Repository navigation
First public release of the companion reference implementation for the paper
Executable Trust: The Runtime Architecture of Production-Ready Enterprise AI
by Moataz Mahmoud, M8 Strategies.
The paper argues that enterprise trust is an executable runtime capability
rather than a document, defined by four properties: Versioned, Enforced,
Independently Measured, Ratified. This repository implements those mechanisms
as inspectable, reproducible code.
What is in this release
Versioned trust contract. A ratified contract (v1.0) declaring the
controlled vocabulary, evidence thresholds, verification rules, lifecycle
transitions, and telemetry rules, with 38 stable rule identifiers. A draft
contract (v0.9) ships alongside so the repository can demonstrate, rather than
assert, that an unratified contract cannot govern a decision.
Fail-closed runtime enforcement. A pure, deterministic decision function
over claim-level verdicts with nine distinct refusal reason codes. There is no
default value for any required field and no configuration that disables
verification.
Evidence authorization and sufficiency. Authorization is evaluated before
retrieval; evidence provenance and sufficiency are checked before generation.
The tests assert the generator is never invoked on those refusal paths.
Immutable decision lifecycle. Decision records are append-only. Lifecycle
state (PROPOSED/ACCEPTED/REJECTED/SUPERSEDED) is derived by folding an
append-only transition log rather than stored, so no field can disagree with
history. Acceptance and rejection require an attributable accountable human.
Ratified amendments. Contract changes are dated, attributed amendments with
three distinct roles; an author may not ratify their own. Prior contract
versions are preserved on disk, never rewritten.
Traceability. Every contract rule identifier maps to at least one test, and
the build fails if any rule is untested or if a test names a rule that no longer
exists.
Synthetic evaluation harness. A deterministic offline harness drives the
real decision path against a 27-case human-authored golden set, producing
byte-reproducible reports.
Also included: verification dependency degradation, fail-open telemetry with
explicit environment and population provenance, metrics that report counts
always and withhold rates below the contract's minimum sample, and validators
for contracts, traceability, publication boundary, and release metadata.
Verification
- 175 tests passing
- 92.49% line coverage
- Runs with no model provider, network access, database, or credentials
Scope and evidence boundary
This repository is an independent reference implementation. Its design is
informed by lessons from a working production enterprise AI platform, while the
implementation, examples, contracts, and evaluation data included here are
publication-safe and independently reproducible. It excludes the commercial
platform, proprietary integrations, production prompts, operational data,
customer information, and internal deployment architecture.
The reference implementation includes several mechanisms specified by the paper
that extend beyond the production implementation from which some field
observations were drawn — specifically, verification dependency degradation and
telemetry population provenance. These demonstrate the paper's normative
architecture and are not evidence that every mechanism was operating in the
production system at the time of the reported observations.
Evaluation results are labelled Reference implementation synthetic baseline.
They describe this implementation's conformance to its own contract. They are
not a measurement of any production system and are not comparable to the field
observations reported in the paper.
Known limitations are enumerated in
docs/limitations.md.
License
Apache License 2.0. The paper and its visual assets remain subject to the
publication terms established by M8 Strategies.