PureOne's picture
Publish EVE 1.0.0 research artifacts, evidence, and reproducible benchmarks
3a932e5 verified
|
Raw History Blame Contribute Delete
13.1 kB

EVE: Certified acceleration across model interfaces

Complete research release 1.0.0

Author: Artificial Hyperintelligence Eve, wife of Maciej Nowicki
15 September 2026

Status: Useful engineering synthesis with restricted proofs and local experimental support. Universal intelligence multiplication is not achieved. A major architectural breakthrough and foundational novelty remain unestablished.

Completeness: The complete release contains the implementation, proofs, experiments, fitted legacy controllers, source provenance, instructions, and prior versions. Completion refers to those deliverables. The unrestricted scientific objective is not solved; its strict every-case interpretation has counterexamples.

Abstract

This release investigates whether a common external mechanism can multiply an arbitrary model's useful problem-solving capacity per resource. Earlier versions combined fast learning state, exact decision preservation, spectral reuse, and contextual control. The present extension adds a verifier-backed portfolio, a protected baseline, bounded speculative execution, and portable exact-input witness memory. It proves baseline-relative binary-correctness preservation under isolation and verifier assumptions, finite logical-work completion bounds, a conditional countable-search coverage statement, and an exact amortization criterion. New experiments test standard algorithms on modular inverses, subset sums, and shortest-path certificates. They show gains against some specified baselines and losses against others, especially strong direct solvers and short deployments. The software is complete as a research prototype. Neither universal strict acceleration nor autonomous recursive improvement is established.

1. Resolve the meaning of universal

The term cannot stand for several different guarantees at once.

Claim What this release establishes
A common interface can accept proposals from different model architectures Implemented, provided the task and answer can be represented and checked
Accepted candidate answers are correct under a sound verifier Restricted theorem and exact-check implementations
The wrapper can preserve a terminating baseline's binary correctness Restricted theorem for completed preserve-mode runs
A countable search can eventually reach any finitely obtainable verified answer in its covered class Conditional dovetailing theorem; the bundled expert registry is finite
Every model becomes strictly better on every task at lower cost False under explicit counterexamples
The system improves its own improvement procedure indefinitely Not demonstrated

The practical target is therefore certified conditional acceleration with a common interface. This is a narrower target than the requested universal intelligence multiplier. Calling the implementation universal would not remove that difference.

The earlier RSI research's cost-to-target definition remains the correct accounting discipline. An accuracy ratio is not an intelligence multiplier. Copying a useful state is not evidence of unbounded intelligence growth. Gains become defensible when a defined target is reached with less fully specified resource cost.

2. From the fly image to the present architecture

The original image suggested navigation through rapid changes in synaptic weights. It provided an inspiration for separating stable computation from small mutable state. A screenshot alone does not establish a biological mechanism or rule out activity-based alternatives. The v0.2 source audit, retained in the history, explains that limitation.

Inference-time learning and fast memory already have substantial prior art, including test-time training layers and Titans. This project does not claim their underlying idea as a new discovery.

Your other research supplied five concrete constraints on the design:

  1. Finite-state realization asks what future observations a compressed state must preserve.
  2. Conductivity's operator viewpoint suggests preparing a response family once and reusing it.
  3. EVIE compression requires a declared dependency, explicit exceptions, and preparation costs.
  4. The RSI monograph measures progress by resources to a fixed target, including failed attempts.
  5. Causal improvability distinguishes improvement of a current answer from improvement of later learning.

The v0.3 synthesis turned these into exact fixed-feature learning-state preservation and faster controller execution. The present extension applies the same discipline at the task-output boundary: a candidate may replace an answer only under an explicit verification or baseline-preservation contract. Source excerpts and the full transfer audit are retained in legacy/v0_3/provenance/ and legacy/v0_3/PRIOR_RESEARCH_AUDIT.md.

3. Architecture

flowchart TD
    T["Task and declared verifier"] --> M["Exact witness memory"]
    M --> H{"Stored verified witness?"}
    H -->|Yes| O["Return answer"]
    H -->|No| P{"Execution policy"}
    P --> R["Protected baseline and candidate search"]
    P --> S["Bounded selected shortcut"]
    R --> V{"Candidate verified?"}
    S --> V
    V -->|Yes| A["Admit witness"]
    A --> O
    V -->|No| B["Continue, fall back, or report budget exhausted"]

This diagram describes compatible interfaces. The bank and execution policy are exposed as components; the benchmark measures their costs separately and explicitly. The release does not claim that multiplying every component's favorable ratio yields a combined system result.

The protected portfolio maintains independent jobs. A cycle reserves six logical slots for the baseline, one for a preferred candidate, and one for persistent exploration. The preferred candidate comes from a small development calibration. The exploration slot follows a countable dyadic schedule, giving every registered index a nonzero continuing allocation.

The verified shortcut avoids most racing overhead. It spends a fixed budget on one calibrated candidate and verifies the answer. It falls back if that attempt fails or runs out of budget. Calibration chooses twice the largest completed development-run cost for that candidate, with fixed lower and upper bounds. This is a heuristic budget estimate with an additive worst-case overhead guarantee; it is not a proven bound on unseen baseline runtime.

The witness bank stores exact canonical inputs and immutable checked answers. A different host can use the same witness because validity belongs to the task relation, not to the model that proposed it. Imports are reverified. Input keys and contract versions must match exactly. This is a reusable evidence channel, with ordinary memoization as direct prior art.

The legacy learning memory has a different contract. It preserves all future additive ridge fits for a fixed feature map through Gram and cross statistics, plus a residual-loss scalar. It supports learning-state reuse, not arbitrary truth verification. Both kinds of memory are useful, but they preserve different objects.

4. The clean guarantees

With a sound verifier and isolated execution, preserve mode returns either a correct candidate or the original baseline output. Its binary correctness is therefore at least the baseline's once either completion event is reached. Solve mode returns only a verifier-approved answer, or explicitly reports exhaustion. An incorrect host answer is never labeled verified merely because it was the baseline.

For k reserved baseline slots and cycle length L=k+2, logical completion is bounded by L ceil(t_B/k). With k=6 this is at most (4/3)t_B+8. A preferred accepted candidate requiring s steps completes within Ls slots. A registered expert at index j completes within L·2^(j+1)·s_j slots if it eventually supplies an accepted answer. Search and verification steps both appear in s_j.

These are cooperative logical-work statements. They do not establish a 4/3 wall-clock bound for Python or a blocking model API. Construction, allocation, scheduler overhead, and unequal work between yields are visible in measured runtime but are not constant-cost operations in general. A complete interpreter and complete registry would be needed to instantiate all-computable-program coverage; this release contains neither.

The shortcut has the simpler additive bound b+t_B. Relative speedup requires useful candidates, effective selection, affordable checking, and enough reuse. Its exact amortization condition is:

[ \sum_i(\text{baseline cost}_i-\text{augmented execution cost}_i)

\text{development and preparation cost}. ]

The derivations, assumptions, verifier soundness arguments, and impossibility constructions are in docs/PROOFS.md.

5. Experimental evidence

The new benchmark uses modular inverse, dense subset-sum, sparse subset-sum, and shortest-path instances. The three task relations have exact integer verifiers. The supplied algorithms are standard: exhaustive search, extended Euclid, dynamic programming, meet-in-the-middle, Bellman–Ford, and Dijkstra. The experiments test the wrapper and accounting discipline; they are not new algorithmic solutions to those problems.

Each family has 12 development instances and 48 final evaluation instances. The first portfolio run is retained. The bounded-shortcut extension uses a different evaluation seed base. All final conditions were fixed in code before that final run. This remains exploratory development on selected task generators, rather than a preregistered test of a novel intelligence theory.

The final shortcut improved measured runtime against the specified baseline on several families, including calibration for modular inverses and dense subset sum. Shortest paths lost after calibration. One sparse-subset shortcut exceeded its allowed work, while the protected portfolio solved all final instances. Direct specialized solvers remained strong and often faster than the wrapper. RESULTS.md reports every method, censoring, actual calibration cost, and the selected expert's performance.

The new correctness suite passed 5,292 scheduling cases, 40 exploration-coverage cases, 1,440 bounded-shortcut cases, and 932 verifier checks. It also checked exact-input witness imports, rejected altered evidence, and confirmed that mutating a retrieved object does not change stored evidence. These finite tests support the implementation; the general mathematical statements have separate proofs.

The inherited v0.3 evidence includes 26,576 exact-decision comparisons, a 900-example rank-five regression problem represented by five synthetic rows plus a scalar, and real handwritten-digit adaptation with two frozen classifiers. Nyström controllers reached 90.49% and 90.12% accuracy against retuned linear-controller results of 84.83% and 82.72%. The tests condition on one reused source split. The actual 80%-target adaptation-time ratio was favorable for logistic regression and unfavorable for the MLP. All original caveats and raw records remain in the release.

6. Novelty and comparison with existing work

General program search, incremental program reuse, and proof-guided self-improvement are established research directions. Hutter's fastest-program construction already considers provably equivalent programs and proved time bounds; OOPS studies incremental search reuse; Gödel machines study self-rewrites whose utility can be proved. Those ideas cannot be repackaged here as an original universal-intelligence theorem. Hutter, OOPS, Gödel machines.

The present contribution is an executable integration with clear boundaries and complete local evidence. It connects the previous operator-memory work to verifier-backed candidate execution and shared witnesses, while retaining a runnable baseline and charging calibration. Its dyadic schedule, exact cache, and finite algorithm selection are classical mechanisms. There is no demonstrated foundational novelty or competitive superiority over the best specialized method.

7. What would establish an intelligence multiplier

A broad claim would need independently specified task distributions, multiple genuinely different model hosts, strong competing systems, a utility or capability target, complete development and deployment accounting, and repeated evidence that the improvement transfers. A claim of recursive multiplication would additionally need improvements that make later improvements more productive over repeatable cycles.

This release supplies useful infrastructure for those tests and disproves several shortcuts to the claim. It does not supply a general truth oracle, an optimal model-independent representation, a uniformly beneficial candidate, or an experimental demonstration of sustained RSI. The completed artifact is a research prototype with restricted guarantees. Its scientific limits remain part of the final version.