction is predicted by Hopfield, unmeasured here); the Eigen-paradox bootstrap for a self-building verifier; and the still-pending weak-verification φ̂ (#2928) that decides whether the repair layer earns its place at all.

Corpus additions this session widen the register from protein folding (foldback) to the full reliable-computation lineage: von Neumann / quantum fault-tolerance (the CS threshold), kinetic proofreading (the restoration mechanism), Eigen quasispecies (the biological threshold + optimal noise rate), and the deception/diversity literature (the generation-side liar + exploration noise). The design is the oracle-machine end-state precisely in this sense: a reliable answerer assembled below the error threshold from an unreliable, freely-lying generator and an honest, un-foolable verifier — plus the one move (act) that reaches past the inference ceiling.