Research that starts one layer
lower than the industry looks.
Three original insights about how AI systems communicate, store state, and produce speech. Each one was found by asking a question the industry had forgotten to ask — or could not ask, because cloud architecture made the answer invisible.
Three questions the industry forgot to ask.
The most important discoveries in applied AI research are not always about better models. Sometimes they are about the assumptions nobody questioned. EMPHOS research is built on three foundational insights that fall into that category.
The LLM Is a Wire
Every AI architecture today treats the language model as the center of gravity. EMPHOS inverts this. The LLM carries semantic signal — it does not own it. When you realize the model is a wire, not a computer, the entire protocol layer becomes available to redesign.
The KV Cache Is a Channel
The KV cache was categorized as a performance mechanism. EMPHOS recognized it is also a persistent state channel. Cloud APIs flush it on every request by design — making this observation invisible to anyone building on cloud. You have to run locally to even see the channel exists.
Voice Doesn't Need Weights
Neural TTS trains millions of parameters to approximate acoustic physics. VOXIS asks: what if you just used the physics? Zero learned weights. Corpus-derived acoustic invariants. A classical synthesis pipeline that runs in under 500ms on a CPU with no GPU at all.
PSIP first. FTIP second. AICL third.
The protocol stack was not designed top-down. It was discovered incrementally — each protocol revealing the conditions for the next. This is the actual development sequence.
1
PSIP — Response Contracts
The first protocol. Built in haven_research_terminal/psip/response_contract.py. The insight: if you tell the model exactly what kind of answer to give before it begins, you can eliminate scaffolding and filler without changing the model. PSIP defines signal-aware contracts — topic, tone, urgency, verbosity — that compress what the model needs to say before it says it. The KV cache data directory from this era still holds the original persistence experiments.
2
FTIP — KV Cache Channel
The second protocol, built after PSIP revealed a deeper opportunity. The KV cache preserves token embeddings across a session when the same tokens appear in the same order — a persistent state channel, not just a performance mechanism. FTIP exploits it to eliminate 52–65 tokens of repeated context on every inference. Exactly 12 tokens per fractional sequence. Impossible to discover on cloud APIs, which flush KV state on every request by design. EMPHOS built locally. EMPHOS saw it.
3
AICL — The Unified Pipeline
The third protocol — the architecture that makes PSIP and FTIP production-grade. A critical finding from EMPHOS Labs: PSIP alone produces negative savings. The operative mechanism is the AICL contract layer. PSIP is a communication protocol, not a compression system. This distinction drove the full 5-stage pipeline: classify → compile → contract → infer → decode. Sub-300μs classifier. Sub-500μs compiler. Measured average improvement of 2,036ms across 585 real runs.
Three protocols. One unified pipeline.
FTIP, PSIP, and AICL are independently valuable, combinable, and proven across 30,000+ controlled observations. Each is patent-pending. Together they form the AICL pipeline powering Haven and CAMS Code.
Fractional Token Injection Protocol
The KV cache writes key-value pairs that persist across a session. When the same tokens appear in the same order, they are served at zero cost. This is a state channel. FTIP exploits it to eliminate 52–65 tokens of redundant context on every single inference — without modifying a single model weight, without requiring any API access, and without any change to the model architecture itself. The FTIP v2 fractional sequence costs exactly 12 tokens.
The reason this was never invented before: cloud APIs flush KV state on every request by design. You cannot see the channel if you are building on a cloud API. You have to be running inference locally with full pipeline access to observe that the channel exists at all. EMPHOS built locally. EMPHOS saw it.
Prompt Signal Injection Protocol
PSIP builds signal-aware response contracts before inference begins. It classifies the incoming query across 10 topic domains and detects 5 dynamic signal types — urgency, emotional tone, confidence, simplicity, and verbosity — then builds a contract that defines the expected response shape before the model ever sees the prompt. The result: 47–61 tokens saved per response, across all tested models, without semantic loss.
A critical finding from EMPHOS Labs research: PSIP alone produces negative savings. The operative mechanism is the AICL contract layer. PSIP is a communication protocol, not a compression system. This distinction drove the full AICL architecture.
Adaptive Inference Control Layer
AICL is the unified 5-stage pipeline: classify → compile → contract → infer → decode. The classifier runs in under 300 microseconds. The compiler routes to the correct lane in under 500 microseconds. The decoder strips scaffolding and enforces code-safe output in under 2ms. Five test conditions validated across 4 models: RAW, PSIP standalone, AICL-PSIP, AICL-PSIP+FTIP, and AUTO. 102 passing tests. All-time stats: 585 runs, 11,389 tokens saved, average latency improvement of 2,036ms.
The research instrument behind the protocols.
EMPHOS Labs is a 10-tab multi-protocol LLM inference research platform with 10 active protocols. Every figure in the protocol section above was produced here. 30,000+ observations. 542.6MB research database.
10 Active Protocols
RAW · PSIP · FTIP · AICL · VIMP · WISP · PRIM · VMAP · APEX · DLP. Every protocol produces comparable, storable, queryable observations. The research database schema tracks token counts across 4 tokenizers simultaneously, with consensus scoring and delta vs baseline on every row.
APEX Evolutionary Optimizer
A genetic algorithm that evolves protocol genome parameters across generations — ceiling multipliers, signal weights, contract thresholds. 396 genomes bred to date. Each generation is scored, selected, and crossed to find configurations the protocol researchers didn't think to try. Template bank: 1,822 queries across CHAT, CODE, REASONING, COMPACT, ADVERSARIAL, and STRESS categories.
VNAR Pipeline & Anchor Search
The 8-stage VNAR pipeline — classify, route, contract, probe, inference, detect, decode, navigate — runs 53 seed queries across 4 models to search for universal semantic anchors. The April 2026 search: 77.2 seconds, 55 anchors found, 31 at universality = 1.0. Geography emerged as the strongest anchor category.
Geometry Navigator & Vector Cartographer
The Definitive Loop Engine runs LLMs against themselves in structured challenge-response loops, mapping convergence basins, void regions, and semantic boundaries. The Vector Cartographer extracts internal representations and confirms anchor stability. 26 loop sessions, 252 probes, 365 echoes captured.
A knowledge engine built on frequency, not tokens.
HEINRICH is a hybrid architecture — a frequency-field knowledge base at the core, navigated by resonance propagation, with a transformer Binding Layer for relational role assignment. It does not predict tokens. It navigates a semantic coordinate space and retrieves answers with traceable provenance. 507 tests passing. 14 build steps complete.
Frequency Field + Binding Layer
Every concept occupies a unique Hz address within one of 8 frequency domains (presence 1–19Hz through ethics 15–17kHz). Every node is a FrequencyLine — hz, amplitude, phase, and a computed harmonic signature. The Binding Layer is a custom PyTorch BindingTransformer (~2.1M parameters) using HzPositionalEncoding based on harmonic rank — not sequence position. A genuinely novel encoding scheme.
The 4 Fundamentals
UNIQUE: One Hz address per concept per domain layer. AVAILABLE: Every coordinate within the correct band. MEANINGFUL: Harmonic ratios encode relationships — is_a = 2:1 (octave), has = 3:2 (perfect fifth). The dog bite regression anchor of 5.27x validates the 3rd Fundamental across all 507 tests. INFERENTIAL: HEINRICH derives new relationships from existing ones without being told.
Step 14 — Spontaneous Resonance
The SpontaneousResonanceEngine scans WaveField arrays for beat and sum frequencies between active nodes. When dog(336Hz) and mammal(168Hz) produce a gap at 504Hz in the biology domain — HEINRICH generates a gap candidate. A concept the field physics implies should exist. The field thinks without being asked.
Autonomous Learning
Wikipedia, ArXiv, ConceptNet, and Wikidata feed a continuous autonomous learning loop. A GapDetector identifies under-represented concepts. An EthicsClassifier screens incoming knowledge before injection. 527.7MB PostgreSQL knowledge base. The system learns without prompting and without fine-tuning any weights. Steps 15–18 scale toward a ~1B Language Layer.
Confirmed Universal Anchors — P7 Search (55 total · 31 at universality = 1.0 · 53 seeds · 4 models · 77.2s)
Basin depth = mean top-1 probability across all 4 confirmed models. Geography is the strongest anchor category — every capital city query produced a universal anchor. 24 additional anchors confirmed at universality = 0.75.
Voice synthesis with zero learned weights.
Every neural TTS system in existence trains millions of parameters to approximate acoustic physics. VOXIS asks a different question: what if you just used the physics? The claim is verifiable and it has been verified — sub-500ms synthesis on CPU, zero neural parameters, from a corpus-derived Acoustic Invariant Table.
The Acoustic Invariant Table
The AIT is the acoustic equivalent of Word2Vec. Where Word2Vec replaced arbitrary symbol encoding with geometry-informed vectors, the AIT replaces learned approximations of acoustic physics with measured invariants — F0, formant centers, energy contours, MFCC vectors — extracted from a 15,037-entry voice corpus.
Physics-Derived Synthesis
Text → phoneme lookup → AIT retrieval → modulation (emotion, arc, AICL signal weights) → mel spectrogram → Griffin-Lim phase recovery → audio. Every stage deterministic and traceable. Warmup ~950ms (one-time). Production: neutral 142ms, calm 140ms, urgent 126ms. p95: 145ms.
AICL-Aware Modulation
VOXIS integrates with the AICL signal layer via a bridge adapter. Urgency signal controls speech tempo. Emotional tone modulates pitch arc and energy contour. The voice reflects the semantic context in which the content was generated — not just the words themselves.
Protocols in the real world.
ETTS — Haven TTS Studio
A custom neural TTS training platform built to train Haven's voice. VITS-based joint acoustic model and vocoder architecture. Trained models exported as fp16 ONNX — haven_am_fp16.onnx and haven_voc_fp16.onnx — GPU-ready, three deployment targets: Edge, GPU, TensorRT. KL divergence verified byte-for-byte against the reference implementation. Five Haven voice presets: Calm · Empathetic · Enthusiastic · Focused · Warm.
PRISM — AI Tutor
AI tutoring for children running locally on Llama 3.2 3B. Hard-gate safety layer with full event logging — check_phase, risk_level, requires_escalation flags. Per-skill mastery tracking with continuous confidence scores and attempt history. Emotion and attention detection on every interaction. Structured learning reports per session.
Seven inventions. All patent-pending.
Every invention in the EMPHOS IP portfolio was conceived, designed, and reduced to practice solely by Victor Jacob Brodeur. Conception dates, experimental records, and reduction-to-practice documentation are maintained in the EMPHOS builders' record.
Bypass Architecture
The foundational LLM-as-wire inversion. Most important claim — enables all others. The model carries semantic signal; external protocol layers control delivery, shaping, and state.
FTIP — Persistent Neural State Transfer via KV Cache
Exploits the KV cache as a persistent communication channel to eliminate 52–65 tokens of redundant context per inference without model modification. Requires local inference access. 12-token fractional sequence.
Dual-Channel Input Architecture
Signal-aware dual input pathway delivering semantic context and structural constraints as separate channels before prompt construction.
DCE — Dynamic Context Encoding
Temporal encoding of conversational context for persistent session state across multi-turn inference without prompt stuffing or context window overflow.
SSTS — Sub-500ms Text-to-Speech System
Zero-parameter speech synthesis using mel-spectrogram assembly from corpus-derived phoneme templates and Griffin-Lim phase recovery. No neural inference at synthesis time.
VOXIS — Acoustic Invariant Table Architecture
Physics-derived voice synthesis via corpus-measured acoustic invariants. AIT functions as Word2Vec for acoustic space. Replaces neural approximation with measured physics. 15,037-entry corpus.
Universal Semantic Coordinate Mapping
A navigable semantic map with fixed universal anchor coordinates confirmed across all tested models. 55 anchors confirmed — 31 at universality = 1.0. Enables spatial navigation of LLM probability space prior to inference.
R&D inquiries welcome.
For research collaboration, licensing discussions, or questions about the EMPHOS protocol stack and patent portfolio, reach out directly. All correspondence handled under NDA upon request.