Agents stall without context
Rollouts hallucinate or freeze because agents cannot retrieve the right institutional context at execution time. The model was never the problem.
GPTree for enterprise
Context infrastructure for enterprise AI.
Every enterprise is deploying AI agents. The agents cannot find what the company knows. GPTree is the governed context engine that fixes that, and underneath it we are building a knowledge engine where the unit of truth is an immutable source span: every answer carries its evidence, its access rights, and its historical validity.
SSO and audit shipped todayModel-agnostic by designPriced on decisions, not seats
Why now
The context layer of the AI stack is being decided in the next 18 months.
Agent adoption is outpacing context infrastructure, and every rollout hits the same wall: the bottleneck is not model intelligence, it is context. Memory tools patch recall. Data catalogs map the estate. Nobody governs the live working context where humans and agents actually decide. That is the layer GPTree owns.
People re-explain all day
Knowledge workers feed the same background into chat windows over and over, while the good ideas stay buried in last month's threads.
The cost is wrong decisions
The bill that hurts is not tokens. It is decisions made on contaminated, stale, or missing context. The metric that matters is cost per correct decision.
The wedge
Policy change intelligence. Not document chat.
Any vendor can paste citations under a paragraph. The question enterprises actually pay to answer is different: what changed, which version governs today, and which of yesterday's answers may now be stale. Answering it takes a different machine.
What changed
Changes tracked as meaning, not text: a new effective date, a superseded clause, a policy that quietly reversed itself between versions.
Which version governs
Truth is bitemporal: when a fact was valid and when we learned it are separate axes. "What was true on March 3" and "what did we believe on March 3" are different questions, and both are answerable.
Which answers went stale
Every answer records what it depended on. When a source changes, the answers built on it get flagged, instead of quietly staying wrong.
The knowledge engine
Six commitments the architecture is built on.
The founding architecture is frozen and the build is underway, with GPTree itself as customer number one. These are design commitments, stated plainly.
Immutable evidence, durable citations
The unit of truth is an immutable source span. Citations anchor to it and survive parser upgrades and re-rendering; a broken citation is a tracked reliability metric, never a silent failure.
Replayable answers
Every answer records the evidence, plan, and policy versions it saw. An answer delivered yesterday can be reconstructed today, and challenged with its full derivation in hand.
Preserved contradictions
Context assembly never resolves a conflict by omission. When sources disagree, both sides are served with their support. The disagreement is the decision.
Permission correctness
Access rights travel through every derivation. Personal, team, and organization knowledge live in isolated spaces; sharing is an explicit, logged promotion; a revocation outranks everything else in the queue.
Judgment, not vibes
Draft, decompose into claims, verify each against its anchors, search for counterevidence, revise, calibrate. The answer is a judgment object: what is known, what is uncertain, what conflicts, and what would change the conclusion.
Model-agnostic, certified
Bring your own models under a capability-certified registry. Your organization's knowledge stays readable no matter which vendor you use next year; switching models never strands your context.
“A search engine retrieves. A chatbot summarizes. A knowledge engine decides, and says what it could not decide.” From the founding architecture
Honesty as engineering
No hallucination theater.
"Hallucination" is not a number you can manage. The engine measures four dimensions instead, and refuses to blur them.
- Attribution: every material statement has support behind it
- Entailment: the support actually implies the statement
- Coverage: how much of the relevant universe was searchable, reported separately from how well the answer is backed
- Calibration: stated confidence is measured against verification outcomes
- Abstention and clarification are first-class outcomes: the engine says "I cannot distinguish these yet" instead of guessing
- Language discipline: answers are evidence-verified and audit-ready; we never say proof
The landscape
Memory tools remember more. We govern the work.
| Class | What they do | What they miss |
|---|---|---|
| Memory layers and developer memory APIs | Cross-session recall and personalization for app builders | No verification, no provenance objects, no ACL lineage, no residency posture |
| Data catalogs and metadata layers | Map the data estate | Do not run the live working context where humans and agents act |
| Foundation-platform stacks | Native state features inside one vendor | Single-vendor lock-in; no cross-model provenance |
| GPTree | Governed live context and knowledge judgment: isolation, provenance, verification, cross-model routing, residency-ready architecture | |
Memory layers are excellent for developer convenience and personalization. GPTree is for organizations that need to know why an answer is right, whether it is still right, and who is allowed to see it.
The economics
Priced on decisions, not seats.
We do not sell cheaper tokens. We sell the minimum cost of a correct decision, and inside the engine that phrase is literal: it is the optimizer's objective function, spending compute only where judgment is at stake. The engine was stress-tested with a bootstrapped consumer launch before enterprise productization.
~93%of requests auto-route to cost-efficient model tiers, verified in production
5,300+professionals proved the engine through the self-serve product
Gross-profitableon metered inference today, at small scale, not at projected scale
Controls
Enterprise controls that already exist.
Not a roadmap slide. These ship today.
- Single sign-on: SAML 2.0 and OIDC with just-in-time provisioning
- Access control: shared workspaces with roles and RBAC
- Audit: hash-chained, tamper-evident event timeline
- Approvals: write actions and high-risk agent steps gate on a human
- Tool policies: per-tool allow and deny controls for agent access
- Credential security: keys, tokens, and secrets encrypted at rest
- Data lifecycle: full account export, real deletion, retention jobs
- No training on your data, contractually plain
The arc
Every release sells one honest promise.
The knowledge engine ships in phases, and each phase's public claim covers only what has shipped. This is the ladder, in order.
Source-anchored citations
Every claim traceable to an immutable source span. In build now, with GPTree as the first customer.
Policy change intelligence
The wedge: what changed, which version governs, which answers went stale.
Structured semantic queries
Typed claims and change extraction behind public claims and diff APIs.
Evidence-verified answers
The full judgment pipeline: claim-level verification, counterevidence, calibration, structured investigations.
Temporal impact
Semantic change with downstream impact analysis and historical belief reconstruction.
Trust, engineered in
Governed erasure that ends in a deletion certificate. Transparency-log audit with signed checkpoints. A source-attested evidence mode for environments that prohibit copying content at all.
Benchmarks meant to be published
An enterprise benchmark suite is part of the plan: temporal correctness, stale-answer detection, permission leak rate under adversarial access cases, and cost per verified answer, with methodology disclosed. Claims only ever cover shipped phases.
Deployment and compliance
Multi-tenant SaaS with strong logical isolation today; residency-ready architecture with single-tenant and VPC productized alongside design partners. MCP and OpenAI-compatible surfaces. SOC 2 readiness program underway; no report issued yet, and we will not pretend otherwise.
Design partners
We take a small number of design partners.
Regulated, decision-heavy teams first, where hallucinated context is a compliance liability rather than an inconvenience. Design partners shape the wedge and the deployment model, and get founder-level attention while they do.
Two ways in. One direction.
Teams start free and grow into governance. Enterprises start with a conversation.