yjz211/hajimi-math-model-agent/modeling-workflow-core

Mathematical modeling workflow

Use this skill for the entire modeling task. It is a working method, not a rigid state machine.

Start every task

1. Inspect the frozen files under input/. Do not search other cases or historical solutions.
2. Call hajimi_task_status, retain its revision, and pass that value as expectedRevision to the next state-changing HaJiMi tool. On a revision conflict, read status again and deliberately rebase the intended change.
3. Read these references before selecting a model:

• references/lab-core/1.0.0/methodology/modeling_workflow.md
• references/lab-core/1.1.0/methodology/validation_framework.md

1. Separate problem facts, assumptions, inferred facts, and computed results.
2. Make the simplest executable baseline before adopting a complex method.

Required modeling loop

• Define each requested output, variable, unit, objective, and constraint.
• Audit data quality and leakage before fitting or optimization.
• Compare plausible approaches when the choice can change the answer.
• Run all code. Never invent a numeric result, figure, citation, or successful validation.
• Test a hand-checkable small case, boundaries, invariants, and applicable V0-V5 checks.
• Register important experiments and artifacts with the HaJiMi tools.
• Keep paper numbers and figures traceable to machine-readable results.
• If evidence contradicts the current route, revise the plan rather than defending it.

Governed workflow state

• The bundled 0—9 milestones are revisitable acceptance boundaries. The MicroPlan is the small dynamic plan for the current objective; never replace the milestones with a flat checklist.
• In stage 4, persist a dependency-valid QuestionWorkPacket for every subquestion. Material route uncertainty requires an accepted route gate. A question is complete only after its governed checkpoint gate is accepted.
• Use hajimi_rollback when a contract, model, implementation, or validation changes. It invalidates only the affected question dependency closure, then propagates stale status through downstream evidence and publication bindings.
• Record formal computation with hajimi_record_managed_experiment, then Evidence and Claim records. Ordinary discussion may use a clearly labeled draft estimate; figures, papers, and final answers may consume only supported claims in an active evidence freeze.
• Stage 8 automatically activates HaJiMi's bundled modeling-paper-standard and modeling-plot-suite capabilities. Follow their routed stage cards and task-specific reading: preparation loads structure/layout and the relevant question route; prose production loads its applicable language and section requirements before writing; whole-paper delivery retains every review gate. Use active frozen Claim/Evidence, the native skeleton, the paper config, publication bindings and hajimi_validate_delivery together. Do not look for global Codex skills or bypass the governed workflow.

Workspace convention

Use these directories:

• input/: frozen read-only problem statement and attachments.
• src/: reproducible source code.
• work/: scratch analysis and intermediate files.
• data/derived/: generated datasets.
• figures/: generated figures.
• paper/: LaTeX and the compiled paper.
• output/: final deliverables.

HaJiMi owns .hajimi/; do not create or edit its files directly.

Checkpoints and delivery

Create a checkpoint after a material route change or a completed major experiment. Before final delivery:

1. Require a recorded successful run of the chosen pipeline on the current frozen inputs, with verified managed experiments, supported evidence/claims and an active freeze. Re-run affected computations if inputs, formulas, code or dependencies changed; reuse still-valid records for prose and layout changes. A previous run over different inputs or invalidated evidence cannot satisfy this requirement.
2. Bind every figure and paper candidate to frozen claims. For papers, set equationBaselineArtifactIds in the paper config to the pre-writing, self-contained technical formula outputs covered by active frozen Evidence; reuse valid records. Then call hajimi_validate_delivery with the current revision. Formula changes require affected technical validation and a new freeze, not replacing the baseline to match rewritten prose.
3. Stage 8 follows figures → writing → review → ready with hajimi_stage8_phase. Preserve writing standards and evidence. Resolve serious validation errors; record advisory language/layout issues for human review without mandatory rework loops.
4. Call hajimi_request_gate with gate="final_delivery" and a concise evidence-backed summary.

The files under references/ are the approved MathModeling-AI-Lab capability snapshot. Read only the reference needed for the current action.