CodexGAS is a manual, LLM-driven model governance workflow.

modelgas/ is not executable code. It is a prompt + schema library: • modelgas/skills//skill.md: what to ask the LLM to do (rules + intent) • modelgas/skills//schema.json: the JSON shape the LLM must output (when required) • modelgas/skills//data/: supporting templates and rubric YAMLs

You (or an agent) run the “skills” by feeding the relevant project context to an LLM and saving the resulting artifacts to disk.

⸻

Critical rules for agents and reviewers

1. Ignore out/

Do not read out/ for evidence, analysis, or inputs.

out/ contains generated artifacts and may be stale, partial, or from a different run. Treat it as output-only.

Only use input// (and other source folders if explicitly instructed) as inputs/evidence.

2. Evidence-only claims

Any assertion, conclusion, or control must be backed by evidence from input//. • Cite the filename (and line range where practical). • If something cannot be supported by evidence, explicitly write “Not evidenced”.

3. No fabrication

Do not invent: • files • interfaces • behaviors • controls

If required information is missing, stop and list exactly what is missing.

4. Existing vs proposed controls

When discussing controls: • Clearly label existing controls (evidenced in inputs) • Clearly label proposed controls (recommendations only)

Do not blur the two.

⸻

How to run CodexGAS (manual)

1. Pick an input project, the user should have told you

Choose an input// directory (example: input/FXChat/).

2. Build a single evidence pack

Create out//aggregate.md by concatenating (copy/paste is sufficient): • input//model_description.md (or equivalent) • Relevant source files (.py, .md, configs) from input// • Example scripts that demonstrate expected usage

Notes: • Do not include anything from out/. • Keep the evidence pack focused on purpose, interfaces, behavior, constraints, and failure modes. • Observed outputs may be included as examples only; do not infer undocumented behavior from them.

Evidence pack checklist • Model purpose and non-goals • Entry points / APIs / scripts • Input and output schemas • Determinism controls (temperature, seed, retries) • Error handling and fallback behavior • Example inputs and expected outputs • Production assumptions (auth, rate limits, logging, storage)

⸻

3. Run skills (ask the LLM)

Skills must be run in the following fixed order: 1. risk_tiering 2. methodology 3. alw 4. tests 5. opm 6. prod_controls 7. documentation 8. remediation_pack 9. iteration

For each skill in modelgas/skills/<skill_name>/: • Read skill.md for purpose and rules. • Provide the evidence pack plus any relevant data/* files (rubrics/templates).

⸻

JSON outputs (schema-bound)

For the following skills, output JSON that matches schema.json exactly: • risk_tiering • methodology • alw • tests • opm • prod_controls

Rules: • No extra keys • No missing required keys • If a field cannot be populated, use an explicit null or sentinel value as allowed by the schema

Naming convention: • out//skills__.json

Example: • out/FXChat/skills_FXChat_risk_tiering.json

⸻

Final report skills (Markdown)

These skills produce Markdown artifacts, not required JSON.

documentation • Write out//docs/model_doc.md • Use modelgas/skills/documentation/data/doc_template.md • Do not change headings or structure; only fill placeholders

remediation_pack • Write artifacts under out//remediation/ • Include patch instructions, test recommendations, and acceptance criteria

iteration • Write out//human/human_responsibility.md using responsibility_template.md • Write out//plan/plan_next.md using planning_template.md

If optional JSON summaries are produced for these skills, keep them clearly labeled and separate. Markdown artifacts are the required deliverables.

⸻

Document versioning and diffs

If a previous version of any output document exists: • Do not overwrite or destroy it • Create a new version alongside it • Produce a Markdown diff file: • out//diffs/diff_to.md

The diff must: • Clearly show additions, removals, and changes • Preserve the original document intact

⸻

FXChat example (input/FXChat)

Evidence pack contents

Include at least: • input/FXChat/model_description.md • input/FXChat/FXChatProcessor.py • input/FXChat/batch_run_50.py • input/FXChat/example_openai.py • (Optional) input/FXChat/batch_run_50_results.json as observed example outputs

Outputs to produce

JSON (per skill, per schema): • out/FXChat/skills_FXChat_risk_tiering.json • out/FXChat/skills_FXChat_methodology.json • out/FXChat/skills_FXChat_alw.json • out/FXChat/skills_FXChat_tests.json • out/FXChat/skills_FXChat_opm.json • out/FXChat/skills_FXChat_prod_controls.json

Markdown (final reports): • out/FXChat/docs/model_doc.md • out/FXChat/remediation/ • out/FXChat/human/human_responsibility.md • out/FXChat/plan/plan_next.md