Idea Spark Skill

Convert an under-specified research direction into ONE reviewer-defensible Oral-level research proposal — grounded in 1947 ICLR/ICML/NeurIPS papers (2021-2025) — via a 5-phase workflow: retrieve recent literature, diagnose the bottleneck, select + generate a candidate using corpus-derived ideation pattern cards, run it through a quality gauntlet, expand into an idea card.

This file is the operational runbook. Design rationale (the 7 design principles, why each contract is shaped this way, removed-check history) lives in references/design-notes.md — read it only when modifying or evaluating the skill, never needed to run it. When MODIFYING the skill, also replay the cross-shape regression set in references/regression-directions.md (deterministic subset: python3 "$SKILL_DIR/scripts/regression_check.py" <run_dir>; routing/merger branch fixtures — including every guard/retry branch that real runs rarely exercise: python3 "$SKILL_DIR/scripts/selftest_routing.py"). First-time installation lives in references/setup.md.

When to use

• "Give me a research idea in {area} I could pursue." / "What's the most impactful next step in this direction?"
• "Help me sharpen this vague direction into an Oral-level proposal."
• "What's the bottleneck of this problem?" / "Run a novelty audit on this idea."

When NOT to use

• Code review, debugging, refactoring. Summarizing one paper. Cross-decade survey writing.
• Free-association brainstorming with no research context. Engineering integration tasks ("ship this feature in our system").
• Pure benchmark / dataset construction work — the 15-pattern vocabulary handles benchmark audit (controlled_diagnostic_design) but not benchmark construction.

Setup (first use only)

Follow references/setup.md. Quick version — set two shell variables, install deps, verify:

SKILL_DIR=<path to this folder>                  # e.g. ~/.claude/skills/idea-spark (Claude Code), ~/.codex/skills/idea_spark (Codex CLI), or any clone location
RUN_DIR="$PWD/ideaspark_run/<topic-slug>" && mkdir -p "$RUN_DIR"   # convention below; any absolute dir works
python3 -m pip install feedparser openreview-py beautifulsoup4 pymupdf
python3 "$SKILL_DIR/scripts/run.py" check_connectors   # from the SAME shell you'll run phases from

Credentials go in .env (OpenReview user/pass + Semantic Scholar key — see setup.md); the orchestrator auto-loads it. Optional: xelatex/tectonic for PDF cards.

How to run: the next loop

The canonical way to drive a run is the run-state navigator:

python3 "$SKILL_DIR/scripts/run.py" next --dir "$RUN_DIR" --query "<user's research question>"

Run-dir convention (one run = one directory, named by the host BEFORE the first command): $PWD/ideaspark_run/<topic-slug> — a short kebab-case slug distilled from the user's direction (e.g. ideaspark_run/diffusion-watermark); if the slug is taken, append _2, not a timestamp. NEVER reuse a directory that already contains a phase0/ — every phase writes into $RUN_DIR and would clobber the prior run. The skill itself never names the dir (any absolute path works); this convention exists so runs from different harnesses land in one predictable place instead of each agent improvising.

next inspects the artifacts already on disk and prints EXACTLY one next step — either a Bash command to run verbatim, or an LLM sub-agent spec (system-prompt path + input file paths + output path + the routing signal to report back). It is read-only and idempotent (safe to re-run anytime, including to resume an interrupted run). The host loop is:

1. Run next.
2. Do what it says (bash → run the command; llm_subagent → execute in an ISOLATED context per the Context discipline rules below).
3. Run next again. Repeat until it reports a terminal state (DONE, do_not_generate, or phase_3_failed).

Consume every emit WHOLE — never grep/filter/truncate the block. INPUT lists span unlabeled continuation lines; a label-keyed grep (grep -E "STEP|INPUT|...") silently drops them. This exact failure occurred in a live run: the navigator listed the coherence gate's blocking findings as an audit input, the host's grep dropped the line, and the audit issued a false advance without ever seeing the executed evidence. If the emit must be captured into a bounded tool result, head -c 4000 the WHOLE block — never filter by line labels.

next encodes the full phase graph — the mandatory full-text gate, the citation gate, the abandon-retry branch, the falsification re-audit branch, and the correct Phase 4 flags per path — so you do not need to memorize the reference tables below; they exist for deviation and debugging.

If your host exposes a task/todo tool (e.g., TodoWrite), seed it with this checklist and tick phases as next moves past them:

- [ ] Phase 0: Literature grounding → lit_table.md, then Phase 0+ full-text fetch (MANDATORY — Phase 1 hard-gates on it)
- [ ] Phase 1: Bottleneck identification → phase1_output.json (routing: proceed | do_not_generate)
- [ ] Phase 2: Gap×pattern selection + candidate generation (ONE isolated context, TWO output files) → citation gate → coherence gate (dry-run trace, fresh context)
- [ ] Phase 3: Collision retrieval (signature@10mo + alias@48mo — launchable in parallel with 2.3) → audit (5 checks) → [revise → merge → re-audit if falsification rewritten] | [abandon → information-gain retry: regenerate while each failure yields NEW binding lessons, ≤3 candidate cycles; repeated subsumption lesson → 1 bottleneck re-diagnosis + 1 attempt; no new information or cap → phase_3_failed]
- [ ] Phase 4: skeleton → fill (technical) → assemble partial → derive (plain, fast-tier) → assemble final + method view → implementability audit → validate → render → return 3 cards inline

Three outcomes per run: the rendered idea-card markdown returned inline (LaTeX + per-phase JSON left under $RUN_DIR), a do_not_generate.md (Phase 1 OOD), or a phase_3_failed.md (audit abandons twice). Never ask the user mid-flow — missing intake fields are inferred; revision, falsification re-audit, and the single internal retry all run without user re-invocation.

Invocation contract

No cd is required. scripts/run.py self-locates its skill root, so every orchestrator command can be invoked from ANY working directory by absolute script path: python3 "$SKILL_DIR/scripts/run.py" <subcommand> --out "$RUN_DIR/<phase>/" .... The legacy form cd "$SKILL_DIR" && python3 -m scripts.run <subcommand> ... works identically. Do NOT use relative script or --out paths — CWD is not stable across host-LLM Bash invocations, and the orchestrator rejects a relative --out outright.

Exit codes 10 and 11 are NOT errors — they are sentinel handshakes. When the orchestrator can't call an LLM itself (no NOVELTY_LLM_CLASSIFY_FAST_CMD), it writes a sentinel JSON describing what the host LLM should do, then exits rc=10 (intent / pattern-summary) or rc=11 (signature_terms). Read the sentinel ($RUN_DIR/<phase>/.<step>_pending), read the file at its rubric_file field (absolute path), produce the expected output, re-invoke per its re_invocation field. Do not stop on these codes. (The default Phase 0 flow below avoids the rc=10 intent sentinel entirely by passing --queries up front.)

Context discipline (read BEFORE running any LLM-driven phase)

A full run accumulates ~180-250k tokens of intermediate state. If the host LLM carries that in its own conversation context across phases, the Phase 1 / 2.2 / 4.fill calls routinely hit the backend request timeout ([API Error · Request timed out · Retrying...]) and the retry times out again. Apply ALL three rules on every run:

Rule 1 — Run every LLM-driven phase in an ISOLATED context. Phases 1 / 2 (2.1+2.2) / 2.3 / 3.2 / 3.3 / 4.fill / 4.1.5 each have file-path inputs and one JSON output; no phase needs the conversation that produced an earlier one. Use the FIRST isolation mechanism your harness supports:

• (a) Subprocess LLM — set NOVELTY_LLM_REASONING_LARGE_CMD / NOVELTY_LLM_CLASSIFY_FAST_CMD (see § Configuration); each phase runs as its own subprocess, fresh context by construction, on any harness.
• (b) Sub-agent tool (Claude Code Agent or equivalent) — spawn one per phase, passing ONLY the file paths the phase prompt lists — not conversation history, not file contents inline. The sub-agent reads from disk, Writes to disk, returns ≤ 250 words (output path + routing signal). Exception by design: Phase 2.1 and 2.2 run in ONE sub-agent writing both output files — both are generation-side; the adversarial separations (3.2 vs 3.3, 4.fill vs 4.1.5) must stay separate calls.
• (c) Manual context reset — run inline but clear/compact at the four points in Rule 3.

Whichever mechanism, the parent context stays ≤ ~30k tokens for the whole run because it never holds a phase's structured output.

Rule 2 — Write every phase artifact directly to disk; never paraphrase it into chat. Output convention: $RUN_DIR/<phase>/<phase>_output.json. Use your harness's file-write tool (Claude Code: Write) — no Bash heredocs (permission prompts + silent truncation), no echo, no pasting JSON into replies. Bound tool-result captures from large files to ≤ 4 KB (head -c 4000 / jq / sed); never Read a >10 KB intermediate dump into the parent context — the dump gets cached into every subsequent turn (this exact anti-pattern caused prior timeout runs).

Rule 3 — Compact between phases. Natural compact points: after Phase 0+, after Phase 1, after Phase 2, after Phase 3.2. Every phase re-reads its disk inputs, so compacting loses nothing. With /compact, use it there; Rule 1 mechanisms (a)/(b) achieve the same on their own.

Diagnostic for "Request timed out" mid-phase: inspect your harness's session transcript/log (Claude Code: ~/.claude/projects/<project-slug>/<session-id>.jsonl, look for isApiErrorMessage: true; other harnesses: their session-log equivalent); the prior tool call shows which prompt got too big. The fix is one of the three rules — usually Rule 1.

Phase reference

next prints each of these steps at the right moment with concrete paths; the tables below are the full contract for deviation/debugging.

Orchestrator entry points

The LLM-driven phases (1 / 2.1 / 2.2 / 2.3 / 3.2 / 3.3 / 4.fill / 4.1.5 / falsification re-audit) have no orchestrator subcommand (a cat prompt | llm wrapper would add fragility without determinism): read the prompt at references/system-prompts/<phase>.txt, gather the inputs listed at its top, Write the JSON described under Output: to $RUN_DIR/<phase>/<phase>_output.json. Run each under the Context discipline rules — Phase 4.fill is the largest output and the most timeout-prone; never in the parent context.

Phase 0 — Literature grounding

Phase 0 and 3.1 require real external retrieval via the bundled connector scripts (scripts/search_*.py) — never WebSearch or ad-hoc fetch (downstream phases reject unstructured output). Gate sentinel: .lit_grounding_mode = real vs connector_failure (halt with diagnostic; --allow-webfallback exists as a flagged, lower-confidence escape).

Default flow (skips one sentinel round-trip): BEFORE invoking phase0, read references/intent-recognition.md (Map mode) yourself and produce 4-6 search queries — including one ESCAPE-MECHANISM query phrased in solution vocabulary (recalls papers that already fixed the bottleneck and title themselves by their fix; problem-keyed queries miss exactly those). Phrase queries mechanism-first: generic "<topic> challenges/overview/landscape" phrasings are survey magnets that dilute corpus density — the orchestrator demotes survey-titled hits to the bottom of lit_results (kept, never dropped), but each connector's cap slots are still spent on them. Also apply the OOD short-circuit (intake-routing.md triggers #1 Too-broad / #2 No-anchor → route to do_not_generate instead of retrieving). Then invoke with BOTH flags:

python3 "$SKILL_DIR/scripts/run.py" phase0 --query "<user's research question>" --queries "q1|q2|q3|q4" --out $RUN_DIR/phase0/

The rc=10 sentinel path still exists as fallback when --queries is omitted. The orchestrator: asserts a sane clock; probes 4 connectors and retrieves role-based (arxiv 0-6mo cap 10 / openalex 6-24mo cap 12 published-only / semanticscholar 6-24mo cap 13 published-only / openreview 0-6mo cap 10 in-review; ~40-45 papers; non-overlapping windows; SS-priority dedup); extracts URL/ID user-refs from the query into phase0/user_refs.json; emits .pattern_summary_pending for the host.

Retrieval takes 3-10 min (the openreview connector alone budgets 600s) — set your Bash timeout ≥ 600s or run it in the background.

Pattern tagging (host step): classify each lit_results.json paper per references/pattern-summary-rubric.md into 1-3 of the 15 patterns → write lit_table.md with columns paper_id | year_month | venue | title | ideation pattern tags | bottleneck this paper targets | open issue / unresolved gap | resolves_problem | retrieved_via. Pure classification — it does not need the large reasoning model: route it to a cheaper/faster model tier or a lower reasoning effort BY DEFAULT (the same tier NOVELTY_LLM_CLASSIFY_FAST_CMD names in § Configuration; spending the large model here is pure waste — a real run measured ~85k large-model tokens for 33 papers); only when no cheaper tier exists, run it isolated on the host model. Rows are per-paper independent, so for 40+ papers the tagging MAY be sharded across 2-3 parallel fast-tier sub-agents (contiguous slices; assemble with the deterministic lit_table_merge --out $RUN_DIR/phase0/ --shards <row-file>... — it validates 9-column shape, row count == paper count, and paper_id coverage against lit_results.json, replacing hand-checks) — wall-clock win, no quality delta.

Title-named user refs: if the user query names anchor papers by TITLE ("based on the LoRA paper" — anything the URL/ID regex can't catch), register each BEFORE phase0_fulltext via add_user_ref (entry-point table). It does a deterministic dedup-merge into user_refs.json — do NOT hand-edit that file (some harnesses' file-write tools refuse to overwrite a file that was never read, and a malformed edit silently drops the U fetch tier).

Phase 0+ full-text fetch — MANDATORY. The instant lit_table.md lands, run phase0_fulltext (entry-point table) before touching Phase 1. Pool = U (user refs, never capped) + T2 (top-10 published on-topic) + T3 (top-5 arxiv on-topic), ceiling 15 excluding U, method-first ordering, concurrent fetch (HTML path first, pymupdf PDF fallback; per-paper budget so one slow PDF can't stall the step). Output fulltext_cache.json keyed by paper_id ({tier, intro, method, source_used, warning}); fetch failures degrade to abstract + warning. Phase 1 hard-gates on this file (error: fulltext_not_fetched). Alongside the blob it writes a derived per-paper read view (phase0/fulltext/index.json + one .md per paper, same content, index carries tier/source_used/warning) — Phase 1 reads the index then only the papers it needs; the blob stays canonical and phase1_fulltext_topup refreshes both. Fetches hit a cross-run content cache (~/.cache/ideaspark/fulltext/, successful fetches only, 30-day TTL; disable with IDEASPARK_FETCH_CACHE=off), so papers recurring across runs on adjacent topics cost zero network time.

Phase 1 — Bottleneck identification

One isolated LLM call. Prompt: references/system-prompts/bottleneck_identify.txt. Inputs: user query + intake, phase0/lit_table.md, phase0/fulltext_cache.json (all-failed cache → continue with fulltext_degraded: true, abstract-level residue confidence), phase0/lit_results.json. Output phase1/phase1_output.json: intake (+_inferred_fields[] — missing fields are inferred, never asked), bottleneck_statement (≥2 paper_id cited inline), closest_adjacent[] ({paper_id, summary_and_residue}), what_phase_0_did_not_address[] (each gap ends with an inline stakes clause — practitioner costs and intellectual costs are coequal; gaps are framed as structural properties of a problem class when honestly true, with the anchor as primary instance, else declared class-of-one), state ∈ {proceed, do_not_generate}.

Routing: proceed (literature-groundable, no OOD trigger) or do_not_generate (too-broad / no-anchor OOD, <5 truly-relevant papers, genuinely blank space, or benchmark/system construction) → write do_not_generate.md with concrete remedial steps — terminal.

Phase 2 — Selection + generation (ONE isolated context, TWO outputs)

Run 2.1 and 2.2 back-to-back in one isolated context, writing BOTH output files (they are both generation-side; only adversarial pairs need separate calls):

2.1 — prompt references/system-prompts/ideate_select.txt; inputs phase1_output.json, references/ideation-patterns/overview.md (all 15 patterns' Definition / Operational signature / When to apply — selection at WHAT/WHEN level), references/ideation-patterns/companion-combos.md, lit_table.md. Pick the anchor gap (type-bound to intake.contribution_type); commit the PATTERN COMPOSITION — ≥2 distinct patterns by default, realized preferentially as a CHAIN on the anchor (attested second pattern from companion-combos, with a named intermediate object) and/or via a sibling that EARNS its seat through the removal test ("the anchor's story is incomplete without it") — anchor-only is legitimate and common (1-3 gaps total, ≤2 siblings); a single-pattern selection requires a composition_note defense (validator presence-checks it; audit weighs it). Record saturation (transparency, not a filter). Output phase2_select/phase2_select_output.json: selected_gaps[] (index 0 = anchor) + coherence_thread_type (or n_a when anchor-only) + composition_note + pattern_saturation + deferred_gaps[]. Retry mode: when $RUN_DIR/attempt_1/ exists, the prompt's OPTIONAL retry input applies — the archived audit + selection become negative constraints; anchor-only is a valid retry outcome. Cross-run dedup: next scans sibling run dirs and appends a soft-negative-anchor input line (titles + signature terms of their canonical candidates, 5 most recent) so adjacent-direction runs don't silently re-invent the same mechanism family; soft by design (see ideate_select.txt's OPTIONAL cross-run input), disable with IDEASPARK_CROSS_RUN_DEDUP=off.

2.2 — prompt references/system-prompts/ideate_generate.txt; inputs 2.1 output, phase1_output.json, lit_results.json (closest_adjacent entries ONLY — filter to those paper_ids before reading; the prompt forbids pulling the full dump into context), plus for each gap ONE picked sub-pattern card from references/ideation-sub-patterns/ (compare when_to_pick_this_one + differentiation_within_parent via its overview.md; then read the picked card's tactical_pattern + Step-by-Step). Output phase2_generate/phase2_generate_output.json — ONE candidate, 12 flat fields: title / hook / core_mechanism / core_mechanism_reasoning / core_mechanism_steps; gap_closure[] ({gap, main_pattern, sub_pattern: "C## (parent pattern name)", how_closed}, mirrors selected_gaps one-for-one); falsification_prediction (single paragraph: minimal experiment + metric-with-direction + ONE named load-bearing variable + negative control on that variable predicting the DOWNSTREAM outcome metric returns to baseline — non-tautological); compute_budget (user-relative, GPU-day line + API-dollar line when the campaign calls paid APIs; default intake envelope = 80GB-class GPUs, ≤8 concurrent, ≈150 GPU-days / 5 months, ~$10k API — overridable per user via IDEASPARK_DEFAULT_COMPUTE, see § Configuration); differentiation_from_lit[] (substantive deltas, not "different pattern"); almost_prior_paper_id + what_step_was_missed; signature_terms[] (own vocabulary — recent collision channel); alias_terms[] (other communities' names for the same mechanism, from parametric knowledge — multi-year alias collision channel); composition_note (echoed verbatim from 2.1). Reality constraints at writing time: core_mechanism_reasoning carries three mandatory blocks — a PREMISES ledger (load-bearing premises + domain justification; bets tagged untested — falsification target; WHEN the mechanism consumes sampled/observed data, at least one OBSERVATION-MODEL premise — how are the inputs sampled, is that unbiased here — else the literal observation-model: n/a; WHEN the mechanism estimates any quantity, declare the ESTIMAND precisely), a NAIVE-BASELINE AUDIT (state the naive version; why doesn't it already work — branch (i) naive relies on a false premise → confronting it IS the contribution, with a STANDARD-TOOL FOLLOW-UP: if the confrontation is a textbook tool from another field, name the domain-specific structure that makes this instance unsolved or declare the contribution application-grade, (ii) naive suffices → incremental signal surfaced honestly, (iii) naive works but the field disbelieves → minimalism with evidence), then design rationale — plus every invoked dataset / model-access level / annotation / tool must be a nameable EXISTING artifact (hard rule 3; modest self-built resources allowed with cost counted in compute_budget) and every claim stated at its defensible strength — guarantee-grade wording only with assumptions stated where it appears (hard rule 4; the coherence gate grades this, and a strong claim with honest assumptions beats a hedged weak one). Both kill-switch fields (falsification_prediction, compute_budget) are locked from here on — see Phase 3 for the single audited exception.

Citation gate (deterministic, MANDATORY before Phase 3):

python3 "$SKILL_DIR/scripts/run.py" validate --phase2 $RUN_DIR/phase2_generate/phase2_generate_output.json

Any fail = a sub_pattern citation was guessed from the parent's gist, not read from overview.md. Fix against references/ideation-sub-patterns/overview.md (or regenerate 2.2 with the card open) and re-run until clean — the gate proves parent-consistency only; whether core_mechanism performs the cluster's actual tactic is Phase 3.2's recipe_application_check. (next runs this gate automatically.)

Coherence gate (2.3 — one isolated LLM call, MANDATORY after the citation gate, before the 3.2 audit; 3.1 collision may run CONCURRENTLY with it): prompt references/system-prompts/coherence_trace.txt; inputs: the 2.2 candidate + the 2.1 spec; MUST be a FRESH context, never the 2.1+2.2 agent (the context that wrote a logic bug rubber-stamps it). It verifies internal procedural validity by EXECUTION, not review — five trace actions: formalize the dataflow (undefined symbols, missing producers, circular deps), numeric dry-run on one small concrete instance (magnitude/probability absurdities — logic bugs read fluently and only surface when computed; when a code-execution tool exists the gate WRITES AND RUNS a stdlib Python script and pastes script+output, else hand-computes marked unexecuted), degenerate probes (empty/k=0/ties), claim→step mapping AND grading (every asserted property mapped to the step that establishes it AND graded established / conditional — assumptions listed / overclaim — wording downgraded / empirical; statistical claims settled by an executed Monte Carlo with the measured number, theorem-shaped claims recorded as proof obligations — at best conditional; grading never rewards vagueness: all-hedged claims are themselves a weakness finding), and a NAIVE-BASELINE COMPARISON (independently construct the naive version — never the candidate's own stated one — run it on the same instance, and judge against the candidate's declared branch: confronts_obstacle / equivalent_to_naive / n_a). OBSTACLE HOLES — findings whose honest fix is a redesign, or which coincide with the obstacle the candidate declares it exists to solve (incl. equivalent_to_naive) — must NOT be patched around with avoidance-style repairs (abstain/clamp/skip); they go to unrepaired[] as blocking, verbatim. Verdict pass | patched; repairs are patch-only via the SAME merger (--out-name refined_candidate.json), scoped to making the written procedure sound (core_mechanism*, how_closed narrative, signature/alias terms when the repair changed what the mechanism is) — novelty surface, pattern bindings, and kill-switch fields are out of scope; unfixable-without-redesign findings go to unrepaired[] for the audit to weigh. Single pass, never abandons. The 3.2 audit stays blind to this report's trace/verdict/patches — with ONE exception: blocking unrepaired[] entries are executed evidence, so the gate ALSO writes them to phase2_coherence/blocking_findings.json (self-contained: verbatim step quote + script excerpt + measured numbers + reading_dependence per entry), which next lists as a 3.2 input; the audit must disposition each (uphold, or refute by naming a concrete modeling/arithmetic flaw — next deterministically bounces an audit report that skipped dispositions or advanced over an upheld finding). Formalization ambiguity is controlled by the gate's dual-reading protocol (every defensible reading of an ambiguous operative term is executed; findings are tagged reading_robust vs reading_dependent). When refined_candidate.json exists it is the canonical candidate for every later phase (next wires this automatically). It validates that the algorithm survives on paper — NOT that it works empirically (falsification experiment) or is novel (audit).

Phase 3 — Quality gauntlet

3.1 collision (orchestrator, no LLM): entry-point table. May be LAUNCHED IN PARALLEL with the 2.3 gate (on the 2.2 output — next emits the background launch alongside the 2.3 step): collision reads only signature_terms[]/alias_terms[], a .collision_terms.json sidecar records the terms actually used, and next re-issues collision if a coherence patch changed those terms (rare — term repairs are scoped to mechanism-changing patches). TWO retrieval channels over all 4 connectors, merged into collision_hits.json with a per-hit collision_channel tag: signature — the candidate's signature_terms[] over a 10-month window (contemporaneous scoop risk); alias — the candidate's alias_terms[] (other communities' names for the same mechanism, produced from parametric knowledge at 2.2) over a 48-month window (renamed-ancestor risk — the "goal-conditioned success detector vs goal-image conditioned scorer" blind spot is lexical, not temporal, so widening the signature window alone cannot catch it). Missing signature_terms[] → rc=11 sentinel: produce BOTH term sets per intent-recognition.md Collision mode (terms 3-7 words each — long sentences break URL encoding), edit the candidate JSON, re-invoke. Missing only alias_terms[] → loud warning, alias channel skipped (add the field and re-run to close the blind spot). The audit-facing pool is relevance-truncated per channel (≤120 hits/channel by lexical overlap with the channel's own terms; zero-relevance BM25 noise dropped unconditionally; drops printed; untruncated pool preserved as collision_hits.full.json), so the audit can consume collision_hits.json in a few sequential Read chunks — no jq two-pass triage needed.

3.2 audit (one isolated LLM call): prompt references/system-prompts/critique.txt; inputs: candidate, 2.1 spec, lit_table.md, collision_hits.json, references/anti-patterns.md, phase2_coherence/blocking_findings.json when it exists (the 2.3 gate's executed blocking evidence — the report must disposition each entry, and advance is forbidden while one is upheld; a refuted disposition additionally triggers a bounded refutation re-check (refutation_recheck.txt, fresh call) that next requires before trusting the verdict — an invalid refutation counts as upheld and bounces the audit), and each cited sub-pattern card references/ideation-sub-patterns/<C##>.md (strip the leading code from sub_pattern; typically 1-3 cards, others NOT loaded). Five corpus-anchored checks:

Verdict is two-layer. Hard floor (LLM cannot override) → abandon: triggered Reject lesson / unmitigatable anti-pattern / exact-mechanism collision. Soft judgment otherwise → advance (only trivial borderlines; concerns surface in Phase 4's reviewer_concerns) or revise with concrete revision_targets[] (scopes: tactical / sub_pattern / falsification). verdict_rationale must cite specific check findings. The audit judges only — it never modifies the candidate.

Routing on verdict:

• advance → Phase 4 reads the 2.2 candidate directly.
• revise → 3.3 (one isolated LLM call, prompt references/system-prompts/revise.txt): reads candidate + 2.1 spec + the revision brief (phase3_revise_brief — next materializes revise_brief.json, the audit minus the bulky reject-lesson quotations; full report stays on disk for lesson-specific lookups); emits patch-only applied_revisions[] — one entry per revision_target, ops replace / append_sentence / append_items / swap_sub_pattern / rewrite_falsification, never echoes the candidate, never re-judges the verdict. Then run the merger (entry-point table, WITH --critique) → writes phase3_revise/final_candidate.json + back-injects it into the patch file. Kill-switch fields are merger-refused with ONE audited exception: a scope=falsification target from falsification_structure_check is applied via the dedicated rewrite_falsification op (authorization verified against the audit report via --critique; same experiment/metric/claim, structure repaired; max one per run). When the merger prints falsification_rewritten, run the falsification re-audit (self-contained prompt references/system-prompts/falsification_reaudit.txt, reading the phase3_falsification_view slice of final_candidate.json → phase3_critique/falsification_reaudit.json): advance → Phase 4; abandon → phase_3_failed.md. compute_budget has no revision route under any scope. No composition scope — gap-level changes route through the abandon-retry below, never through patches.
• abandon → information-gain retry rule (ONE rule, no death-type taxonomy; the one-shot guarantee bars asking the user, not internal regeneration). Every failed attempt's audit is reduced to a LESSON SET — upheld executed blocking findings (mechanism-level positive directives), unaddressable subsuming papers (mechanism-family negative anchors), triggered reject/anti-pattern/recipe lessons (generation-quality constraints). A retry must carry NEW lessons the previous generation did not have: the first abandon always retries (generation ran with zero audit information); a later abandon retries only if its lesson set adds entries beyond the union of all archived attempts' lessons, with the new lessons injected as directives. A REPEATED unaddressable-subsumption lesson across attempts binds at the FRAMING level instead → one bottleneck re-diagnosis (Phase 1 retry mode; do_not_generate remains a legitimate exit), whose new framing gets exactly ONE candidate attempt. Termination: no new lessons, or the candidate-cycle cap (3 per framing; worst case 4 gauntlet cycles per run). next computes all of this deterministically from the archived critique reports. On the first abandon, archive the attempt and regenerate —

  mkdir -p "$RUN_DIR/attempt_1" && \
  mv "$RUN_DIR/phase2_select" "$RUN_DIR/phase2_generate" "$RUN_DIR/phase2_coherence" "$RUN_DIR/phase3_collision" \
     "$RUN_DIR/phase3_critique" "$RUN_DIR/phase3_revise" "$RUN_DIR/attempt_1/" 2>/dev/null; \
  touch "$RUN_DIR/.retry_used"
  

  
  then re-run Phase 2 in retry mode (archived audit + selection = negative constraints; blocking obstacle findings = POSITIVE directives the new mechanism must confront), citation gate, 3.1, 3.2. Phase 0/1 artifacts are reused as-is.
  Later abandons → the same rule, deterministically (next compares this attempt's lesson set against the union over attempt_N/ archives): new mechanism-level lessons → directed retry to the next attempt_N (all archived audits + selections ride along as constraints; the fresh lessons are named in the emit); repeated subsumption lesson → bottleneck re-diagnosis (archive incl. phase1/, touch .bottleneck_retry_used; the re-diagnosed framing gets ONE attempt); no new lessons / cap reached / post-bottleneck failure → write phase_3_failed.md citing EVERY attempt's verdict_rationale + triggering checks + user-side options — terminal.

Phase 4 — Expansion + packaging

Six steps in order (next emits each with the correct flags for the advance vs revise path — on the revise path --candidate is final_candidate.json and --phase3-revise is passed; on advance it's the CANONICAL candidate (refined_candidate.json when 2.3 patched, else the 2.2 output) and the flag is omitted):

1. skeleton (orchestrator): populates every mechanical field — kill-switch echoes (byte-identical from the candidate), differentiation_from_lit venue_years, almost_prior_venue_year, why_prior_stopped[].paper_id/venue_year, domain_landscape (pattern_distribution + candidate_uses), literature_breakdown, reviewer_concerns_and_responses[].attack/severity/fields_changed_to_address (lifted from audit + patch), feasibility_validation.compute (bucketed against intake.compute) — and marks every prose field <TODO[path]: hint>.
2. fill (one isolated LLM call, prompt references/system-prompts/expand.txt): author the TECHNICAL TODO paths as one flat {path: value} map → phase4/fill_map.json. The derive-owned paths (title_zh + all plain_*) are explicitly EXCLUDED — the derive step owns them. No calendar projections; no experiment matrix / ablation plan / baseline table — the skill produces IDEA + falsifiability + feasibility judgment, not experimental engineering.
3. assemble (partial) → derive → assemble (final): the orchestrator assembles the technical map first (phase4_expansion.json, derive-owned placeholders remain — the WARN is expected); then derive (one isolated LLM call, prompt references/system-prompts/derive_plain.txt) mechanically rewrites the finished technical fields into the plain register (title_zh + plain_* → phase4/derive_map.json) — register transformation + translation with NO new facts, so it runs on the CLASSIFY_FAST tier by default (fallback ladder: NOVELTY_LLM_CLASSIFY_FAST_CMD → host cheap model → host model isolated; even the last rung keeps it cheap since the input is only the finished prose); then the final assemble merges both maps (phase4_assemble --fill-map <tech> --fill-map <derive> — overlapping paths are a hard error) and extracts phase4/method_view.json (phase4_method_view — the method-only slice the 4.1.5 audit reads: method_flow + plain steps + key_equations + claims). The assembler validates every path resolves to a real TODO and refuses kill-switch roots per map.
4. implementability audit (4.1.5, one isolated LLM call, default on): prompt references/system-prompts/implementability_audit.txt — fresh skeptical-engineer persona (separate from the 4.fill author) reads method_view.json (fallback: the full expansion) and rewrites each method step into a buildable spec: enriched_steps[] (one per step, same ids/order, what_changes + what_to_do_en + what_to_do_zh) + underspecified_points[] ({step_id, hole, fill, severity: filled|open} — unfillable holes stay honest as open). Compute-agnostic by design (resource feasibility is 4.1's job); never adds/removes/renames steps; never carries kill-switch fields. Output phase4/phase4_implementability.json.
5. validate + render: run the validators (below), then phase4_render — templating only, no model call; auto-detects the sibling implementability file and merges enriched_steps by step_id into the rendered Method (deterministic; no-op when absent). Writes idea.std.zh.md (plain Chinese, domain-newcomer register) + idea.std.en.md (plain English) + idea.detail.en.md (rigorous English — the novelty + validity defense) + idea.std.{en,zh}.tex (auto-compiled to PDF when xelatex/tectonic is on PATH; skipped with a hint otherwise).

Final response: read all three markdown cards and return them inline under headings 中文版 / English / Reviewer version. Other phase outputs stay on disk for inspection, not echoed.

Validators

# advance path: --phase3 = phase3_critique_output.json; revise path: --phase3 = phase3_revise_output.json
# --phase2 = the CANONICAL candidate (refined_candidate.json when 2.3 patched, else the 2.2 output)
python3 "$SKILL_DIR/scripts/run.py" validate \
  --phase2 <canonical candidate file> \
  --phase3 <see comment> \
  --phase4 $RUN_DIR/phase4/phase4_expansion.json \
  --phase4-impl $RUN_DIR/phase4/phase4_implementability.json   # optional; enables implementability checks

Retry budget on fail (cap = 2). Fix only the named contract, re-validate; still failing after the 2nd retry → stop revising, render as-is, and append a short note listing the failing validators (a flagged-imperfect card beats a watchdog-killed run with zero output). Never "fix" kill_switch_integrity or subpattern_citation_consistency by editing a guarded field — surface them as the headline caveat instead.

Configuration

By default every model-driven phase runs on the host LLM. To route phases to a different backend (Gemini, open-weights, custom):

• NOVELTY_LLM_REASONING_LARGE_CMD — Phase 1 / 2.1 / 2.2 / 3.2 / 3.3 / 4.fill (needs ≥ 200k context, JSON output)
• NOVELTY_LLM_CLASSIFY_FAST_CMD — Phase 0 intent extraction + per-paper pattern tagging (smaller context, JSON output)

Each is a CLI taking a stdin prompt (<<SYSTEM>>...<<USER>>...) and emitting JSON on stdout. When unset (the default when running inside any host LLM), the orchestrator emits sentinel files and the host LLM handles those steps natively.

• IDEASPARK_CROSS_RUN_DEDUP — set to off to disable the sibling-run soft-negative-anchor scan in the Phase 2 emit (default: on).
• IDEASPARK_DEFAULT_COMPUTE — optional standing compute profile for the user (free text, e.g. "8×H100 node, ~300 GPU-days, $50k API budget"). Put it in .env (auto-loaded); next surfaces it to Phase 1 as intake context. Precedence: compute stated in the user's query > this value > the factory default (80GB-class GPUs, ≤8 concurrent, ≈150 GPU-days / 5 months, ~$10k API campaign). Use this instead of editing the factory default — the default is the feasibility yardstick for users who state nothing.