Skip to content

Latest commit

 

History

93 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Open Institutional Compiler

  • Status: OWNER-AUTHORIZED BOOTSTRAP - PRE-EXTERNAL-REVIEW
  • Bootstrap date: 2026-07-29
  • Governing design: TDD-OIC-001 v1.1

Open Institutional Compiler (OIC) is developing a reviewable path from the human-readable sources that govern a regulated enterprise to the explicit controls its software and authorization systems can evaluate.

Target capability:

A regulated enterprise should be able to automate a larger class of consequential actions and regulated workflows without lowering the evidentiary standard required to authorize, explain, audit, or correct those actions.

What exists today: a tested, non-semantic Python infrastructure package for offline schema validation, historical bootstrap verification, current manifest verification, environment and gate diagnostics, and reproducible CI and SBOM checks. It is not a functioning institutional compiler.

What is blocked: semantic implementation remains blocked pending the corpus provenance and ZTL/VEIP interface evidence listed in STATUS.md. Nothing in this README opens that gate or promotes the repository's maturity.

Capability being developed

Regulated enterprises govern consequential work through heterogeneous sources: regulations, enterprise policies, procedures, approval matrices, contracts, amendments, definitions, exceptions, and prior decisions. Runtime authorization requires a different form: explicit actors, actions, conditions, authority limits, evidence requirements, effective periods, exceptions, and failure behavior.

The conversion between those forms remains substantially manual:

governing sources
-> human interpretation
-> control matrix
-> developer translation
-> Rego / policy-as-code / workflow logic
-> runtime authorization

Teams reconcile definitions and amendments, identify applicable authority and exceptions, specify evidence, translate the result into runtime artifacts, create tests, and later reconstruct lineage for audit or change analysis. Each translation can create another interpretation of the same governing source.

OIC is intended to support a different control-production path:

governing sources
-> source-grounded candidate meaning
-> explicit ambiguity and missing support
-> authorized institutional admission
-> portable executable control
-> runtime authorization
-> evidence, reliance, and correction

The development hypothesis is that more of this transformation can be automated without automating institutional authority itself.

Development hypothesis

OIC is intended to produce reviewable, source-grounded candidate controls; route unresolved meaning to authorized enterprise reviewers; compile only institutionally admitted meaning; and retain the evidence needed to reconstruct why a runtime action was permitted, denied, escalated, or could not be determined.

Machine extraction may propose meaning. It does not create authority. Confidence does not constitute institutional admission. Ambiguity, contradiction, unsupported meaning, missing authority, unresolved exceptions, and insufficient evidence must remain visible. They must not be coerced into an executable ALLOW.

Authorized enterprise reviewers remain responsible for admitting candidate meaning for a stated scope and use. Target code is a projection. The proposed canonical admitted representation is Institutional IR, and the proposed portable enforcement boundary is the Open Control Envelope.

If this approach succeeds, a larger set of consequential regulated actions may become safely automatable without reducing the authority, evidence, review, or correction obligations that govern them. This is a development hypothesis. It is not an established release, benchmark, adoption, or compliance claim.

Why now

Runtime authorization infrastructure can evaluate structured policy at high volume. Document parsing and language models can extract, classify, anchor, and propose candidate structures. Agentic systems can propose and execute consequential actions faster than repeated undocumented human interpretation can scale.

These capabilities do not remove the control-production boundary. Generated candidate meaning is still untrusted. A regulated enterprise must determine which source applies, what it authorizes, what evidence is required, which exceptions govern, and who may admit the result.

The control-production bottleneck is the creation of an authorized, versioned, source-grounded control with explicit uncertainty and a correction path.

Runtime authorization starts downstream

OPA evaluates policy over structured input and data and returns policy decisions to the integrating application. Rego is OIC's first proposed executable target. OIC operates upstream of that evaluation boundary; it does not replace OPA.

Cedar, relationship authorization, and comparable systems provide mature runtime authorization approaches. They expect policies, models, relationships, entities, or equivalent control inputs to be supplied. OIC addresses the reviewable path by which heterogeneous regulated-enterprise sources could become those institutionally admitted inputs.

Rules-as-Code and computational-law systems address important parts of the upstream problem. Many depend on deliberate human formalization or focus on a specific domain. Document AI and language models can assist extraction and proposal, but generated output does not establish institutional authority.

The interface is:

policy engine: execute or evaluate the control
OIC:           establish the reviewable path from governing sources to that control

Existing systems are not deficient for stopping at this boundary. That boundary is the interface OIC is designed to supply. See the prior-art and build/reuse treatment in TDD-OIC-001 v1.1, especially Section 3.

The differentiating hypothesis concerns the integrated transformation and its boundary discipline, not novelty of individual components. OIC does not claim standalone novelty for policy evaluation, Rules as Code, provenance, delegation, semantic preservation, authority versus logical warrant, issuance versus evaluation, counter-evidence, or the distinction between institutional meaning and representation.

The proposed conjunction is:

regulated-enterprise governing sources
-> source-grounded candidate meaning
-> explicit uncertainty
-> authorized institutional admission
-> target-independent control representation
-> runtime policy
-> evidence
-> correction

Whether this conjunction is useful and differentiated remains unproven until comparative benchmarks, independent review, and external consumption establish it. See CLAIMS.md for the evidence required before stronger statements are permitted.

How OIC will be measured

Generating syntactically valid Rego is not the OIC success criterion. OIC-Bench is intended to test whether the source-to-control transformation preserves source support, authority, uncertainty, exceptions, temporal and currentness state, evidence requirements, runtime behavior, change impact, and human-review efficiency.

The relevant question is whether OIC can automate more of regulated-enterprise control production while preserving the evidence and authority properties that make the resulting controls safe to rely on. OIC is successful only if increased control-production automation does not come from converting unsupported, ambiguous, stale, or unauthorized meaning into executable controls.

The following development gates are preregistered in TDD-OIC-001 v1.1. They are targets, not observed results, calibrated thresholds, independent validation, or current product claims.

Benchmark Target Status
Source-supported executable fields >=99% TARGET - NOT MEASURED
Unsupported executable-field rate <=1% TARGET - NOT MEASURED
Unknown-to-false conversions 0 TARGET - NOT MEASURED
Authority Reconstruction F1 >=0.80 TARGET - NOT MEASURED
Ambiguity recall >=0.85 TARGET - NOT MEASURED
False-resolution rate <=0.05 TARGET - NOT MEASURED
Behavioral conformance >=0.95 TARGET - NOT MEASURED
Change-impact recall >=0.95 TARGET - NOT MEASURED

Provisional comparative target

Separately, TDD-OIC-001 v1.1 preregisters a provisional comparative-claim threshold: at least 40% reduction in active human time to first admitted envelope versus the manual baseline, with no more than 2 percentage-point loss in adjudicated behavioral quality.

Status: PROVISIONAL TARGET - NOT MEASURED - NOT CALIBRATED.

This threshold is subject to calibration, retention, amendment, or withdrawal under the benchmark governance process before it can support a public comparative claim. It is not an absolute safety or release gate.

Full OIC-Bench v0.1 is designed around public, frozen-test, private held-out, adversarial, change-impact, runtime, and practitioner-review partitions. Its proposed aggregate scope includes 21 governing documents, approximately 140 pages, at least 260 adjudicated normative clauses, at least 85 candidate controls, at least 34 known ambiguities or conflicts, at least 160 runtime cases, and at least 18 source-version pairs. PROPOSED BENCHMARK SCOPE - NOT A MEASURED SUFFICIENCY CLAIM.

Comparative baselines

OIC-Bench is designed to compare OIC with named control-production paths, not only with earlier OIC versions:

  • Direct LLM-to-Rego: the same source inputs and, where architectural comparison requires it, the same declared model.
  • Modular document-to-policy pipeline: extraction, validation, and target compilation without OIC's admission and authority representation.
  • Human policy engineer: an experienced practitioner using ordinary policy and control-production tools.
  • Source-grounded knowledge graph: source entities and relations without a formal admission or control envelope.
  • RAG plus rule generation: retrieved source passages followed by target-policy generation.
  • Commercial capability review: public documentation and controlled trials where evidence is available. Unknown capabilities remain UNKNOWN; they are not assumed absent.

No comparative outperformance claim is permitted until the full benchmark, baseline implementations, annotation protocol, and held-out evaluation are frozen and independently reviewed.

Proposed future metric: Safe Automation Coverage

PROPOSED - NOT YET PART OF THE FROZEN OIC-BENCH SPECIFICATION.

Safe Automation Coverage would measure the percentage of in-scope, gold-adjudicated regulated-control cases that reach an admitted executable state while satisfying all applicable source-support, authority, evidence, ambiguity, currentness, and behavioral gates.

CANNOT, escalation, or human judgment may be the correct outcome. The metric must not reward falsely converting ambiguous or unsupported work into executable controls. Safe Automation Coverage is not equivalent to ALLOW rate. No numerical target is authorized.

A related operational measure is admitted automation yield under fixed expert-review budget: the number of valid admitted executable controls produced under a fixed amount of qualified reviewer time, compared with the manual baseline. This measure is also PROPOSED - NOT YET PART OF THE FROZEN OIC-BENCH SPECIFICATION. Formal adoption of either measure requires benchmark governance, an ADR, and a traceability update.

Current measured evidence

TARGET identifies a preregistered or proposed benchmark threshold, not an observed result. MEASURED identifies an observed result tied to a reproducible repository or benchmark composition. BLOCKED or NOT YET RUN identifies an evaluation that cannot proceed because semantic implementation or evidence prerequisites remain incomplete.

Exact repository composition, CI execution composition, and run provenance for these measurements are recorded in the corresponding pull-request and CI evidence record. The measurements below are infrastructure evidence, not semantic OIC-Bench results.

Evidence class Current state
Non-semantic infrastructure verification MEASURED
Schema validation MEASURED - 9/9
Bootstrap integrity MEASURED - 52/52
Infrastructure falsification harness MEASURED - 4/4
Full repository test suite MEASURED - 1232 passed, 1 intentional skip
Manifest MEASURED - INCOMPLETE, required exit 3
Semantic implementation gate BLOCKED
OIC-Bench preflight design PREREGISTERED
Semantic OIC-Bench preflight NOT YET RUN
Full OIC-Bench v0.1 NOT YET RUN
Comparative outperformance NOT ESTABLISHED
Practitioner usability benchmark NOT YET RUN
Regulated-enterprise pilot evidence NOT ESTABLISHED

These results verify repository infrastructure. They are not semantic OIC-Bench results. The benchmark preflight metrics remain proposed or preregistered and not measured. Experimental branch results are not accepted benchmark evidence.

Development roadmap

The roadmap is evidence-gated, not schedule-driven. No stage is complete because its features exist. Advancement requires the named benchmark, verification, operational, and independent-review evidence.

Stage 0 - Infrastructure and benchmark readiness

STATUS: CURRENT

Current capabilities include non-semantic schemas and contracts, manifest/integrity verification, the falsification harness, supply-chain controls, and the benchmark preflight specification.

Exit requires complete governing-source rights and provenance plus sufficient ZTL and VEIP provisional-interface evidence to open the semantic code-start gate.

Stage 1 - Source-to-control reference

STATUS: NEXT - BLOCKED by the semantic code-start gate

Target path:

governing sources
-> source anchors
-> candidate normative meaning
-> ambiguity and review docket
-> authorized admission
-> Institutional IR / Open Control Envelope
-> Rego
-> runtime result
-> lineage

Advancement requires a reproducible semantic preflight, passing critical invariants, published raw benchmark outputs, and visible failures and limitations.

Stage 2 - Evidence release

Target evidence includes a frozen OIC-Bench v0.1; public, frozen, held-out, and adversarial partitions; named comparative baselines; a human policy-engineer baseline; change-impact evaluation; a practitioner usability study; and independent technical and security review.

Exit permits only bounded benchmark claims for the tested scope.

Stage 3 - Regulated-enterprise pilot

Target capabilities include real regulated-enterprise governing sources, SSO and identity, tenancy, evidence-system integration, an on-premises or offline profile, change propagation, rollback, and operational monitoring.

Advancement requires a controlled design-partner pilot, security evidence, runtime and reliability evidence, and scoped pilot-readiness claims.

Stage 4 - Continuous regulated control compilation

Target capabilities include governing-source monitoring, amendment and supersession detection, control-dependency impact, invalidation, re-admission, controlled publication, CI/CD integration, and historical replay.

Advancement requires change-impact benchmarks, currentness and revocation tests, operational incident exercises, and replay and correction evidence.

Stage 5 - Dependable regulated agents

Target capabilities include agent identity and mandate, dynamic grounding, ZTL warrant, VEIP execution continuity, delegation and revocation, reliance and correction, and multi-agent consequential-action chains.

Advancement requires a real consequential-runtime pilot, measured consequence-control behavior, and independent audit.

Later - Open infrastructure

Target capabilities include interoperable control-envelope profiles, independent compiler and adapter implementations, domain packs, conformance suites, and federated authority and control ecosystems.

Evidence requires independent implementations or consuming projects, external conformance, and community governance.

Current phase

This repository authorizes contract-first, non-semantic infrastructure work. It currently implements four infrastructure CLI commands:

  • oic validate-schema
  • oic verify-bootstrap
  • oic verify-manifest
  • oic doctor

It does not implement document interpretation, candidate extraction, institutional admission, Institutional IR production, Open Control Envelope generation, Rego compilation, or runtime semantic decisions.

Run the safe non-semantic checks after the hash-locked installation described in docs/operations/CI.md:

make verify
make falsify

oic verify-manifest --all currently reports INCOMPLETE and exits 3 because the local corpus evidence is incomplete. That result is intentional. It must not be normalized into success.

First executable objective

The first semantic objective remains blocked. When its prerequisites are authorized, a bounded set of public or synthetic procurement governing sources is intended to flow through:

documents
-> source anchors
-> candidate normative units
-> review docket
-> admitted record
-> Open Control Envelope
-> Rego
-> ALLOW / DENY / CANNOT
-> lineage

This objective is limited to the named procurement profile. It does not imply support for arbitrary enterprise policy, jurisdictions, or regulated workflows.

What OIC does not do

OIC does not create institutional authority. It does not turn model confidence into admission, eliminate authorized review, guarantee legal or regulatory compliance, or make automation inherently safer.

The proposed design must fail closed. Missing source support, authority, admission, evidence, current versions, or grounding facts must block, escalate, or return CANNOT. Open-textured standards remain human-judgment boundaries unless an admitted decision procedure exists.

ZTL and VEIP remain provisional project-controlled interfaces. ZTL is intended to evaluate logical warrant from admitted grounds; it does not establish source authority or interpret prose. VEIP is intended to preserve execution and correction continuity; it does not create institutional meaning. Their required evidence is listed in STATUS.md and LIMITATIONS.md.

Governing invariants

The complete set is in docs/requirements/INVARIANTS.md. Core boundaries include:

  • no executable field without a source anchor or an explicit institutionally admitted addition;
  • no enforcement artifact from a state earlier than admitted;
  • candidate confidence never substitutes for institutional admission;
  • unknown never becomes grounded false;
  • missing authority, admission, evidence, current version, or facts never silently yields ALLOW;
  • every executable artifact binds its compiler, schema, source, admission, and test versions; and
  • every published runtime verdict must be replayable to its exact control, admission, sources, warrant, inputs, and versions.

Public limitations

The repository is a governance, contract, and non-semantic infrastructure foundation. It is not a functioning institutional compiler. Current scope, provisional interfaces, corpus restrictions, human-judgment boundaries, and benchmark limitations are recorded in LIMITATIONS.md.

No license grant or SPDX identity is established. Licensing remains pending counsel review.

Claims discipline

Architecture statements, measured results, and release claims are different things. Passing tests or CI does not establish semantic correctness, institutional validity, compliance, comparative advantage, or owner acceptance.

The permitted, evidence-gated, and forbidden claims are recorded in CLAIMS.md. Semantic implementation remains blocked by STATUS.md.

About

Open-source reference implementation for compiling institutionally warranted policy into machine-operational controls with preserved provenance, versioning, authority boundaries, and verifiable source-to-execution traceability.

Topics

Resources

Contributing

Security policy

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Used by

Contributors

Languages