mathodology-agent-pipeline

Name: mathodology-agent-pipeline
Author: sweetcornna

36stars5forks

How to install mathodology-agent-pipeline?

sweetcornna/mathodology/mathodology-agent-pipeline

$npx -y skills add sweetcornna/mathodology --skill mathodology-agent-pipeline

Installs into the current project.

Summary

Use when maintaining archived knowledge about Mathodology's former Python agent pipeline, worker roles, prompts, Coder execution, HMML, MATLAB, search, or critic behavior.

Files · 1

View on GitHub

SKILL.md

1	# Mathodology Agent Pipeline Archive
2
3	## Scope
4
5	This skill preserves architectural knowledge about the former Mathodology Python worker and multi-agent pipeline.
6
7	The current GitHub branch does not contain the worker source. Use this skill to update archived guidance, explain the former design, or help reconstruct context from history. Do not instruct agents to edit or test missing worker files in this checkout.
8
9	## Archived Concepts
10
11	The former worker coordinated specialized modeling agents around a mathematical modeling run:
12
13	- problem interpretation and task decomposition
14	- literature and evidence search
15	- model selection and method grounding
16	- code execution for numerical work
17	- draft generation and revision
18	- critique, audit, and final evidence checks
19
20	The worker also carried knowledge about HMML-style method retrieval, MATLAB or Octave execution, web and scholarly search tools, figure generation, and runtime skills used by the Coder role.
21
22	## New Orchestration Use
23
24	This skill now also carries the reusable workflow pattern that replaced the removed worker source.
25
26	The gate schemas (`handoff` / `gate` / `scorecard` / `decision_memo`), the judge-panel protocol, and the figure/PDF QA scripts live in `mathodology-award-gates`; reference that skill rather than restating those contracts here.
27
28	For Claude Code:
29
30	- Use `.claude/workflows/mathodology-award-submission.md`.
31	- Use `.claude/workflows/mathodology-contest-variants.md` when the contest is not the default MCM/ICM or CUMCM-style paper workflow.
32	- Dispatch `.claude/agents/mathodology-*.md` roles.
33	- Lead owns phase gates and synthesis.
34
35	For Codex:
36
37	- Use multi-agents mode.
38	- Dispatch separate agents for problem analysis, evidence, model design, experiments, critique, writing, and packaging.
39	- Keep a phase log and run an independent critic gate before advancing.
40	- Treat execution as phase-sized and resumable. Ask the user only for contest-critical details, then continue automatically when the gate passes.
41	- In Phase 0, classify the contest type and apply the adapter from `docs/WORKFLOWS.md`.
42
43	## Human Confirmation Checkpoints
44
45	The workflow should not stop for every ambiguity. Most modeling details can be handled by conservative, explicit assumptions and later critic review.
46
47	Ask the user only when the answer would change official requirements, private data access, model route selection, compute or reproducibility constraints, or final submission decisions. The question should state the current phase, why the detail blocks progress, the recommended default, and the consequence of common answers.
48
49	For all other ambiguity, record the assumption in the phase log and proceed. If Codex must stop because a response boundary is reached, preserve the current phase, completed gates, unresolved risks, artifact paths, and next action so the next response resumes rather than restarts.
50
51	## Phase Responsibilities
52
53	- Phase 0: problem analyst maps prompt clauses, scoring surface, contest type, deliverables, official constraints, dependencies, and material ambiguities; critic checks requirement coverage, adapter selection, and whether user questions are truly contest-critical.
54	- Phase 1: evidence researcher builds source ledger, data dictionary, proxy logic, benchmark inventory, citation plan, and evidence gaps; critic checks traceability and proxy defensibility.
55	- Phase 2: at least two model agents propose alternatives; lead compares at least three routes and selects one with rejection reasons; critic checks route fit, novelty, time feasibility, and absence of generic method stacking.
56	- Phase 3: modeler writes notation, assumptions, units, objectives, constraints, algorithms, pseudocode, validation metrics, baseline, ablation, sensitivity, and robustness plan; critic checks implementability and mathematical coherence.
57	- Phase 4: coder produces reproducible computation, environment notes, raw outputs, tables, figures, baseline, ablations, sensitivity, robustness outputs, and run logs; the figure/table set must cover model structure, primary results, sensitivity, robustness, decision tradeoffs, and final recommendations; critic checks number traceability, source data, density, and no cherry-picking.
58	- Phase 5: modeler and paper editor translate results into prompt-level answers, captions, recommendations, limitations, uncertainty notes, claim-source links, and a figure/table coverage map; critic checks answer coverage, support, and whether visual and tabular evidence is sufficient for a top-tier paper.
59	- Phase 6: paper editor builds summary, coherent paper narrative, references, appendix, AI-use statement when required, and final figure/table placement; critic checks summary quality, paper coherence, notation, citations, page/format risk, and whether r

People who liked this also viewed

6 picks

Type

mathodology-agent-pipeline

36stars5forks

View on GitHub

How to install mathodology-agent-pipeline?

sweetcornna/mathodology/mathodology-agent-pipeline

$npx -y skills add sweetcornna/mathodology --skill mathodology-agent-pipeline

Installs into the current project.

Summary

Use when maintaining archived knowledge about Mathodology's former Python agent pipeline, worker roles, prompts, Coder execution, HMML, MATLAB, search, or critic behavior.

Files · 1

View on GitHub

SKILL.md

1	# Mathodology Agent Pipeline Archive
2
3	## Scope
4
5	This skill preserves architectural knowledge about the former Mathodology Python worker and multi-agent pipeline.
6
7	The current GitHub branch does not contain the worker source. Use this skill to update archived guidance, explain the former design, or help reconstruct context from history. Do not instruct agents to edit or test missing worker files in this checkout.
8
9	## Archived Concepts
10
11	The former worker coordinated specialized modeling agents around a mathematical modeling run:
12
13	- problem interpretation and task decomposition
14	- literature and evidence search
15	- model selection and method grounding
16	- code execution for numerical work
17	- draft generation and revision
18	- critique, audit, and final evidence checks
19
20	The worker also carried knowledge about HMML-style method retrieval, MATLAB or Octave execution, web and scholarly search tools, figure generation, and runtime skills used by the Coder role.
21
22	## New Orchestration Use
23
24	This skill now also carries the reusable workflow pattern that replaced the removed worker source.
25
26	The gate schemas (`handoff` / `gate` / `scorecard` / `decision_memo`), the judge-panel protocol, and the figure/PDF QA scripts live in `mathodology-award-gates`; reference that skill rather than restating those contracts here.
27
28	For Claude Code:
29
30	- Use `.claude/workflows/mathodology-award-submission.md`.
31	- Use `.claude/workflows/mathodology-contest-variants.md` when the contest is not the default MCM/ICM or CUMCM-style paper workflow.
32	- Dispatch `.claude/agents/mathodology-*.md` roles.
33	- Lead owns phase gates and synthesis.
34
35	For Codex:
36
37	- Use multi-agents mode.
38	- Dispatch separate agents for problem analysis, evidence, model design, experiments, critique, writing, and packaging.
39	- Keep a phase log and run an independent critic gate before advancing.
40	- Treat execution as phase-sized and resumable. Ask the user only for contest-critical details, then continue automatically when the gate passes.
41	- In Phase 0, classify the contest type and apply the adapter from `docs/WORKFLOWS.md`.
42
43	## Human Confirmation Checkpoints
44
45	The workflow should not stop for every ambiguity. Most modeling details can be handled by conservative, explicit assumptions and later critic review.
46
47	Ask the user only when the answer would change official requirements, private data access, model route selection, compute or reproducibility constraints, or final submission decisions. The question should state the current phase, why the detail blocks progress, the recommended default, and the consequence of common answers.
48
49	For all other ambiguity, record the assumption in the phase log and proceed. If Codex must stop because a response boundary is reached, preserve the current phase, completed gates, unresolved risks, artifact paths, and next action so the next response resumes rather than restarts.
50
51	## Phase Responsibilities
52
53	- Phase 0: problem analyst maps prompt clauses, scoring surface, contest type, deliverables, official constraints, dependencies, and material ambiguities; critic checks requirement coverage, adapter selection, and whether user questions are truly contest-critical.
54	- Phase 1: evidence researcher builds source ledger, data dictionary, proxy logic, benchmark inventory, citation plan, and evidence gaps; critic checks traceability and proxy defensibility.
55	- Phase 2: at least two model agents propose alternatives; lead compares at least three routes and selects one with rejection reasons; critic checks route fit, novelty, time feasibility, and absence of generic method stacking.
56	- Phase 3: modeler writes notation, assumptions, units, objectives, constraints, algorithms, pseudocode, validation metrics, baseline, ablation, sensitivity, and robustness plan; critic checks implementability and mathematical coherence.
57	- Phase 4: coder produces reproducible computation, environment notes, raw outputs, tables, figures, baseline, ablations, sensitivity, robustness outputs, and run logs; the figure/table set must cover model structure, primary results, sensitivity, robustness, decision tradeoffs, and final recommendations; critic checks number traceability, source data, density, and no cherry-picking.
58	- Phase 5: modeler and paper editor translate results into prompt-level answers, captions, recommendations, limitations, uncertainty notes, claim-source links, and a figure/table coverage map; critic checks answer coverage, support, and whether visual and tabular evidence is sufficient for a top-tier paper.
59	- Phase 6: paper editor builds summary, coherent paper narrative, references, appendix, AI-use statement when required, and final figure/table placement; critic checks summary quality, paper coherence, notation, citations, page/format risk, and whether r

People who liked this also viewed

6 picks

Type