Skip to content

Thesis context

This software is the artifact of a master's thesis:

Diseño y evaluación de una arquitectura de harness engineering para el desarrollo de software asistido por inteligencia artificial, con supervisión humana y retrospectiva basada en evidencia. Juan Sebastián Barrera Pulido. Trabajo de grado, Maestría en Informática, Escuela Colombiana de Ingeniería Julio Garavito, Bogotá, Colombia.

The thesis asks to what extent a harness engineering architecture based on normative phases, verifiable gates, controlled memory, human supervision and evidence-based retrospective improves the quality, traceability and control of AI-assisted software development in greenfield and brownfield scenarios. It follows Design Science Research: the harness is the designed artifact, and its evaluation is formative and descriptive (no statistical generalization).

Evaluated versions

The thesis keeps chapters 1 to 4 of its manuscript; chapters 5 to 8 are rewritten with the results of 2.0.1, the version of the controlled evaluation (campaign in progress, evaluation/campaign/). v0.8.0 is an earlier, obsolete Design Science Research iteration: it stays documented and reproducible byte for byte (see provenance), its tag is never moved, and its behavior may be changed in new versions. Results of earlier versions are in the git history and the v0.8.0 tag and are not part of the 2.0.1 analysis. v0.8.1 adds infrastructure, documentation and quality fixes without behavior change. v0.9.0 fixes seven behavioral defects; the only difference visible in the thesis cases is the brownfield closure, which is now a plain APPROVE instead of an APPROVE_EXCEPTION (see the brownfield guide). v1.0.0 adds, after the evaluation, governed memory operations, decisions on retrospective recommendations and the usage reported by agent providers (see memory and retrospective decisions); with no memory records the run path is that of the earlier iteration. v2.0.0 adds the governance found missing by that evaluation and by use on real projects: clarification and criteria at intake, independent acceptance tests, a verification ladder with certification, a review panel with layered rule catalogs, capabilities per phase, an agent write sandbox, stop the line, routing anchored at the invoking model, low-friction small changes, local metrics, forges and embedded mode (see the changelog). Every new behaviour sits behind a configuration key that harness init writes.

The controlled evaluation installs the harness from the wheel built from its release tag, not from a source tree, so the version under test is the one each environment.json records (with the wheel's SHA-256 and the commit). The deterministic suites run on GitHub Actions; the runs with a model run on the author's Windows PC, declared as such. Pilot and precheck attempts made before the campaign are apparatus checks and are listed, with their cause, in the preregistration (evaluation/campaign/PREREGISTRO.md, amendment of 2026-11-12).

Snapshot of the earlier iteration (v0.8.0)

The figures of the earlier iteration v0.8.0, reported by its thesis manuscript (they describe that frozen cut, not the current state of develop nor the 2.0.1 evaluation):

Fact Value
Python modules in src/governed_harness 74, with 6,746 lines
JSON Schema contracts 25
CLI commands 21
API routes 9 (seven queries, one decision POST, the dashboard at /)
Tests 86: core 79, E2E Python 4, E2E Node.js 1, performance 2
Full suite in one process exit 0 in three runs, 19–21 s
Core coverage lines 80.60 %, branches 59.29 % (77.48 % combined), 3,268 statements
Ruff 0.16.5 140 findings in src, 11 in tests (60 F401, 40 B008)
Mypy strict package mode blocked by the missing py.typed; 54 errors in 15 files on sources
Wheel in a clean environment installs its runtime dependencies (18 counted during publication) and harness runs
Exit codes 0, 1, 2, 3, 4, 5, 6, 130
Brownfield case pallets/itsdangerous 2.2.0; baseline broken by a missing freezegun; 298 tests after installing it; ChangeSet digest invariant

The re-verification of these figures during publication is recorded in the v0.8.0 release notes.

How to cite

Use the citation metadata in CITATION.cff (GitHub shows a "Cite this repository" button), and cite the version you used, for example v0.8.0 for the earlier, documented iteration or v2.0.1 for the controlled evaluation.