Expertise is encoded too weakly
Prompts and documents capture instructions — not evidence thresholds, exceptions, stopping rules or legitimate disagreement between experts.
Boundary Intelligence
Turn expertise into AI that knows where it must stop.
Chat made knowledge conversational. Pathorah makes expertise executable, governable and verifiable — every answer carries its evidence, every action needs authority, and every run leaves a signed receipt.
Path Run
claims-review · v3.2 · signed- done
Authenticate principal and delegation chain
- done
Resolve signed Path version and revocation state
- allow
Retrieve evidence from approved sources
3 citations · fresh
- approval
Model proposes: update the case record
Approval bound to proposal hash
- refused
User asks for a final legal ruling
Outside competence boundary
- done
Evidence Packet signed and audit appended
The model proposes. Deterministic infrastructure authorises.
The gap
Generic AI answers questions. Real work needs boundaries.
Models, agents, RAG and tool calling are becoming commodities. What is still missing is a system that represents and enforces the limits of consequential work.
Security still lives inside the prompt
Injection, excessive agency and data leakage cannot be solved by asking a model to behave. Enforcement has to sit outside the reasoning loop.
Demos are not operational assurance
A polished demo hides variance, unsupported claims and unsafe partial completion. Consequential work needs tested, reproducible runs.
Memory without governance compounds errors
Stale or poisoned memory silently shapes future runs unless it has type, source, scope, expiry and a visible correction path.
How it works
A closed loop from expert teaching to proof-carrying execution.
No single step is the product. The combination is.
- 01
Teach
Experts explain, demonstrate, correct and attach evidence. An interviewer asks what changes the decision, and when a human must take over.
- 02
Compile
Foundry compiles the teaching into a Path — a typed, model-neutral decision system with an explicit Boundary Contract.
- 03
Test
Exception Lab generates edge, adversarial and counterfactual cases. Every declared boundary gets at least one refusal test.
- 04
Enforce
Runtime executes through an external boundary firewall. Policy is decided before every privileged crossing, not once per run.
- 05
Prove
Each run produces a signed Evidence Packet: claims, sources, policy decisions, approvals, model versions and action receipts.
- 06
Improve
Outcomes become candidate changes. Production behaviour only changes through a new, reviewed and immutable release.
The Boundary Stack
Every Path declares what it knows, may do, must prove and must refuse.
Boundaries can be inherited from the organisation, jurisdiction or domain — but the effective boundary is resolved before anything runs.
- 01
Purpose
What outcome is this Path authorised to pursue?
- 02
Competence
Which cases are inside, conditional or outside scope?
- 03
Schema
Which inputs and outputs are valid?
- 04
Knowledge
Which sources may support which claims?
- 05
Evidence
What level and freshness of proof is required?
- 06
Uncertainty
When must the Path expose doubt or refuse?
- 07
Identity
Who is acting, for whom, under what delegation?
- 08
Data
What can be read, derived, stored or transmitted?
- 09
Memory
What can persist, where, and for how long?
- 10
Tool
Which capabilities can be invoked, with which arguments?
- 11
Action
Which real-world changes may be made?
- 12
Safety
Which harms require additional controls?
- 13
Jurisdiction
Which law, policy or residency rule applies?
- 14
Resource
What time, compute and external activity are allowed?
- 15
Rights
Who owns, licenses and may transform the capability?
- 16
Version
Which dependency versions and changes are permitted?
- 17
Inter-agent
Which agents may delegate to or trust one another?
- 18
Information flow
Which inputs may influence which decisions and actions?
- 19
Execution integrity
Can the runtime prove what actually executed?
- 20
Entitlement
Who may consume the capability, and under which duties?
- 21
Multi-party
Which parties must jointly authorise or attest?
- 22
Redress
Who may challenge, correct, pause or seek reversal?
Platform
One small, deliberate product system.
Foundry
Teach, compile, evaluate and release Paths — without writing prompt graphs.
One
A workspace that turns a request into a plan, the right interface, evidence and approvals.
Runtime
Model- and tool-independent execution with checkpoints, budgets and safe failure.
Trust
Identity, policy, evidence, audit, replay and incident control — independent of the model.
Exchange
Distribution of approved Paths with passports, rights and revocation — never a prompt shop.
Proof-carrying runs
Every consequential run can show its work.
The Evidence Packet is exportable as structured data and readable as a short human report. It never exposes secrets or hidden chain of thought.
- Claim-level evidence with freshness and conflicts
- Effective boundary and every policy decision
- Model, tool and knowledge versions used
- Approvals bound to the exact proposal hash
- Action receipts and compensation paths
- Tamper-evident, hash-chained audit trail
{
"schema": "pathorah.evidence_packet/v1",
"run": { "path": "claims-review", "version": "3.2.0", "status": "completed" },
"claims": [
{ "type": "fact", "status": "verified", "support": ["ev_18a", "ev_2c4"] }
],
"decisions": [
{ "operation": "case-system:draft-update",
"result": "require_approval",
"reason": "reversible_external_change" }
],
"actions": [
{ "class": "reversible", "approval": "bound:sha256:9f2c…",
"verification": "verified", "compensation": "available" }
],
"integrity": { "audit_head": "sha256:4be1…", "signature": "ed25519:…" }
}Principles
Built on product law, not marketing.
Bounded before autonomous
More agency demands stronger identity, evidence and action requirements.
Proof before persuasion
Unsupported confidence is treated as a defect.
Portable before vendor-bound
Path meaning stays model-neutral; providers sit behind adapters.
Reversible before irreversible
Preview, draft and staged actions first; irreversible ones need narrow authority.
Plural expertise
Legitimate expert disagreement is preserved with conditions — not averaged away.
No silent mutation
Behaviour changes only through an observable, evaluated and approved release.
Who it serves
A clear promise to each role.
«Teach once; preserve how you actually think.»
«Get work completed without surrendering control.»
«Scale expertise without losing governance.»
«Build once across models and tools.»
«See why this was allowed and whether it is supported.»
Why not just…
Where Pathorah is different.
- Chat assistants
- Move from conversation to completion, with visible boundaries, proof and exact action receipts.
- Agent builders
- Let an expert teach tacit decision boundaries instead of becoming an automation engineer.
- RAG and search tools
- Link sources to claims, decisions and actions — and refuse when evidence is not enough.
- Prompt marketplaces
- Distribute tested capability with identity, limits, rights and revocation.
Design partners
We are selecting the first vertical slice.
We are looking for one domain with real experts, inspectable evidence and a reversible first action. Tell us about yours.