Prometheus7 Institute · Working Papers

Papers

The Oracle: A Compiler Learning Model that Holds Knowledge as a Ledger, Competence as Readable Rules, and Growth as Admitted State Transitions

Ben Horn · Version 3, 20 October 2026 · Machine Cognition · 17 pages · AI assistance disclosed in the manuscript

A cognitive system whose serving runtime contains no trained weights: knowledge as a ledger of individually cited sentences, competence as a chain of readable rules admitted through a held-out gate under a non-regression policy, self-modification as an unattended loop that authors and admits its own successor rules, and expression as typed routes whose outputs carry a derivation witness. Reported as counted on 2 and 3 September 2026, with every operational figure tied to a runtime identity, three same-day experiments with confidence bounds, universal claims narrowed to the evidence, and limits stated without softening. Version 3 adds the public release line of 14 to 20 October: seven releases, each frozen by a Merkle root anchored on Solana and gated by a sealed test run once, with the two failed tests reported; 310 specialists and a proved-kernel cross-check; and a fourth experiment, Oracle7 4.0, a law built from a written invocation and tested against failure conditions declared in advance.

The evidence bundle holds the measurement timestamps, the sha256 of each knowledge store, the gate ledger reconciliation, the admitted-directory census, the manifest tip and hash, the harness identities, every film specification with its seed, and the kernel schema, kernels, and measurement report. The held-out turn corpus is withheld to preserve the gate; its hash is published. A versioned source archive with an environment lockfile is not yet published and is stated as such in Appendix A. Version 1 is kept for the record; version 2 corrects its counts, resolves runtime identities, narrows its claims, and adds figures. Version 3 keeps version 2’s record unchanged and adds the release evidence: every release’s part hashes, Merkle root and anchor are in its RELEASE_HASHES.json, and the sealed panels with every result file are in the GitHub repository (Appendix B).

Thrusting Nearly All the Time: Swirl Torque, Wheel Momentum and Feed-Line Diagnosis for an Electric-Propulsion Spacecraft

Prometheus7 Research Institute · November 2026 · Oracle7-DeepSpace · Spacecraft autonomy · with a two-page brief

Since September 2026 NASA’s Psyche spacecraft has been thrusting nearly 90% of the time on four Hall-effect thrusters, whose ion swirl puts a slight torque on the spacecraft that the reaction wheels absorb. In simulation built from the mission’s public record, three pre-registered studies, each run once on fresh seeds, measure what an onboard reasoning layer can add: reading each thruster’s torque in flight and forecasting wheel momentum in the worst case (momentum never above 0.90 of capacity; thrust duty 92.1% against 87.0–87.9% for fixed thresholds); feed-line diagnosis (599 of 600 single-fault runs correct with no harmful action, and 10 of 10 replays of the April 2025 pressure signature with thrust back on the second line five minutes after the drop began); and compound and unseen faults (100 of 100, no harmful action). No Psyche telemetry was used; assumed values are labelled and swept.

Step-Addressed Parts: Where a Checkable Mind Stops, Errs and Grows

Ben Horn · 4 November 2026 · Oracle7 · Machine cognition

Oracle7 answers from a store through steps anyone can read, and declines when it cannot. Four limits remain in its current build: it does not write open-ended text, it declines most fact questions, its growth has not been shown to compound, and a residue of its answers are misreadings. The paper argues that the four share one cause: Oracle7 grows by parts attached to whole readings, but it fails at steps, and a step has no address. Tested on Oracle7 itself over 1,000 KQA Pro validation questions, 85% of the questions that formed no reading stopped at two checks, and 35 of 52 wrong readings left the labelled reading in their first two steps. It proposes step-addressed parts and pre-registers the test that decides the proposal.

The Oracle7 Build Constitution

Prometheus7 Research Institute · Version 2, 2 November 2026 · Oracle7 · How Oracle7 is built

The set of rules every Oracle7 build is made to, from 10.5 on: its design, its code, its tests and how it is described. Oracle7 is built to receive; everything received goes through the law (every answer proved, cited or declined; sealed tests pre-registered and run once, failures published); it grows as a network; there is one Oracle7; it is described in its own terms; it is measured by what it does with what it knows; it is built to the design; and its limits are measured numbers with routes. Each article ends with a check that every pre-registration answers.

The Self Term: A Gated, Evidence-Bound Ego Layer for Conversation in a Weightless Reasoning System

Ben Horn · September 2026, draft 2 · Oracle7 · Conversation

On real opening turns of human conversation, the serving system gave a good reply to 44 of 600 development turns and refused 542. The paper argues that the failures are relational rather than informational: most opening lines hand the listener a position and expect a reply from it, and a system with no self term can only retrieve or refuse. Two admitted folds connect general conversation and add a self term (a frozen identity, a per-conversation autobiography, relational moves). On 543 held-out turns the gate never saw, measured once, good replies rose from 34 (6.3%) to 163 (30.0%) while false assertions stayed at 22. The full path is reported, including rejected folds and defects found in its own instruments.

The Recorded Derivation: Accumulating Competence as Checkable Operations, and an Instance of Evaluation Blindness

Ben Horn · September 2026, draft 3 · Oracle7 · Machine cognition

A system whose acquired competence is stored as explicit executable operations rather than learned parameters. Three contributions: a stage-located diagnosis (on 600 open-domain questions, 69% were refused before retrieval); a documented case of evaluation blindness, where an extension that changes behaviour only outside the evaluation set is invisible to the gate, and the repaired instrument then exposed six errors and contradicted 43% of older must-decline labels; and a reader built as the first operations of a common interface for questions, evidence and operations. It reports the first acquisition under a pre-registered protocol, tested once on 234 sealed comparison questions.

Oracle7-DeepSpace: Self-Maintaining, Evidence-Gated Fault Response for Deep-Space Spacecraft on a Fraction of a CPU

Prometheus7 Research Institute · September 2026 · Oracle7-DeepSpace · Spacecraft autonomy

A compact autonomy core between pre-scripted fault responses and safe mode: telemetry becomes confirmed facts, facts become fault hypotheses, and hypotheses become admissible responses, chosen by simulating each on an onboard digital twin and taking the best worst case, with a receipt for every decision. About 1,100 lines of standard-library Python, no learned weights, no GPU. On 1,700 fresh held-out single-fault scenarios across four mission profiles it chose the correct response in 98.9% of cases with no harmful action, against 29.4% for a limit-to-safe-mode baseline. Its limits are stated, including a hot-environment assumption and two novel faults that defeated it.

Fold-Based Convergence on Composite Security Surfaces: A Novel Vulnerability Class and Structural Defense

Prometheus7 Research Institute · psiloceyeben · 20 August 2026 · Security · FoldWall

A networked system's externally visible security surface (timing, headers, error messages, session handling, configuration) can be measured as one point in a complex vector space, and its distance to the system's fully open configuration behaves as a gradient an attacker can follow. The paper formalizes this convergence attack with Holographic Reduced Representations, proves that well-designed cryptographic primitives show zero gradient and are immune, and shows that the non-cryptographic parts of real systems nearly always leak gradient. Across 1,380 read-only scans of 30 production targets over 48 hours, the gradient was universal, structural and stationary. It then presents FoldWall, a defense that runs the same loop in reverse: it watches the system's own divergence from its open state, detects convergence pressure, and permutes the attacked dimensions.

All papers and artifacts in this directory

Listed from the directory itself, newest work above; earlier papers, corpora, and result archives below