Skip to content

jonathan-kellerai/trust-boundary-protocol

TRUST-BOUNDARY-PROTOCOL — Layered Invariant Mediating Epistemic Norms

A unified epistemic boundary layer specification — a confidence-state mediator that lets cooperating systems gate cross-domain actions on a shared view of every participant's epistemic readiness.

Specification only. No implementation in this repository.


TRUST-BOUNDARY-PROTOCOL is a design specification for a boundary layer that sits between independently-developed systems and mediates whether a cross-domain action may proceed. It does not move data, transform content, or run business logic. It answers one question, fail-closed: given every participating domain's current epistemic state, is this action admissible right now?

The specification is v0.1-draft. It is published as a design artifact — to be read, cited, and critiqued — not as a released library.


1. What TRUST-BOUNDARY-PROTOCOL is

TRUST-BOUNDARY-PROTOCOL is an epistemic boundary layer: a confidence-state router, not a content filter.

It mediates the epistemic readiness of a domain to act — whether that domain currently holds enough well-founded confidence to be allowed into a cross-domain action — and it does not inspect, validate, or transform the content of the action itself.

The distinction matters. A content filter asks "is this payload acceptable?" TRUST-BOUNDARY-PROTOCOL asks "is the system proposing this action in a state where it should be trusted to act at all?" Those are orthogonal concerns, and TRUST-BOUNDARY-PROTOCOL owns only the second.

Every decision TRUST-BOUNDARY-PROTOCOL makes is ternaryADMITTED, DENIED, or INDETERMINATE — and every caller is required to treat INDETERMINATE identically to DENIED. There is no fourth state, no retry-into-admission, no default-allow.

2. Why this exists — the convergent-evolution finding

TRUST-BOUNDARY-PROTOCOL was not designed top-down. It is the formalization of a pattern that was observed.

Nine independently-developed systems — built by different efforts, at different times, for different purposes — were each analyzed for how they decide whether to act under uncertainty. All nine had independently converged on the same abstract shape:

evidence → accumulate → compare_to_threshold → gate_action

Each system had reinvented some portion of: an append-only evidence log, a confidence score, a threshold, a fail-closed default, and a policy gate. None of them shared infrastructure. Each maintained its own version, with its own bugs, its own audit gaps, and no cross-system visibility.

TRUST-BOUNDARY-PROTOCOL is the unification of that convergence: one epistemically-aware boundary layer that the nine systems — and systems like them — can share instead of re-deriving. The empirical evidence behind this claim is in docs/pattern-report/trust-boundary-collector-synthesis.md.

3. Scope

Dimension Count
Interfaces 5
Methods 14 (13 empirically derived + 1 from correctness analysis)
Correctness constraints 14 (each with a stated testability criterion)
Fail-closed enforcement layers 5
Substrate Dolt — bitemporal provenance ledger

The specification fixes the contract — interfaces, types, exception hierarchy, and correctness invariants — and deliberately stops short of implementation. Thirteen open questions are surfaced explicitly in specs/TRUST-BOUNDARY-PROTOCOL-SPEC.md §11 rather than papered over; three of them (signal-type taxonomy, the no-confidence-mechanism adapter, and build phasing) are resolved in this draft.

4. The five interfaces

The interfaces form a dependency DAG, not a linear stack. UniversalProvenanceTrace is the shared write substrate every layer audits into.

Layer Interface Responsibility
L1 UniversalProvenanceTrace Append-only event substrate — the shared write sink for all layers
L2 TemporalQueryInterface Bitemporal query — valid-time and transaction-time axes
L3 EpistemicBoundaryProtocol Per-domain confidence inspection — the primary domain contract
L4 TrustAccumulationBus Signal accumulation and the authoritative 5-gate admission chain
L5 PolicyCompositionLayer OPA/Rego policy composition with self-validation

The critical hot path is TrustAccumulationBus.request_admission: it touches all five layers, has a maximum call depth of four, and includes an external policy-engine dependency. All five gates must pass. The aggregation satisfies a veto property — a single registered domain below its threshold denies the action regardless of every other domain's confidence.

5. Integration targets

TRUST-BOUNDARY-PROTOCOL is designed for stacks composed of independently-developed domains that need to coordinate cross-domain actions under shared epistemic constraints.

It does not assume a particular language, framework, or deployment topology. A domain participates by implementing EpistemicBoundaryProtocol and registering with the TrustAccumulationBus; the rest of the layer treats it uniformly.

See specs/TRUST-BOUNDARY-PROTOCOL-SPEC.md §9 (Domain Mapping) for the nine reference systems whose convergence motivates the design, together with each system's current confidence mechanism, fail-closed posture, and gap analysis. The §9 table is the empirical basis of the specification — this README declines to re-render it here, to avoid drift between the two documents.

6. Status

v0.1-draft — pre-/spec-init.

Open questions are tracked in specs/TRUST-BOUNDARY-PROTOCOL-SPEC.md §11. Three are resolved in this draft:

  • OQ-3 — Signal type taxonomy. A closed, peer-level enumeration of five signal types — GROUNDEDNESS, POLICY, SAFETY, STABILITY, AUTHORITY — where no type overrides, supersedes, or escalates another.
  • OQ-5 — No-confidence-mechanism domains. A domain with no native confidence source is given a stub adapter that returns INDETERMINATE for every query; it can be registered but never admitted until it ships a real confidence mechanism.
  • OQ-6 — Build phasing. Full-stack bottom-up (UPT → TQI → EBP → TAB → PCL) is the only supported delivery order; phased delivery was evaluated and rejected.

Ten open questions remain — method-signature decisions, token TTL and presentation, payload encoding, constraint-snapshot versioning, domain lifecycle, distributed-deployment concurrency, signal normalization, and the policy bootstrap path.

7. Substrate

TRUST-BOUNDARY-PROTOCOL's provenance ledger is specified on Dolt, a version-controlled SQL database.

Provenance is bitemporal: every record carries both a valid-time interval (valid_from, valid_to) and a transaction-time stamp (tx_time), so the layer can answer both "what was true at time T" and "what did we believe was true at time T."

Append-only integrity is enforced in five independent layers, so that no single bypass defeats the guarantee:

  1. SQL privilege grants
  2. Dolt branch protection
  3. Remote-side hooks
  4. Signed commits
  5. An application-level hash-chain

See specs/TRUST-BOUNDARY-PROTOCOL-SPEC.md §7 for the bitemporal schema and the enforcement design.

8. Disciplines this specification exercises

This document is also a worked artifact for several engineering disciplines:

  • Mission Autonomy — Policy & Behavior. Cross-domain, confidence-gated action selection: deciding which domain may act, and whether it may act, from a composite epistemic state rather than a single local score.
  • Staff Software Systems Safety. A five-gate, fail-closed admission chain; fourteen testable correctness constraints; and the ternary INDETERMINATE = DENIED discipline that removes silent-allow failure modes.
  • Technical Program Management. A five-interface decomposition with an explicit dependency DAG, thirteen open questions surfaced before implementation, and a documented phased-delivery analysis (OQ-6).
  • Senior Human-Machine Teaming Engineer. At its core TRUST-BOUNDARY-PROTOCOL is human-machine teaming confidence handling: a shared, inspectable model of every participant's epistemic readiness, so that cooperating systems — and their operators — gate joint action on the same evidence.

9. License

This repository is licensed under the Apache License, Version 2.0 (Apache-2.0).

It is a documentation-only repository — a Markdown specification, a pattern report, a white paper, and supporting docs. You may use, share, and adapt the material, including for commercial purposes, under the terms of the license. Full text in LICENSE; the attribution and copyright notice is in NOTICE.

10. How to navigate this repository

Path What it is
specs/TRUST-BOUNDARY-PROTOCOL-SPEC.md The specification — interfaces, shared types, exception hierarchy, correctness invariants, domain mapping, open questions
docs/pattern-report/trust-boundary-collector-synthesis.md The empirical convergence evidence — how nine systems independently arrived at the same pattern
docs/TRUST-BOUNDARY-PROTOCOL-whitepaper.md The white paper — TRUST-BOUNDARY-PROTOCOL explained in narrative form
docs/TRUST-BOUNDARY-PROTOCOL-WORKFLOW.md The spec-authoring pipeline — how this specification was produced

Start with specs/TRUST-BOUNDARY-PROTOCOL-SPEC.md. The pattern report supplies the empirical evidence; the white paper gives the narrative overview; the workflow document records the method.


For agents

Agents reading this repository should start at AGENTS.md, not this README. Claude Code users: see CLAUDE.md, which imports AGENTS.md. The agent files document the conventions, vocabulary, and contribution discipline that agents are expected to follow.

About

Multi-interface trust boundary protocol for AI agent systems

Topics

Resources

License

Contributing

Security policy

Stars

Watchers

Forks

Packages

 
 
 

Contributors