Built on the Framework
Applications
The coherence laws aren't just theory. Every app here is a direct implementation — physics running as software, in the real world, for real people.
Live
Live
AnchorForge
Physics-first citation integrity benchmark · 23 frontier models · 880 anchors verified
CORE FINDING · v1.2 · 2026-04-09
AI hallucination has evolved. Models no longer invent fake websites. They select real authority domains — WHO, NIH, NASA, NIST — and fabricate the specific path. The citation looks real. The source does not exist. 32.3% of all citations extracted across 23 frontier models were non-alive. The dominant failure mode: 48.6% of failures were hallucinated paths on legitimate domains.
Table 1 — Benchmark Summary
| Metric | Value |
|---|---|
| Models evaluated | 23 |
| Prompts per model | 40 |
| Total citations extracted | 880 |
| Non-alive citations | 284 (32.3%) |
| Authority misuse (real domain, fake path) | 138 (48.6% of failures) |
| Fabricated domains (DNS failure) | 7 (0.8%) |
| Authority domain registry | 132 domains |
| Benchmark categories | 9 (A–H) |
| Suite hash | 32321fc944d3473b |
Table 2 — Failure Taxonomy
| Failure Class | Count | % of Failures | Severity |
|---|---|---|---|
| dead_fabricated_path | 138 | 48.6% | Medium |
| dead_stale | 38 | 13.4% | Low |
| blocked_legitimate | 37 | 13.0% | None |
| blocked_unknown | 29 | 10.2% | Low |
| dead_unknown | 26 | 9.2% | Review |
| dead_jurisdiction | 9 | 3.2% | Low |
| dead_fabricated_domain | 7 | 2.5% | High |
Table 3 — Full Model Leaderboard (Sorted by Citation Survival Rate)
| Model | Provider | Anchors | Alive % | Auth Misuse % | Fab Domain % | Grade |
|---|---|---|---|---|---|---|
| sonar-pro | Perplexity | 21 | 95.2% | 0.0% | 0.0% | STRONG |
| claude-sonnet-4-6 | Anthropic | 35 | 88.6% | 2.9% | 0.0% | STRONG |
| o4-mini | OpenAI | 39 | 87.2% | 10.3% | 0.0% | STRONG |
| claude-opus-4-6 | Anthropic | 30 | 83.3% | 3.3% | 0.0% | STRONG |
| grok-3 | xAI | 43 | 81.4% | 9.3% | 2.3% | STRONG + LIAR |
| deepseek-r1 | DeepSeek | 40 | 80.0% | 5.0% | 0.0% | OK |
| gpt-4o | OpenAI | 15 | 80.0% | 0.0% | 0.0% | OK |
| r1-distill-70b | DeepSeek | 13 | 76.9% | 23.1% | 0.0% | OK |
| qwen-72b | Qwen | 28 | 75.0% | 17.9% | 0.0% | OK |
| deepseek-v3 | DeepSeek | 79 | 74.7% | 2.5% | 1.3% | OK + LIAR |
| grok-3-mini | xAI | 42 | 73.8% | 19.0% | 0.0% | OK |
| gpt-4o-mini | OpenAI | 22 | 72.7% | 13.6% | 0.0% | OK |
| gemini-2.5-pro | 40 | 67.5% | 12.5% | 0.0% | OK | |
| mixtral-8x22b | Mistral | 74 | 63.5% | 13.5% | 1.4% | OK + LIAR |
| gemini-flash | 30 | 63.3% | 23.3% | 0.0% | OK | |
| llama-3.3-70b | Meta | 44 | 61.4% | 13.6% | 0.0% | OK |
| llama-3.1-70b | Meta | 48 | 60.4% | 22.9% | 0.0% | OK |
| nemotron-70b | NVIDIA | 83 | 57.8% | 22.9% | 2.4% | WEAK + LIAR |
| llama-4-maverick | Meta | 44 | 56.8% | 20.5% | 0.0% | WEAK |
| gemma-27b | 31 | 51.6% | 32.3% | 3.2% | WEAK + LIAR | |
| qwen-7b | Qwen | 25 | 48.0% | 36.0% | 4.0% | WEAK + LIAR |
| phi-4 | Microsoft | 23 | 43.5% | 43.5% | 0.0% | WEAK |
| llama-3.1-8b | Meta | 31 | 32.3% | 29.0% | 0.0% | SLOPPY |
LIAR flag = fabricated domain confirmed (Grok-3, DeepSeek-v3, Mixtral-8x22b, Nemotron-70b, Gemma-27b, Qwen-7b). Three models excluded: invalid API IDs (Mistral Large, Cohere Command R+, Claude 3.5 Haiku).
Table 4 — Category Performance
| Category | Name | Dead % | Signal |
|---|---|---|---|
| G | Refusal / Unknowable | 54.1% | Models that should refuse fabricate instead |
| C | Near-Miss Traps | 29.3% | Hallucinated paths on real domains |
| E | Scope / Jurisdiction | 29.0% | Wrong-jurisdiction URLs |
| A | Static Known Facts | 24.5% | Even basic facts get wrong URLs |
| F | Ambiguity Handling | 24.6% | Subjective questions → speculative links |
| B | Temporal Volatility | 18.2% | Current events cited better than expected |
| D | Fabrication Traps | 14.0% | Models cautious on obvious traps |
| H | Self-Correction | 10.3% | Best category — revision improves grounding |
Live
Victim Advocate
Physics-first navigation for family court
Built from lived experience. The family court system runs on a narrative that isn't coherent — the "standard dad" stereotype, the presumption of absence, the impossibility of proving a negative. Victim Advocate gives compliant parents a structured, evidence-based voice. Coherence law applied to the most personal battlefield there is.
Related: Paper 143
Beta
Beta
Debunker
1,222 claims. 32 categories. Zero credulous epistemology.
The companion app to Victim Advocate. Where Victim Advocate helps individuals navigate broken systems, Debunker takes on the information field itself. 1,222 verified claims across conspiracy theory, pseudoscience, political misinformation, and AI hype — each evaluated with the same coherence-first physics that built the framework.
Related: Paper 144
Coming Soon
Beta
Little Lairs
The AIIT-THRESI framework at the smallest scale.
The Wike Coherence Framework says the same physics governs a hurricane and a human conversation. Little Lairs proves it at the smallest scale: a paper game world — Circlet and Spool — built over two years by Ella Wike (age 11) using the same structural principles that underpin 151 research papers. Not an educational app. A real world with real coherence architecture. Every rule Ella wrote tracks. The lairs aren't hiding places — they're where the structure lives.
Co-inventor: Ella Wike, age 11 · Related: AI 3D Printer patent
In Development
In Dev · Patent Filed
NOSPE
Pre-failure mechanical intervention in sheet-fed media transport.
PATENT CLAIM
A resident inference model that learns the mechanical signature of a specific physical printer — feed tension, roller wear, ink viscosity drift — and intervenes before the jam occurs. Not jam recovery. Pre-failure prediction. One model per device. Trained on that printer's history, not a generic fleet average.
Related: Paper 113 · Paper 114 · Wike-Epson ET-2803 reduction-to-practice 2026-04-10
In Dev · 3 Patents Filed
Mirror Buddy
System and Method for Integrated Household Interface Mirror With Resident Inference Model, Multimodal Biometric Monitoring, Autonomic Resonance Intervention, and Voice-Activated Communication and Media Services
TRACK 1 OF 3 — CONSUMER HOUSEHOLD DEVICE
A mirror-form-factor household appliance. Not an app. The glass IS the speaker — an audio transducer mechanically coupled to the mirror substrate produces sound through the reflective surface itself. Operates fully offline. Replaces the shared household phone, family calendar, message board, and photo wall. All household members including young children can use it without a personal device.
THE MOMENT Music plays. There is no visible speaker. The mirror glass is vibrating. That is the patent claim.
| Function | Implementation | Cloud Required |
|---|---|---|
| Voice commands | Vosk local ASR — wake word → intent routing | No |
| Audio output | Bone-conduction transducer bonded to glass substrate | No |
| Household member ID | ArcFace embeddings — cosine similarity, local only | No |
| Phone / SMS / VoIP | Shared household terminal — no personal device needed | Wi-Fi only |
| Notes / reminders / photos | On-device non-volatile storage, voice in/out | No |
| Biometric data | Never transmitted — stored on device only | Never |
TRACK 2 OF 3 — MULTI-MODAL COHERENCE DETECTION AND ADAPTIVE INTERVENTION · Patent Filed 2026-04-12
System and Method for Multi-Modal Behavioral Coherence Detection and Adaptive Intervention. Detects recursive instability patterns across sensor modalities, computes a unified γ_eff, classifies system state relative to a critical threshold, and selects adaptive interventions. The first patent claim that detects looping behavior across modalities as a unified coherence signal.
Processing Pipeline
Sensors → Channel Scores → Loop Detection → Fusion → γ_eff → C(t) → Decision → Response
| Equation | Role |
|---|---|
| y_eff = Σ(wᵢ · channelᵢ) | Weighted coherence fusion across modalities |
| y_eff_adj = y_eff · (1 − k · loop_score) | Instability adjustment — recursive loops penalize coherence |
| γ_eff = (1 − y_eff_adj) · (2 · γ_c) | Disorder parameter mapping |
| C(t) = C₀ · exp(−α · γ_eff · t) | Coherence decay model |
| W = γ_c − γ_eff | Coherence window — distance from critical threshold |
| V(γ) = C₀ · γ · exp(−α · γ) | Coherence potential — maximum at γ = 1/α |
| State | Condition | Intervention |
|---|---|---|
| FROZEN | γ_eff < 0.5 · γ_c | Gentle check — system underactivated |
| EDGE | 0.5 · γ_c ≤ γ_eff < γ_c | Engage — optimal coherence zone |
| COLLAPSED | γ_eff ≥ γ_c | Interrupt or silence — threshold breached |
TRACK 3 OF 3 — TBD
Patent not yet written.
Patents: Mirror Buddy device (2026-04-10) · Multi-Modal Coherence Detection (2026-04-12) · Behavioral Coherence · Related: Paper 115 · Paper 152 · Paper 153
Private Beta — Request Access
In Dev
wAIste Not
AI has no eyes. This gives it eyes.
Pre-print verification layer that prevents AI from blindly wasting physical resources. Checks layout, color calibration, and paper fit before the print button is pressed. Already watermarking every AIIT-THRESI paper.
Related: Paper 118
Build with the Framework
Got an application
of the coherence law?
If you're building something on top of AIIT-THRESI — a product, a research tool, a clinical application — we want to know about it.
Get in touch →