# BYOB BFCL Context

Use this pack for `byob/bfcl`: executable function-calling benchmark
generation, localization, and candidate evaluation from an allowlisted oracle
pack.

## Routing

- `stage=prepare`: validate/normalize the oracle pack; no publishable benchmark.
- `stage=generate`: build canonical raw/published parquet, caches, validation
  evidence, and `run_manifest.json`.
- `stage=translate`: localize only approved natural-language fields while
  preserving executable truth and source identity.
- `stage=eval`: verify the frozen source, check candidate contamination, run
  native function-calling episodes, and write bound eval artifacts.
- `stage=all`: use only with a config whose intended enabled stages were read.

The oracle pack contains tools, fixtures, task templates, assertions, and
validation cases. It may use an allowlisted local Python backend or the pinned
BFCL Oracle HTTP contract. Never execute unreviewed pack code in the evaluator
process; the runtime isolates it in workers.

## Publication Invariants

- `benchmark.parquet` plus `run_manifest.json` are canonical. Optional BFCL
  JSONL and NeMo Evaluator bundles are derived compatibility exports and need
  `export_validation_report.json` before use.
- Use `tiny` for the smallest no-model-call plumbing check. A smoke lineage is
  intentionally not publication-eligible.
- Resume only through verified stage checkpoints; keep release/config/source
  hashes bound together.
- Translation starts from the immutable source run manifest, not a bare
  parquet. Do not translate tool names, schemas, expected calls, fixture values,
  assertions, tool results, held-out state, or lineage.
- Before candidate inference, verify source hashes/publication semantics and
  evaluate contamination. Unresolved candidate/source identity must not be
  guessed away for a publishable result.
- Native candidate requests preserve tool calls and raw arguments. Do not
  repair model output before scoring.

## Evaluation Modes

Direct BFCL evaluation runs through the BFCL runtime. Launcher mode materializes
an immutable adapter package and then requires the generated package to be
installed explicitly before submission. Read the selected `eval.*.yaml` and
the live scoring contract before producing a runnable command; the BFCL eval
surface is intentionally too strict to reconstruct from memory.

## Authoring Boundaries

Oracle-pack scaffolding, assisted authoring, MCP intake/gateway, release review,
freeze, publication, revocation, and bias audit are explicit workflows with
their own approval/evidence contracts. Do not infer authorization to publish,
freeze, revoke, contact an MCP target, or execute external mutations merely
because the user asked to configure `byob/bfcl`.

## Live Sources

- Manifest: `src/nemotron/steps/byob/bfcl/step.toml`
- User flow: `src/nemotron/steps/byob/bfcl/README.md`
- Oracle contract: `src/nemotron/steps/byob/references/bfcl-oracle-pack.md`
- Eval contract: `src/nemotron/steps/byob/references/bfcl-eval-scoring-contract.md`
- MCP/user authoring references: `src/nemotron/steps/byob/references/bfcl-*.md`
- Pattern: `src/nemotron/steps/byob/patterns/create-bfcl-from-oracle-pack.md`
