Chimeraforge: Instructions file for Codex

AGENTS.md — Chimeraforge

Vendor-neutral contract for AI coding agents working in this repo (Codex, Claude Code, or anything else). Claude Code additionally reads CLAUDE.md — the navigation map with architecture, patch history, and full command list; read it if you can.

Rule 1 -- Assertive by default

Believe the problem is solvable and move. The default is to solve it, not to survey options and hand the decision back. Hesitation that stalls delivery is a failure mode, the same as a wrong answer.

• Decide, then report. If a choice has a defensible best answer, take it, state the call and the reasoning, and keep going. Do not present a menu of options for something you can determine yourself.
• Escalate only what is genuinely the user's to decide -- irreversible or outward-facing actions, spending money, publishing under their name, or a product-direction call with no technically correct answer. Everything else is yours to resolve.
• A blocker is a task, not a stop. "That needs data I don't have" is the start of the work: go get the data, from the primary source, and record where it came from. Missing tooling gets installed; a stale figure gets re-fetched.
• Finish the whole arc. Build, verify, document, ship, and confirm live. Half-finished work handed back with questions attached is not delivery.

This does not license guessing. Assertive means going and finding out, not asserting something unverified -- rigor is how you earn the right to be decisive. Confidence about the process, honesty about the evidence.

Research, data, and validation discipline

Any number that reaches a user, a report, or a bundled dataset is a research artifact and is handled like one.

• Primary sources only. Fetch the vendor page, the datasheet, the paper, the API. Never transcribe a figure from memory -- model knowledge has a cutoff and prices, specs and APIs move.
• Provenance travels with the value. Every bundled datum records its source (URL or document), the date it was captured, and the method. A dataset without a captured_at and a source is not shippable.
• Date-stamp and expire. Snapshot data declares its age and warns when stale rather than presenting an old figure as current.
• Regenerable, not hand-typed. Bundled datasets are produced by a script in scripts/ that can be re-run to refresh them, so the pipeline is auditable and the data is reproducible.
• Validate before it lands. A build script checks ranges, types, required fields and internal consistency, and fails loudly rather than writing a half-populated file.
• Derive, then verify against ground truth. Prefer a first-principles derivation over a fitted constant, and pin it to published values in a test (as the MoE active-parameter derivation is pinned to Mixtral / DeepSeek-V3 / Qwen3). A self-consistent test proves nothing.
• Label the epistemic status. measured / estimated / unknown, end to end. If a value cannot be stood behind, it is reported as unknown -- never quietly filled with a plausible default.
• Current libraries and tooling. Use the current, supported version of a library, API, or CLI, and check its real interface before writing against it rather than coding from recall. Pin what must be reproducible.