| name | resmax-idea |
| description | Compile a validated ROI-aware ResearchPack into evidence-grounded IdeaCards, lineage, closest-work checks, experiment planning inputs, and structured negative memory. |
resmax-idea
When To Use
Use this skill after resmax-survey has produced either a validated normalizer contract or a legacy ROI-aware research_pack/.
This is not a topic brainstorm skill. It is a deterministic compiler from structured evidence to structured candidates. For the normalizer-first path, ideas must be grounded in downstream/survey_contract.json, claim_graph, gap_map, closest-work candidates, paper assets, EvidenceCards, and missing-evidence records. For the legacy path, ideas must be grounded in gap_map.json, roi_lens.json, paper_roles.json, reviewer pressure notes, and EvidenceCards.
Production/default execution is interactive. Stop before changing the idea portfolio, experiment budget, execution permission, or long-term memory. Full non-interactive execution is allowed only when the user explicitly says test/dev/debug/smoke; persistent writeback still requires explicit opt-in flags such as --confirm-write.
Input
Preferred normalizer contract:
literature_research/<topic>/downstream/survey_contract.json
If the contract exists, read it first and follow its paths for verified paper set, claim graph, gap map, closest-work candidates, paper assets, asset stats, evidence cards, and missing evidence. If it does not exist, fall back to the legacy pack:
literature_research/<topic>/research_pack/
The legacy pack must contain the manifest, claim/gap/evidence artifacts, reviewer-pressure artifacts, ROI lens, and idea seed constraints, and must pass:
python3 .agents/skills/_shared/resmax_core/validators/validate_research_pack.py \
--pack literature_research/<topic>/research_pack
Optional negative memory defaults to resmax_memory/negative_memory.jsonl; missing memory is recorded as memory_status=not_found and is not an error. Use --negative-memory <path> to override it.
Input trust boundary:
verified_fact: accepted-index metadata or materialized source evidence only.
external_claim: external report or seed assertion; never promote directly.
model_inference: deterministic/model extraction or classification; cite provenance and lower confidence.
missing_evidence: blocking or warning record; blocking records prevent review-ready idea generation.
Do not expand this skill into survey normalization or retrieval. If the contract says a gap lacks closest-work candidates or blocking missing evidence is open, keep it out of review-ready idea generation.
Linear Workflow
Generate and validate the idea portfolio:
PYTHONPATH=.agents/skills/resmax-idea/scripts python3 -m resmax_idea generate \
--pack literature_research/<topic>/research_pack \
--out literature_research/<topic>/ideas
PYTHONPATH=.agents/skills/resmax-idea/scripts python3 -m resmax_idea validate \
--ideas literature_research/<topic>/ideas
Show ideas/idea_cards.jsonl, idea_report.md, closest-work checks, and cheapest falsification notes. Ask which ideas enter review, whether to add human seeds, and which ideas to discard.
Compile experiment planning artifacts only from reviewed ideas:
PYTHONPATH=.agents/skills/resmax-idea/scripts python3 -m resmax_idea compile-experiment-plan \
--pack literature_research/<topic>/research_pack \
--ideas literature_research/<topic>/ideas \
--reviews literature_research/<topic>/reviews \
--out literature_research/<topic>/experiment_plan
Stop after writing the experiment planning package. Do not run training, evaluation, code edits, paper-result claims, or mark a block executable until the user approves budget, baseline, dataset, and metric.
Write negative memory only after explicit user approval:
PYTHONPATH=.agents/skills/resmax-idea/scripts python3 -m resmax_idea write-negative-memory \
--reviews literature_research/<topic>/reviews \
--experiment-plan literature_research/<topic>/experiment_plan \
--memory resmax_memory \
--confirm-write
Outputs
ideas/: manifest, idea_cards.jsonl, lineage, closest-work checks, rejection/falsification notes, generation trace, rendered report
experiment_plan/: manifest, blueprint, falsification plan, baseline/metric contracts, ablation/visualization plans, risk register, human gate, claim-to-experiment matrix
resmax_memory/: negative memory, reviewer blockers, failed gap paths, infeasible experiments
idea_report.md is display-only and must be rendered from structured artifacts.
Contracts
- Full field contracts live in
.agents/skills/_shared/resmax_core/schemas/idea_card.schema.json, experiment_blueprint.schema.json, and negative_memory.schema.json.
- Prefer
literature_research/<topic>/downstream/survey_contract.json when present; otherwise use legacy literature_research/<topic>/research_pack/.
- Every non-speculative IdeaCard must cite
source_gap_ids and evidence_ids; otherwise it must be speculative or insufficient_evidence.
- Every IdeaCard must preserve the distinction between verified fact, external claim, model inference, and missing evidence.
topic_direct is forbidden and validator-failing.
- Cards without
closest_work_ids are not ready for review.
- Gaps marked not
downstream_ready in the survey contract are not ready for review.
- Cards without
direct_baselines are not ready for executable experiment planning.
- Experiment planning consumes
resmax-review outputs; a promoted idea without raw ReviewTrace coverage must not enter an experiment block.
- Missing baseline, dataset, or metric contracts produce
insufficient_evidence / follow-up plans, not executable experiment blocks.
- Complete experiment contracts produce an approval gate plan, not permission to execute.
write-negative-memory is append-only and deduplicated by dedupe_key; superseding old judgments requires appending a new structured record.
- Negative memory stores structured failure reasons only. Do not write secrets, tokens, personal identifiers, raw prompts, or raw model responses.
Boundaries
- Do not generate ideas directly from a topic or paper list.
- Do not promote or reject by tournament;
resmax-review owns heterogeneous review, blocker aggregation, and final promotion.
- Do not create
resmax-review, execute experiments, run training, edit research code, write papers, or claim empirical results.
- A future experiment execution skill must consume validated
experiment_blueprint.json; it must not start from free-form idea text.