Work queue #

This file is the canonical project backlog. Update it when work is accepted, split, rejected, or completed. A checked box requires a repository receipt such as a test, frozen artifact, evaluation result, or publication readback. Provider calls and paid training remain separately authorized operations.

Curriculum generation #

• CUR-01: Build the deterministic candidate-generator interface. plan-curriculum-generation emits 228 schema-validated assignments with exact distribution, source, skill, train-demonstration, rubric, and contract receipts. Repeated runs are byte-identical and provider-free.
• CUR-02: Freeze teacher and reviewer prompts. bind-curriculum-assignment freezes exact prompt files and hashes, allowed source and train-demonstration receipts, explicit model profile, sampling controls, acceptance-schema hashes, and a conservative provider cost ceiling.
• CUR-03: Build candidate acceptance. Proposed candidates omit final identity and split; deterministic receipts enforce assignment/schema/provenance contracts, exact and near-template deduplication, separate reviewer identity, rubric gates, and explicit reject/accept outcomes. Acceptance makes a candidate freeze-eligible but does not freeze it.
• CUR-04: Run the first original-synthetic pilot. The one-assignment vertical slice is complete and correctly rejected both sampled candidates; the planned 228-candidate pilot remains unrun. It still requires a printed provider cost bound and fresh authorization. Do not introduce CC BY source text in the first pilot.

Evaluation and training #

• EVAL-01: Establish cumulative base-model baselines for Stages 1–6. Freeze renderer, sampling controls, held-out records, and per-stage gate results before training.
• EVAL-02: Implement the curriculum matrix report. Compare every checkpoint against the matched base and previous promoted phase by gate, task family, skill, and difficulty axis.
• TRAIN-01: Execute one expanding-replay phase. Only after candidate acceptance, exact token rendering, a cost plan, and fresh authorization.
• TRAIN-02: Reject or promote under a separate decision. Lower training loss is never promotion evidence by itself.

Repository trajectories #

• PI-01: Mine historical code/documentation co-change tasks. Materialize parent commits, expose only implementation-side changes, keep historical documentation patches hidden, and preserve repository-disjoint splits.
• PI-02: Compile expected_effects into executable checks. Evaluate required symbols and facts, forbidden claims and mutations, target existence, verification commands, and clean termination rather than merely validating their schema.
• PI-03: Complete private credential delivery. Pass provider credentials through a private file descriptor or equivalent secret mount inside the isolated namespace. Never copy an environment file into a task workspace.
• PI-04: Run one teacher trajectory. Requires an accepted task, isolation receipt, explicit teacher profile, exact tool schema, cost ceiling, and fresh authorization.
• PI-05: Evaluate base and checkpoint through the same Pi harness. Compare filesystem effects, mutation scope, grounding, tool failures, path economy, and termination.

Corpus and rights #

• DATA-01: Expand the audited repository corpus. Favor shorter, diverse targets across Python, Rust, Julia, TypeScript/JavaScript, and C++ rather than another one-repository page pile.
• DATA-02: Strengthen deterministic claim checks. Add signature, default, exception, command-path, and runnable-example gates beyond approximate lexical coverage.
• RIGHTS-01: Decide share-alike treatment. Do not ingest Diátaxis or Red Hat source text until dataset and checkpoint obligations are explicit.
• RIGHTS-02: Implement CC BY receipts before using Google text. Store the exact page revision/hash, retrieval date, attribution, license link, and modification notice.

Project operations #

• OPS-01: Add credential-dark CI. Run Ruff, tests, curriculum/source/model validation, seed determinism, and package audits without ambient provider credentials.
• OPS-02: Add a release checklist. Include source and license scans, artifact exclusion, dependency audit, changelog, and live remote readback.
• OPS-03: Review this queue after every promoted phase. Remove stale tasks, split oversized items, and record rejected directions rather than leaving zombie plans.

Completed foundations #

• Provider-neutral software-source and model manifests.
• Provenance-preserving dataset records and repository-level splits.
• Seven-stage technical-writing curriculum with generation, rubric, provenance, and expanding-replay contracts.
• Forty-four checked-in curriculum seeds with separate supervised and repository train/evaluation lanes and train demonstrations for every planned task family and skill.
• Pinned Pi tool schema, Git-free target masking, Bubblewrap isolation proof, mutation evaluation, trajectory export, and per-assistant-decision Qwen rendering.
• Deterministic curriculum validation and four-file seed-pack compilation.