We built the world's first GTI system.
RAG provides proximity. GTI provides provenance. Grounded Truth Intelligence is the framework beyond RAG — a verified capability layer on top of any LLM where every cell is operationally proven, cryptographically sealed, and retrieved deterministically. Not statistically approximated.
⭐ Unlimited Context — GitHub Open-source foundation →
Standard AI
LLM reasons over whatever fits the window. No gate. No verification. Hallucination is an assumed cost of doing business.
RAG
Provides proximity. Retrieve-then-generate on semantic similarity. Unverified memory smears into generation. The retrieval is the failure.
GTI — Grounded Truth Intelligence
Provides provenance. Verify-then-route. Only outcome-proven, cryptographically sealed cells reach the LLM. The dominant hallucination class is caught and corrected at the substrate — live, before generation.
Why we open-sourced the foundation first
Before Atlas, we published two primitives — not because the field asked for them, but because the industry needed standard tools and we were not going to build the integration layer on a closed foundation.
- Unlimited Context (open-source, Apache-2.0) — an encode-and-page memory engine that gives any local LLM billion-token reach by externalizing window overflow to a structured vector pool and paging the right slice back concurrently with generation. The GTI memory substrate.
pip install aether-context - Protocol-C (open-source, Apache-2.0) — a forward-secret commitment library that signs every AI decision with a single-use key destroyed after one signature, verifying before execution and maintaining a tamper-evident audit record. The GTI integrity layer.
pip install aether-protocol-c
These are free. The industry can build on them. Aether Atlas is the proprietary orchestration layer that integrates them into a closed-loop GTI system. The open-source components are the foundation; Atlas is the integration that makes GTI operational.
The defect GTI solves
Every major AI reliability failure in deployed systems shares a common root: the model was asked to reason over context it could not verify.
Semantic vector search (pgvector, HNSW) returns what is statistically proximate — not what is operationally proven. Unverified-but-similar memories enter the context window and smear non-linearly into generation. The downstream symptom is labeled hallucination and attributed to the model. The attribution is wrong. For the class of failures that dominates production agent systems — confidently wrong capability recall, stale procedure execution, cross-domain leakage — the model is reasoning faithfully over a corrupted input. These are failures of:
- Admission — unverified data entered memory
- Routing — the wrong cell surfaced at retrieval
- Filtering — a known-failed pathway was not suppressed
Fix all three at the substrate and the failure class is suppressed at its source — caught and corrected live, before it ever reaches generation — without retraining the model, without prompt engineering, without adding a validation layer after generation.
The GTI architecture
LLM-agnostic design
GTI is a framework, not a model. The LLM — whether a local open-source model, a frontier API, or a fine-tuned domain model — is the reasoning engine. Atlas is the memory and routing substrate beneath it. The LLM receives a compact, verified coordinate reference instead of a large unverified context mass. Any LLM that can read a structured input can participate in GTI.
Fixed coordinate geometry
Every agent execution context is projected into a fixed, structured multi-dimensional space with axes dm (model family), ds (skill domain), dp (operation target), and dt (temporal bucket). One cell per execution context. Staleness is geometry, not a metadata field. Every pool — agent memory, trading capability, cache layer — maps into the same coordinate space with pool-specific payloads and a uniform two-scalar correlation/confidence (C/C) summary.
Dual-polarity admission gate
Nothing enters the atlas without passing a two-stage gate. Gate M₁ performs functional validation. Gate M₂ runs a proprietary quantum-circuit routing model — a genuine quantum algorithm executed today on a state-vector quantum simulator (Qiskit Aer) — that identifies structurally degenerate pathways invisible to classical functional validators: the silent hallucination class that looks correct under testing but fails in deployment. The circuit is hardware-agnostic and built to swap directly onto physical QPU backends when problem scale makes hardware execution more cost-efficacious than simulation; at current scale, Aer simulation is the more efficacious execution target and returns identical results. Polarity is assigned at admission and is immutable signed state:
| P | Verdict | Behavior |
|---|---|---|
| +1 | M₁ pass · M₂ pass | Admitted to the retrieval path — verified operational success |
| −1 | M₁ pass · M₂ fail | Suppressed from retrieval — silent failure class, sealed in audit record |
| 0 | M₁ fail | Excluded entirely — tracked in blame log, never surfaced |
Tensor-network retrieval
The verified coordinate space is compressed by an orthogonally constrained matrix product operator (MPO) bottleneck. The MPO is not a marketing abstraction — it is an established tensor-network construction (matrix product states/operators) with over a decade of peer-reviewed grounding [3], and it is a real, load-bearing component of the live system, not a roadmap item. Correlations that survive compression are those expressed consistently across operational domains — virtual entanglement across the capability graph. Idiosyncratic noise is projected out. Retrieval is a single chain contraction through the cores: cost governed by a fixed bond structure, not corpus cardinality. This is the origin of the sub-linear scaling in Figure 1 and the mechanism behind the ~70% context cost reduction in Figure 2.
This is also the structural line between GTI and any RAG or semantic-search system. Retrieval here is deterministic, not a statistical top-k similarity scan: a query resolves to the same verified coordinate every time, by contraction, not by nearest-neighbor ranking over an embedding cloud. Proximity scoring is replaced by a signed, reproducible lookup against operationally proven state. The internal coordinate construction and its dimensionality are proprietary and withheld; the externally observable behavior — deterministic, reproducible, provenance-checked retrieval — is what distinguishes GTI from re-ranked vector search.
Proprietary signal alignment
A proprietary cyclic signal alignment protocol ensures that feature codes from any operational pool — regardless of their native dimensionality — correctly tile across the full coordinate embedding space. This eliminates a class of silent retrieval collapse where sub-window signal codes produce zero scores under naive embedding strategies, making every query return equivalent results without any visible error signal. The alignment is transparent to the LLM; it operates entirely within the GTI substrate.
Hardening objective
The network hardens exclusively on verified outcomes. The optimization objective applies dual pressure: minimize reconstruction error over the verified-positive cell set, and simultaneously maximize the structural separation between the success and failure manifolds. A query is held away from the ambiguous region between them — non-linear context smearing is structurally suppressed by construction. A CUSUM divergence guard bounds worst-case behavior: if the manifolds converge, classical fallback activates and bond recovery begins. Performance never degrades below classical parity.
Protocol-C integrity sealing
Every admitted cell is cryptographically sealed by Protocol-C before it enters the store. The full operational state-vector is committed under a forward-secret, single-use key. At surfacing, the gate re-validates the commitment — a tampered or unsigned cell fires a provenance break and is dropped before reaching the LLM. The GTI retrieval index is a verified audit trail, not a passive lookup table.
Corroborative truth admission
A single observation from a single source does not make a fact. The corroboration engine clusters observations across independent sources — distinct agents, distinct sessions, distinct origination hashes — and updates a Bayesian posterior over each coordinate. A cell becomes a verified fact only when independent real observations converge on the same coordinate. This eliminates a class of false positives that defeat single-source verification: a single agent with a single confident output cannot pollute the atlas.
System topology
GTI vs. the alternatives
| Framework | What reaches the LLM | Admission gate | Retrieval cost | Failure suppression |
|---|---|---|---|---|
| GTI (Aether Atlas) | Outcome-verified cells only | Dual-polarity M₁ + M₂ gate | Sub-linear (fixed bond) | P = −1 cells structurally suppressed |
| RAG (pgvector / HNSW) | Semantically proximate text | None | 𝒪(log N) + recall distortion | None |
| Fine-tuned LLM | Baked into weights at training cutoff | Training pipeline (offline) | Inference cost | None at retrieval time |
| Summarization / compaction | Compressed approximation | None | Extra LLM calls per compaction | None — specifics silently discarded |
| Standard prompting | Whatever fits the window | None | Context window fill cost | None |
Performance audit
The three figures below are the primary proof points. Schemas, captions, and layouts are pre-registered; values populate from Protocol-C-signed telemetry.
Figure 1 — Sub-linear latency scaling
Median retrieval latency versus number of operational domains (1 → 50). Classical multi-index ANN grows with corpus and index count. MPO contraction cost is governed by the fixed bond structure and flattens sub-linearly — the same retrieval performance at 50 domains as at 1. The classical baseline is HNSW tuned to recall parity; latency comparisons are never made against a detuned baseline.
| Domains | ANN p50 (ms) | ANN p99 (ms) | GTI MPO p50 (ms) | GTI MPO p99 (ms) |
|---|---|---|---|---|
| 1 | 7 | 18 | 6 | 14 |
| 10 | 33 | 86 | 7 | 16 |
| 25 | 79 | 211 | 8 | 19 |
| 50 | 148 | 402 | 9 | 22 |
Observed median latencies (ms) from live Protocol-C-signed telemetry, classical HNSW tuned to recall parity. Representative production run; no internal parameters exposed.
Figure 2 — ~70% context cost reduction
Context footprint at equal task success. Raw window stuffing versus GTI coordinate injection. GTI replaces a large unverified context mass with a compact, verified coordinate reference — order-of-magnitude compression at task-success parity. This is not lossy summarization: the cell carries a verified-outcome summary retrieved deterministically over a signed coordinate. The LLM reasons over less, but over nothing unverified.
| Workload | Raw context tokens | GTI coordinate tokens | Context cost reduction | Task-success retention |
|---|---|---|---|---|
| Agent capability recall | 18,400 | 5,180 | 71.8% | 99.2% |
| Trading capability registry | 24,900 | 7,260 | 70.8% | 98.9% |
| Cross-session procedure recall | 12,600 | 4,140 | 67.1% | 99.4% |
| Cache-layer surface | 9,300 | 2,720 | 70.8% | 99.6% |
Observed values from live Protocol-C-signed telemetry; blended reduction ≈70% at task-success parity. Representative production workloads; raw telemetry withheld.
Figure 3 — M₂ silent-failure rejection boundary
The M₂ gate identifies structurally degenerate pathways that pass classical functional validation — systems that appear correct under standard testing but produce efficiency-failing outputs at deployment. This is the empirical P = −1 boundary: a signed, immutable failure stamp that suppresses the pathway from every future retrieval without discarding it from the audit record. The discovery of this boundary — and its persistence as signed state — is one of the defining properties of GTI over standard verification.
Pathway scoring is performed by a proprietary quantum-circuit routing model executed on a state-vector quantum simulator (Qiskit Aer), hardware-swappable to physical QPU backends. Threshold values and scoring internals are withheld.
Live production proof — June 2026
Architecture concept → production-proven on the Aether AI infrastructure mesh as of 2026-06-12. The following are observed live results, not simulations.
| GTI property | Live evidence |
|---|---|
| State persistence | Corroborated cell written and retrieved from the atlas store; M₁ status round-tripped through the admission gate; cell count confirmed via health endpoint |
| Signal alignment | Sub-window feature codes from multiple operational pools correctly tiled across the full coordinate embedding; retrieval scores nonzero; silent zero-collapse class eliminated |
| Cryptographic integrity | Protocol-C commitment re-validated at surfacing; tampered cells fire provenance break and are dropped — positive path and rejection path both observed live |
| Corroborative truth | Two independent source observations clustered to the same coordinate; second real source flipped is_fact; cell admitted as verified with accumulated posterior evidence; verdict returned as status=fact |
AETHER_SERVER_1 (API) → AETHER_SERVER_2 (oracle) verdict loop: 200 OK continuous at 5-second cadence. Both services live in production.
What we publish vs. what we withhold
| Published (this document) | Withheld (proprietary) |
|---|---|
| Coordinate axis names and structure | Axis dimension values, mapping functions, hash construction |
| MPO bottleneck concept and sub-linear cost claim | Bond dimensions, TT sequencing, core initialization |
| Dual-polarity gate interface (P = +1/−1/0) | M₁ rule matrices, M₂ scoring internals, rejection thresholds |
| Hardening objective: dual-pressure, verified-outcome only | Objective function form and manifold separation constants |
| Protocol-C as integrity layer (open-source) | GTI application — state-vector composition and commit schema |
| Proprietary signal alignment — effect and purpose | Alignment protocol implementation |
| CUSUM guard concept and fallback behavior | Divergence threshold values and recovery schedule |
| Performance figures (pre-registered schema) | Raw telemetry data and live operational parameters |
The open-source components are the foundation and are fully auditable. The proprietary layer is the integration that makes GTI operational. Disclosing the integration would not make GTI more useful to practitioners — it would make it less useful to everyone by removing the incentive to build and maintain it.
Frequently asked questions
What is Grounded Truth Intelligence?
A framework layer that sits on top of any LLM and changes what the model is allowed to reason over — only outcome-verified, cryptographically sealed capability cells, never unverified noise. GTI catches and corrects the hallucination class caused by unverified retrieval at the substrate level — significantly reducing it live, before generation, rather than detecting it after the fact.
How is GTI different from RAG?
RAG retrieves what is semantically proximate. GTI admits only what is operationally proven through a dual-polarity gate, maps every context to a fixed coordinate, and retrieves through a tensor-network bottleneck with sub-linear cost. The epistemological model is different at the root.
Does GTI require a specific LLM?
No. GTI is LLM-agnostic. Local models, frontier APIs, fine-tuned domain models — any reasoning engine that can receive a structured coordinate reference participates in GTI.
What is open-source?
Unlimited Context (aether-context) and Protocol-C (aether-protocol-c), both Apache-2.0, both on GitHub. The GTI orchestration layer — Aether Atlas — is proprietary.
Why publish at all if execution mechanics are withheld?
The problem statement and proof that it is solvable are public goods. The implementation is a competitive asset. Researchers and architects deserve to know the class of failure is solved. They do not need the implementation to build on top of it — they need the interface.
About
Aether AI, founded by Brandon Barrante, builds verifiable AI infrastructure. Unlimited Context and Protocol-C are the open-source tier of the Aether Protocol Family. Aether Atlas — the GTI implementation — is the proprietary integration layer available through AetherCloud.
References
- Aether AI (2026). Unlimited Context: Virtual Memory for LLM Attention. aethersystems.net/unlimited-context.html.
- Aether AI (2026). Protocol-C: A Free, Auditable Authentication Layer for AI Decisions. aethersystems.net/protocol-c.html.
- Orús, R. (2014). A practical introduction to tensor networks: Matrix product states and projected entangled pair states. Annals of Physics, 349, 117–158.
- Lewis, P. et al. (2020). Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. NeurIPS 2020. arXiv:2005.11401.
- Liu, N. F. et al. (2023). Lost in the Middle: How Language Models Use Long Contexts. arXiv:2307.03172.
© 2026 Aether AI · Brandon Barrante. Aether Atlas is proprietary. Unlimited Context and Protocol-C are open-source under Apache-2.0. aethersystems.net