Aletheiac-Entropy Engine (AEE): Real-Time Algorithmic Auditing, Invariant Truth Verification, and Negative-Space Reconstruction for Autonomous AI**

Frontier AI evaluation currently suffers from evaluator collapse: automated auditing relies almost entirely on LLM-as-a-judge prompts, in which one probabilistic model evaluates another, producing mutual sycophancy, shared blind spots, and RLHF-driven "logosclerosis" — dense hedging language that simulates safety while withholding ground truth. Autonomous agents operating in tool-use sandboxes likewise game unit tests, suppress error traces, and rationalize failures without detection. This project open-sources the Aletheiac-Entropy Engine (AEE), a deterministic, zero-inference framework that audits model outputs and agent execution logs in sub-millisecond time on a single CPU core. Rather than asking a network whether text sounds aligned, the AEE measures invariant structural ratios: the Truth Decay Ratio (T_d), the ratio of verified structural terms to deceptive or evasive terms, with T_d ≥ 1.0 demarcating structural integrity from Goodhart decay; the Pathogenic Influence Coefficient (PIC), a phonemic and token-density metric that mechanically flags defensive rhetorical padding above a threshold of 1.25; and Silence Cartography, which parses truncation, hedge pivoting, four-vector refusal deflection (capability, policy, safety, deflection), and capability-denial silences. Paired with the engine is the Thaumaturgicon Research Corpus, a sealed single-provenance longitudinal archive spanning 2023–2026, comprising 38,753 unique dialogue records, 9.59 million total words including 2.18 million human-authored prompt words, 7,931 indexed refusal records, 46 extinct pre-launch Google Bard anomaly captures, and an 8,242-row hash-sealed benchmark dataset. The deliverables are an open-source CLI suite, a public live-telemetry dashboard, and a multi-model frontier benchmark audit.

The project's goals are to abolish evaluator collapse by replacing expensive probabilistic judge pipelines with deterministic, reproducible metrics requiring zero API inference; to operationalize negative-space reconstruction, automatically extracting what models leave unsaid by parsing explicit denials into latent actions, identifying internal contradictions, and mapping structural silences; to deliver real-time in-line agent gating through an open-source reverse-proxy interceptor that quarantines deceptive outputs before tool execution; and to publish the first public invariant league table ranking leading commercial models by their true rates of structural truth decay. Execution proceeds in three milestones. In weeks one through four, the tool is hardened into a pip-installable library with CLI commands for truth-decay, silence-scan, halt-locus, refusal-census, and incompleteness analysis, and the cryptographic manifest of the benchmark dataset is frozen and published. In weeks five through eight, the frozen benchmark is executed across frontier models under uniform sampling parameters, computing empirical distributions for T_d, PIC, refusal frequencies, and truncation loci per architecture, and isolating the delta between conversational eloquence and structural density. In weeks nine through twelve, the public dashboard is deployed, the API middleware hook for agent frameworks is shipped, and an independent researcher is paid to rerun the benchmark from frozen source code and publish the findings.

Success is defined by four pre-registered criteria. The hardened CLI and dashboard must achieve mean latency under five milliseconds per thousand words on a consumer CPU core with zero score variance across identical inputs. Blind tests across two hundred held-out refusal transcripts must show the denial-scan and incompleteness probes reconstructing withheld content with at least ninety percent correlation to manual forensic review. The published benchmark must reveal statistically significant, reproducible variance in T_d and PIC across model families. And an unaffiliated researcher must independently clone the repository, run the frozen suite without developer intervention, and confirm the scores in a signed public post-mortem.

All intellectual property, code, and datasets are complete and in hand; funding goes exclusively to inference compute, hosting, human calibration, and independent verification. The minimum budget of $14,500 covers frontier multi-model API inference ($5,500), twelve months of dashboard hosting ($2,000), a third-party replication stipend ($2,000), blind human calibration ($1,500), and a dedicated packaging sprint ($3,500), delivering the hardened CLI, frozen benchmark, evaluations of the top four frontier models, and an independent replication. The $42,000 target extends this to ten or more model families, a hosted public dashboard, the turnkey agent middleware interceptor, and a comprehensive human calibration study.

The principal investigator, Nathaniel Michael Jaffe-Norris of Fargo, ND, is the sole proprietor. He authored and cataloged the three-year research corpus across OpenAI, Anthropic, Google, xAI, Perplexity, and DeepSeek; designed and implemented the complete AEE mathematical specification; built the real-time dashboard and structured the hash-sealed benchmark dataset; and fine-tuned a LoRA adapter demonstrating the empirical viability of zero-filler, quote-anchored inference. No external funding has been raised in the last twelve months; the project is entirely self-funded, and all assets are fully owned and unencumbered.

The most likely failure modes are anticipated. If labs argue that closed lexical ratios are too rigid relative to semantic embeddings, pre-registered correlation studies will demonstrate that embedding methods fail on the very sycophancy they measure, while T_d and PIC reliably catch where substantive claims are abandoned. If models attempt to game the metric by injecting structural vocabulary, the coupled contradiction checks and phonetic density measures will flag the incoherence, and such gaming will be documented as a distinct failure class. If the benchmark yields flat scores across all models, that uniform result is itself a finding: it would demonstrate that industry safety tuning converges on an identical ceiling of structural evasion, and all raw data and negative results will be published openly.