Skip to content

Latest commit

 

History

History
211 lines (162 loc) · 12.2 KB

File metadata and controls

211 lines (162 loc) · 12.2 KB

Agentic Engineering Assurance System

DOI

Governance, Auditability, and Release Integrity for AI-Assisted Software Engineering

The Agentic Engineering Assurance System is a governed system for long-running, AI-assisted software delivery. It keeps implementation, read-only technical audit, human authority, evidence, and release distinct so that material decisions remain attributable and reviewable after the conversation that produced them has ended.

This repository is the bounded public technical record. It documents the design, implemented control model, recorded outcomes, release evidence, and known limitations without disclosing the private operating environment or its sealed evidence archive.

In 30 seconds

Question Short answer
What is this? A governed system for long-running, AI-assisted software delivery, documented as a public technical record.
What does it separate? Human authority, implementation, read-only technical audit, durable evidence, and release.
How is it controlled? Bounded work packages, pinned source, explicit finding disposition, owner gates, and a separate release step.
What can a reader inspect? Public claims and their boundaries, Release v1.0.0, publication metadata, and the version DOI.
Who is accountable? Ata Nasseri, System Designer and Assurance Owner.
What is the current boundary? Evidence-backed and publicly released; independent technical review has not yet been performed, and no external certification is claimed.

Choose your reading path

If you are… Start here Then inspect Primary question answered
A product or system leader This README and the System Overview Limitations and Reassessment What is governed, who decides, and where are the boundaries?
A technical reviewer Implementation Record Assurance Method and Public Evidence Index What was recorded, how was it reviewed, and how strong is each claim?
A risk, governance, or assurance reader Assurance Method Public Evidence Index and Limitations Which decisions remain human, and what is explicitly not established?
A release or provenance specialist Current assurance status Signing and Release, publication metadata, and DOI guidance Which exact public object is citable, persistent, and integrity-bound?
A prospective independent reviewer Independent Review Brief Review Checklist, review status, and the immutable v1.0.0 GitHub Release with its signed Git objects What must be independently checked before any external conclusion is published?

Professional accountability

Ata Nasseri, System Designer and Assurance Owner

The accountable scope represented here includes translating objectives and constraints into bounded acceptance criteria; defining authority boundaries; authorizing audit rounds; deciding criteria changes and residual risk; and governing release. AI implementation and audit roles operated within those human-defined boundaries.

This describes accountable system and assurance leadership. It does not claim sole authorship of every implementation artifact or independent certification of one's own work.

Why it exists

AI agents can produce substantial engineering work, but speed alone does not create assurance. Long-running work introduces practical risks:

  • implementation and review can blur together;
  • a reviewer can inspect a moving source tree;
  • corrective work can introduce new defects;
  • authority can become implicit in chat history;
  • operational observations can be overstated as proof; and
  • a release can become difficult to reconstruct later.

The system addresses these risks through explicit authority, exact source checkpoints, a read-only audit role, bounded review rounds, durable decision records, evidence-qualified claims, and controlled release.

System at a glance

AEAS system map showing human authority above a bounded implementation, read-only audit, disposition, approval, and separate release lifecycle, with durable records below

Open the full-size system map.

The owner defines objectives, acceptance criteria, continuation authority, residual-risk decisions, and release approval. The implementation role changes the system. The audit role inspects a pinned source position without modifying it. Git provides the durable evidence surface connecting all three. In text, the governed sequence is: bounded work package → implementation → exact source checkpoint and integrity-bound handoff → read-only audit → finding disposition → owner decision → terminal approval → separate release action.

Recorded outcomes

AEAS recorded outcomes showing bounded report, review, finding, disposition, correction, and per-tier test counts with evidence-class labels

Open the full-size recorded-outcomes visual.

The sealed private evidence baseline supports the following public statements:

  • seven audit reports were preserved across three governed reviews;
  • one attempted round was explicitly recorded as not auditable;
  • the six substantive audit rounds reported 33 findings: 11 HIGH and 22 MEDIUM, with no CRITICAL finding recorded;
  • resolution records mark 32 findings as fixed and one as settled through an owner-ratified criteria revision;
  • each of the six substantive audit reports recorded a changes-required verdict;
  • the final delivery review used two authorized rounds and reported eleven MEDIUM findings;
  • three final-round findings were defects introduced by earlier corrective work; and
  • the committed completion record reports 1,111 tests green in both the hermetic and explicit live-host tiers.

These statements are deliberately bounded. They describe retained records and verified Git facts; they do not prove the absence of undiscovered defects or constitute external certification.

Engineering principles demonstrated

  1. A fix is a new change, not proof. Corrective work receives the same skepticism as the original implementation.
  2. Evidence collection is not enforcement. A signal only matters when the decision logic actually requires it.
  3. Fail-closed controls must run before side effects. A late refusal is not a safe precondition.
  4. Tests must exercise real call shapes. Simplified test vectors can miss production boundary failures.
  5. Approval is not release. Risk acceptance, merge, tagging, deployment, and human-only observations remain distinct events.
  6. Limitations belong in the result. Assurance becomes more credible when non-claims remain visible.

Read the system record

Document Purpose
System Overview Components, boundaries, and system behavior
Implementation Record What was delivered and what the audits changed
Assurance Method Authority, evidence, review, and claim discipline
Public Evidence Index Public claims and their evidence boundaries
Limitations and Reassessment Known constraints and future triggers
Independent Review Scope and status of third-party review
References External standards and platform mechanisms used by the public release process
Publication Charter Public/private disclosure boundary
Visual System Semantic color, typography, connector, accessibility, and change-control rules for public visuals
Visual Manifest Machine-readable claims, sources, dimensions, non-claims, and approval status for every visual asset

Release and external-review materials are kept separate from the system description:

Package Purpose
Publication Gate Separate repository-deployment, Release, archival, and review gates
Signing and Release Signed release commit, signed annotated tag, immutable GitHub Release, and provenance process
DOI and Archival Zenodo and persistent citation process
Independent Review Brief Scope for a qualified external reviewer
Independent Review Checklist Evidence, identity, method, and conclusion checks for the reviewer
Independent Technical Review Statement Template Required structure of the signed external conclusion

Evidence and release chain

AEAS evidence and release chain separating public claim discipline, the later public release record, and independent scrutiny

Open the full-size evidence-and-release visual.

The visual deliberately shows three separate groups rather than one unbroken proof chain:

  1. a public claim is bounded by its identifier, evidence class, source, and explicit non-claims;
  2. the later publication record identifies the signed release commit and signed annotated tag for v1.0.0, the immutable GitHub Release, checksums and provenance workflow, and persistent version DOI without changing those released objects; and
  3. independent scrutiny remains NOT_PERFORMED until a qualified reviewer publishes a scoped conclusion under their own control.

Private evidence relationship

This public record derives from a sealed private baseline identified as PTE-2026-08-19-v1.0.1. The baseline manifest is committed here only by its SHA-256 digest:

ecd87eb9581e535c95dee1beb0871fa5fc1382c9d3847eb5ebba337bc5ef827f

The digest is a commitment to a specific private manifest. It does not reveal the manifest, independently validate its contents, or grant public access to the underlying evidence.

Current assurance status

Release v1.0.0 has verified signed Git objects and an immutable GitHub Release. Its release assets have SHA-256 checksums and GitHub artifact attestations; Zenodo provides a separate persistent version DOI. Independent technical review has not yet been performed, and no external certification is claimed. The independent-review package is prepared so that a qualified reviewer can assess that precise release commit, signed annotated tag, and immutable GitHub Release, then publish a scoped conclusion tied to them.

The immutable v1.0.0 GitHub Release remains the citable release object, bound to a signed release commit and signed annotated tag. The default branch contains later publication-record and documentation refinements; they do not alter those signed Git objects, the Release assets, or the DOI.

Use and citation

Documentation and visuals are licensed under CC BY-NC 4.0. Small software utilities in this repository are licensed under Apache 2.0. See License, Notice, and Citation.

Persistent identifiers: