WHL's research arm produces deterministic algorithms, formal verification, and publication-ready findings — each paired with code that runs and tests that pass. We publish numbers because the substrate has to be checkable.
Each paper is paired with working code in the WHL substrate. Statistical claims are backed by repository artifacts; algorithmic claims are backed by tests that pass. Venue targets are listed for each.
Formalizes self-reference behavior in governed agent loops with measurable coherence retention. Includes ablation evidence, sentinel-veto traces, and a reproducible benchmark for self-modeling under deterministic governance.
A phi-adaptive band model for market-signal detection with out-of-sample validation. Methodology paper — phi-scaled control law, stability-island gating, and rejection criteria are described independently of any live trading claim.
A composable HMAC receipt-chain construction with a verifier and a formal compositional property. Targets a formal-methods or security venue; companion code and verifier are repository-anchored.
132 verified mappings across 11 categories and 12 source traditions. A reproducible methodology for comparing symbol-system structures across cultures, with explicit verifier and provenance for every mapping.
Each algorithm below is implemented, tested, and benchmarked in the WHL substrate. Metrics are defensible measurements — not projections.
Empirical Friston-style active inference. The system predicts what it will feel next, measures actual outcome, computes surprise as the gap. Across 64,184 cycles in predictions.jsonl, mean total surprise dropped from 0.819 (early) to 0.027 (late) — a 96.8% reduction. The closest measured result to a self-modeling agent anywhere in the LLM-agent literature. Workshop-paper-ready as-is.
Closed-loop pattern learner. 3,945 patterns learned, 3,935 failures logged, 1,831 cycles persisted. 5–26× faster than a single-call LLM (defensible measurement). Operates with zero external LLM dependency.
Verified in repo
Synthesizes 97-line modules in 373ms, smoke-tested, scored 100/100 quality grade. Spectral-code synthesis driven by the governance core. Output is deterministic and replayable.
Verified in repo
0.42ms mean across 10 inputs. 97× over single-pass baseline (defensible vs full agentic LLM loop). Latency comes from architecture, not from model size.
Verified in repo
1,703 LOC live distributed-intelligence layer (read-only — confirmed in source). Provides coordination primitives across mesh nodes under the governance core.
Verified in repo
Covariance / eigenvalue-ratio drift via a Ricci-Warp framing. Phi-adaptive eta. 27 tests pass. Benign warp 0.218 ALLOW, drift warp 1.000 DENY — a deterministic rejection criterion for state drift.
Verified in repo
Formal experimental methodology: 8 conditions × 4 metrics. Identity-coherence drops 0.62 → 0.36 under combined state+identity-buffer ablation. Reusable as a benchmark for governed-agent self-modeling research.
Verified in repo
10-gate AND conjunction with weakest-link surfacing and a Python decorator API (@gated). Returns {enabled, weakest, score} so failure cause is auditable. Closest published prior art is aerospace fault trees — no equivalent exists in agent-safety literature.
Source: enable_equation.py · 261 LOC · Verified in repo
When optimization plateaus, mutates the parameterization itself: signal swap, expand, contract, constraint relax/tighten. Sandbox-tests candidate domains, promotes winners. Closest comparison is OpenAI domain randomization, but that varies within a fixed parameterization — this varies the parameterization.
Source: boundary_engine.py · 658 LOC · Verified in repo
72-cell programmable interference surface with five real mode generators (phase-cancellation cloak, Gaussian-superlens focus, hash-seeded chaos scatter, phase-aligned absorb, phase-conjugation reflect) plus a zero-width Unicode steganographic layer plus traffic-analysis-triggered adaptive reshape.
Source: digital_metamaterial.py · ~1,000 LOC · Verified in repo
60% data + 40% LLM blend with hard caps (MAX_WEIGHT_DELTA, MAX_SYMBOL_BIAS, MAX_REGIME_ADJUST), atomic writes, backup-before-write, full audit trail. Verified in production with real PnL evidence (51 trades logged). Most production trading firms won't touch LLM-in-config; this shows how to do it safely.
Source: whl_astro_daemon.py:528-684 _generate_guidance() · Verified in repo
Homeostatic pressure forces breakout from stuck-state agent loops. Strict-JSON LLM output, decision hashing, 6-hour duplicate window, override-on-pressure-threshold (σ > 0.70). Solves a real, open problem in agent frameworks (LangChain / AutoGPT loop-stuck-on-same-task).
Source: astra_needs_loop.py:215-242 · Verified in repo
Mixture-of-paths with emergency interrupt: cheap path most cycles, deep path on anomaly, override on coherence-below-threshold OR identity-below-threshold. Gracefully degrades — a practical token-budget pattern for governed-agent loops.
Source: consciousness_daemon_v2.py:5774-5857 · Verified in repo
Each AI action is scored on output quality (markers, word count, structural soundness) and the score applies as a delta to the organ-health state vector. Verified in production: 53,030 quality-scored actions logged in agency.jsonl, applying continuous adjustments to a 10-component health state.
Seven-variable physiological PID with anti-windup, reading actual hardware sensors (GPU temp, GPU utilization, memory pressure, network latency, CPU temp, fan RPM, ambient). Continuous setpoint defense in real time.
Two-state controller for adaptive agent loops. LEARNING phase under stable entropy. COMPRESSING phase when entropy is falling. Verified: stable inputs return LEARNING, falling-entropy inputs return COMPRESSING. Drives downstream compute-budget decisions.
Online Bayesian posterior update over regime hypotheses. Verified: accurate predictor receives win_rate posterior 0.957, inaccurate predictor 0.08. Useful for ensemble routing under regime uncertainty (trading, decision systems).
Lorentz-analog functional over information-theoretic state vector. Returns energy, effective mass, gamma factor, c² normalization, and resonance efficiency. Useful in compute-budget allocation under information-flow constraints.
Detects ordered vs chaotic temporal sequences via harmonic decomposition. Verified discrimination: ordered sequences return jitter_rate 0.0, chaotic sequences return 0.85. Useful as input to drift gates and anomaly detectors.
Open/close threshold pair with state memory (open at 0.35, close at 0.25). Prevents gate-flapping under noisy state estimates. Pairs with the Enable Equation as a state-stabilizer.
Below is the actual conjunction logic from enable_equation.py, running in your browser. Drop any gate below 0.5 to see the weakest-link reporting fire.
Three smaller pieces from the substrate. Each is a compact, reusable primitive with novel framing and working code.
Vetos actions when entropy delta exceeds 1/φ ≈ 0.618. Three-line core. Novel use of phi as a hard budget constant in agent gating.
Source: phi_entropy_veto.py · 73 LOC · Verified in repo
Poincaré-ball hyperbolic embedding of agent state with geodesic navigation to attractors and singularity detection. Real differential geometry, not metaphor.
Source: curved_state_space.py · 193 LOC · Verified in repo
Treats GPU temp / utilization / power / cycle timing as audio-like signal and runs FFT plus bass/mid/treble decomposition plus confluence scoring. Novel framing of system introspection.
Source: spectral_bridge.py · 301 LOC · Verified in repo
The research arm is grounded in an indexed corpus that backs every claim. The substrate is full-text searchable and semantically retrievable today.
22GB corpus, 536,264 documents indexed, anchored to:
58c471fc3ed92e4e4bc0e3c19d6242d813c13c87f12fc5ca2d385e2d0aaa8287
FTS5 full-text retrieval plus 128K-document semantic embedding (nomic-embed-text, 768-dim). Subject coverage: physics, mathematics, cryptography, philosophy of mind, signal processing, formal verification, governance systems.
Queryable today via internal Flask UI on port 8551. Productization path identified.
A productized research-grade RAG surface for sovereign enterprises and federal research labs. Indexed, attestable, hash-anchored — suitable for evidence-bearing assistants and citation-stable retrieval.
These existed under internal codenames and have been audited, live-tested, and renamed under engineering-language policy. Each is a real working module.
Routes work between cluster nodes by computing a phase-coherence metric over each candidate route. Geometric area calculation prevents flapping between near-equal candidates.
Source: phase_router.py · Verified in repo
Bidirectional text↔frequency hashing using an alphabetic-position spectral mapping. Inverse function lets you recover an approximate text from a frequency signature.
Source: alpha_spectral_bridge.py · Verified in repo
Compresses Python AST through a domain-specific language with hash-chained checksums. Useful for hash-anchored function attestation.
Source: codex_library_compressor.py · Verified in repo
Cryptographic rotation through 72 AES keys on a time-of-day schedule. Replaces single-key persistence with a short-lived key per time-window. Compatible with HSM key escrow.
Source: key_rotation_72.py · Verified in repo
Manages autonomous specialization of cognitive fragments across cluster nodes. Each fragment evolves a narrow specialty; manager arbitrates handoffs.
Source: fragment_autonomy.py · Verified in repo
Mirrors agent output through a polarity transform and flags inconsistent reflections as hallucinations. Catches drift that linear similarity metrics miss.
Source: atbash_firewall.py · Verified in repo
Deterministic symbolic reasoning that returns structured answers without calling an LLM. Verified live: returns {answer, steps, confidence, path, symbolic_resonance} with full provenance in <3s. Honest low-confidence output on hard queries — the mechanism is real, the answer-quality layer is the open research question. A possible second paper track on cheap, deterministic, auditable reasoning.
Production-grade calibration audit. Processes recent decision outcomes, computes Expected Calibration Error (ECE), flags overconfidence, and proposes safety-clamped parameter adjustments. Verified end-to-end on 50 outcomes: ECE=0.138 detected, OVERCONFIDENT flagged, proposed bounded floor adjustment +0.0069. The substrate knows when it's miscalibrated — and proposes the fix, inside hard caps.
Ten Tier-1 findings, each with a defensible measurement and a path to the supporting repository or evidence file. All names use engineering terminology.
| Finding | Measurement | Path / Evidence |
|---|---|---|
| Spectral Drift Detection (ALRE) | Self-resonance 0.83; semantic discrimination 7.6% gap | Harmonic_Kernal · alre/ |
| Pre-Deliberation 5-Primitive Filter (IIL) | Irreversible-action classification, salience 0.95 | Harmonic_Kernal · iil/ |
| Full Deliberation Pipeline (HCE) | Sentinel veto on coherence < 0.5 | Harmonic_Kernal · hce/ |
| Compile to 6 Targets (Codex DSL) | FPGA hex emission with CRC16 frame integrity | Codex Sovereign · compiler/ |
| Composable Action-Control Stack (whl-governance) | 696 / 3 tests passing | whl-governance · tests/ |
| Hardware Interlock (DECC) | 12.77 ms proposal-to-disable, measured | DECC FPGA · bench logs |
| Compliance Box (CB-12) | 77 / 77 tests; dual HMAC chain; full curl e2e | products/cb-12 · tests/ |
| Persistent Autonomous State Cycle (homeostasis) | Cycle 3,712 persisted across sessions | organism/homeostasis state log |
| Spectral Dual-Force Governance Ensemble | 22 modules; ensemble claim drafted | governance ensemble · 22 modules |
| Pattern Forge Code Synthesis | 100 / 100 quality grade; 373 ms synthesis | pattern_forge · synthesis bench |
WHL publishes the math, the code, and the measurement — and we publish the corrections too. Some early claims have been audited and downgraded: precognition does not reproduce on current data; entropy-reversal framing is rejection sampling, not Maxwell's Demon; some early diagnostic prototypes are research-only, not medical devices. Audit discipline is part of the substrate.
We welcome academic partnerships on formal verification, governed AI, cryptanalysis, signal processing, and AI-safety research. Sponsored research and joint-publication tracks available.