arc-skills v4.2

Pure-skills research pipeline for AutoResearchClaw. Executable as a standalone skill chain by any AI agent — the workflow is agent-portable, while successful execution still depends on real local capabilities such as Python execution, network access, and local LaTeX.

Core non-negotiable objective: when the full chain succeeds, it should yield a high-quality, verifiable, reproducible, authentic academic paper with at minimum real LaTeX sources, a locally compiled PDF, verified references, real experiment-backed claims, and submission-format compliance.

Quick Start

Prerequisites Checklist

Before running the pipeline, verify you have:

Requirement	Check Command	Required For
Python 3.9+	python --version	Stages 10-13 (experiments)
PyTorch	python -c "import torch; print(torch.__version__)"	Neural network experiments
GPU/CUDA	nvidia-smi or torch.cuda.is_available()	GPU experiments
Local LaTeX	pdflatex --version	Stages 22, 28 (PDF compilation)
BibTeX	bibtex --version	Bibliography compilation
Network	curl -I https://doi.org/10.1000/182	Stages 3-4 (citation verification)

Install

# Clone or copy the skills
cp -R arc-skills/* ~/.claude/skills/

# Verify installation
ls ~/.claude/skills/arc-00-01-research-pipeline/

Canonical entrypoint — start from an idea:

/arc-00-01-research-pipeline
{
  "topic": "your research idea",
  "target_venue": "neurips"
}

Canonical entrypoint — improve an existing paper/code project:

/arc-00-01-research-pipeline
{
  "mode": "existing_project_improvement",
  "project_path": "/path/to/project"
}

Recovery helper:

/arc-00-03-resume

arc-00-01-research-pipeline is the only user-facing pipeline entry. It supports exactly two modes: default idea_start and existing_project_improvement.

Current Limitations

Template Assets

This repository provides skill definitions only. Some conference templates require official style files from organizers:

Template	Status	Additional Files Required
NeurIPS	Ready	Uses included neurips_2026.sty
ICML	Ready	Uses included icml2026.sty
ICLR	Ready	Self-contained
AAAI	Ready	Self-contained
ACL/EMNLP/NAACL	Partial	Requires acl.sty from official release
CVPR/ICCV/ECCV	Partial	Requires cvpr.sty from official release
IEEE	Ready	Uses standard IEEEtran class
Elsevier	Ready	Uses standard elsarticle class
Springer	Ready	Uses standard svjour3 class

If the pipeline resolves to a template that requires external style files not present in your workspace, Stage 22 (export) will block with E_TEMPLATE_ASSET_MISSING.

Execution Environment

The pipeline BLOCKS (does not degrade) if any of these are missing:

• No execution backend → Cannot run experiments (Stages 10-13)
• No local LaTeX → Cannot compile PDF (Stages 22, 28)
• No network → Cannot verify citations (Stages 3-4)

This is intentional — the pipeline refuses to produce paper-shaped artifacts without real evidence backing.

Known Issues

1. Template resolution (Stage 15.7) runs after hypothesis generation. Ideally, venue-aware decisions (e.g., conference vs journal citation preferences) should influence earlier stages.

2. External review (Stage 24) defaults to degraded local review if Codex MCP is unavailable. For full adversarial guarantees, configure a dedicated MCP reviewer.

3. Experience accumulation (Stage 20.5) stores per-run guidance only. Cross-run knowledge sharing requires manual aggregation.

Quick Start with Templates

Option 1: Use Built-in Templates (Recommended for Testing)

For venues with self-contained templates (NeurIPS, ICML, ICLR, AAAI, IEEE, Elsevier, Springer):

# Templates are auto-copied to workspace by Stage 15.7
# Just specify target venue:
/arc-00-01-research-pipeline
{
  "topic": "your research idea",
  "target_venue": "neurips"
}

Option 2: Use Official Conference Templates

For venues requiring official style files (ACL, CVPR):

# 1. Download official template
mkdir -p /path/to/workspace/templates/acl
cd /path/to/workspace/templates/acl
curl -L -o acl.sty https://raw.githubusercontent.com/acl-org/acl-style-files/master/acl.sty

# 2. Run pipeline
/arc-00-01-research-pipeline
{
  "topic": "your research idea",
  "target_venue": "acl",
  "template_path": "/path/to/workspace/templates"
}

Option 3: Custom Template

To use your own LaTeX template:

# Place your template files in workspace
/arc-00-01-research-pipeline
{
  "topic": "your research idea",
  "target_venue": "custom",
  "template_manifest": {
    "template_id": "custom",
    "asset_status": "user_supplied",
    "asset_dir": "/path/to/your/template"
  }
}

Template Quick Reference

Conference Targets (8 pages main + references):
  neurips, icml, iclr, aaai, acl, cvpr, kdd

Journal Targets (no hard page limit):
  ieee_journal, elsevier_journal, springer_journal

Anonymous submission (default for conferences):
  All conference templates default to anonymous mode

Camera-ready:
  Pass "submission_mode": "camera_ready" for final version

Design Philosophy

Skill = Task Definition + Output Contract + Acceptance Rules.

Hard Paper Standards

The chain target is not merely “a paper-shaped artifact”. Final acceptance should satisfy hard academic-output standards:

• Real local LaTeX compilation is mandatory; no fake or placeholder PDF path
• Verified literature only; fabricated/unverifiable references are blocking failures
• Quantitative claims must remain traceable to real experiment artifacts
• The paper should normally carry at least 30 references
• The paper should normally maintain at least 20% of references from the last 5 years
• Figures must satisfy venue-aware hard standards for count, readability, style, and file quality
• Figures must be traceable to real experiment outputs; fabricated or decorative figures are unacceptable
• Missing critical execution / LaTeX / network capability is blocking, not degradable

These standards should shape every iteration loop, review loop, and re-check flow.

Each SKILL.md describes:

• What must be produced (output schema)
• What quality the output must meet (quality contract)
• What happens on failure (error codes, retry policy)

Each SKILL.md does NOT describe:

• How to generate the output (execution is the Agent's responsibility)
• What tools or models to use (Agent decides based on its own capabilities)

---

Skill Structure

All 28 stage skills follow this mandatory structure:

Section	Meaning	Required
## Purpose	What this stage accomplishes	Yes
## Quality Contract	Hard constraints on the output — failure conditions	Yes
## Inputs	Required input files and their schema	Yes
## Outputs	Required output files and their schema	Yes
## Procedure	Step-by-step execution protocol (numbered)	Yes (all 28 stage skills)
## Validation	Machine-checkable acceptance criteria	Yes
## Error Codes	Failure types this stage can emit	Yes
## Retry Policy	Max retries for each error code	Yes
## Gate Behavior	(gate stages only) approve/reject routing	Gate only
## Rollback Contract	(gate/decision stages) rollback targets	Gate + decision
## State Transition	How this stage affects pipeline status	Yes

Orchestrator skills (arc-00-*) use a superset structure: Stage Map + State Machine + Error Registry + Gate/Rollback/Decision rules — they are the protocol definition, not the execution.

---

Orchestrators

Skill	Role	Has Procedure
arc-00-01-research-pipeline	Canonical user-facing entrypoint. Runs the pipeline in exactly two modes: default idea_start, or existing_project_improvement for paper/code re-check and iterative improvement.	N/A (protocol def)
arc-00-02-status	Read-only state inspector	Yes — 7 steps
arc-00-03-resume	Resume from checkpoint	Yes — 7 steps
arc-00-04-environment-probe	Internal/support orchestrator for environment detection (execution + local LaTeX + network). Automatically required before pipeline progress.	Yes — 11 steps
arc-00-05-project-recheck	Internal/support orchestrator used by existing_project_improvement mode. Diagnoses existing projects, restructures external ones into ARC-compatible artifacts, delegates fixes to main pipeline skills, and iterates until quality target and required late-stage gates pass or max iterations is exhausted.	Yes — 14 steps
arc-00-06-meta-optimizer	Internal/support orchestrator for experience accumulation and reuse. Distills reusable guidance from real artifacts into run-owned outputs without relying on bundled non-skill storage.	Yes — 8 steps
arc-09-01-paper-review-loop	External adversarial review via Codex/GPT-5.4 MCP in nightmare mode, with degraded-local review logged honestly only as fallback.	Yes — 9 steps
arc-09-02-paper-polish	De-AI Polish + Reverse Outline + provenance-preserving final checks	Yes — 7 steps

All 28 stage skills now have complete sections: Purpose · Quality Contract · Inputs · Outputs · Procedure · Validation · Error Codes · Retry Policy · State Transition. Gate/decision stages additionally have Gate Behavior and/or Rollback Contract.

---

Stage Map

Phase	Skills	Description
0: Environment	arc-00-04	Execution + LaTeX + Network detection (BLOCKING on missing capabilities)
A: Scoping	arc-01-01 → arc-01-02	Topic → Problem decomposition
B: Literature	arc-02-01 → arc-02-04	Search → Collect (BLOCKING on unverifiable DOI/arXiv) → Screen → Extract
C: Synthesis	arc-03-01 → 7.5 → arc-03-02	Synthesis → Novelty Gap Gate (BLOCKING) → Hypothesis generation
D: Experiment Design	arc-04-01 → 9.5 → arc-04-03	Design (GATE) → Reproducibility Design Gate (BLOCKING) → Code gen → Resource plan
E: Execution	arc-05-01 → arc-05-02	Run experiments → Iterative refine (with auto-repair)
F: Analysis	arc-06-01 → 15 → 15.5 → 15.7	Result analysis → Research decision → Result Claim Gate (BLOCKING) → Template Resolve
G: Writing	arc-07-01 → 17 → 17.5 → 18 → 18.5	Outline → Draft → Writing Compliance Gate (BLOCKING) → Review → Bibliography Quality Gate (BLOCKING) → Revision
H: Finalization	arc-08-01 → 20.5 → 21 → 21.5 → arc-08-03 → arc-08-04	Quality (GATE) → Meta Optimizer → Archive → Reproducibility Bundle Gate (BLOCKING) → Export → Verify
I: Review & Polish	arc-09-01 → 24.5 → arc-09-02	External nightmare review → Academic Integrity Gate (BLOCKING) → De-AI Polish
J: Late-Stage Gates	arc-10-01 → 27 → 27.5 → 28 → 28.5	Claim-Evidence Trace (GATE, requires arc-10-04 numeric-truth pre-gate) → Numeric Truth (pre-GATE) → Figure Quality (GATE) → Submission Format (GATE) → Final Acceptance Gate (BLOCKING)

---

Supported Venue / Template Families

The pipeline is now venue-aware. Template semantics are resolved before writing/export and preserved through submission gates.

Template/template-family rules are defined directly by the skills chain as text contracts rather than bundled JSON registries or asset folders.

The chain remains venue-aware, but because this repo is pure-skills-only:

• no built-in registry file is assumed
• no bundled schema file is assumed
• no bundled template asset directory is assumed
• if the chosen venue requires official LaTeX assets that are not already present in the execution workspace, stage 22 must block honestly rather than degrade

Family	Example template IDs	Publication type
neurips	neurips_2025, neurips_2024	conference
iclr	iclr_2026, iclr_2025	conference
icml	icml_2026, icml_2025	conference
aaai	aaai_2026	conference
acl_family	acl_2025, emnlp_2025, naacl_2025	conference
cv_family	cvpr_2026, iccv_2025, eccv_2026	conference
kdd	kdd_2025	conference
ieee_conf	ieeetran_conference	conference
ieee_journal	ieeetran_journal	journal
elsevier_journal	elsevier_cas_sc, pattern_recognition	journal
springer_journal	springer_lncs_journal_style	journal

Conference targets should carry explicit page-limit rules and usually default to submission_mode: anonymous. Journal targets may use page_limit_policy instead of a hard main-body cap and usually default to submission_mode: preprint unless camera-ready delivery is requested.

Template-contract validation rules also carry hard figure standards, including minimum figure count, recommended figure-count range, and style/file-quality requirements (caption_required, legend_required_for_multiseries, units_required_when_numeric_axes, color_consistent_across_figures, font_readable_min_pt, vector_preferred, raster_dpi_min, subfigure_label_style).

---

Gate Stages (17 gates + 1 pre-gate)

Stage	Skill	Rollback on Reject
0	arc-00-04-environment-probe	N/A (blocking) — pipeline cannot start without execution/LaTeX/network
4	arc-02-02-literature-collect	N/A (blocking on hallucination) — unverifiable citations rejected per-entry
5	arc-02-03-literature-screen	arc-02-02-literature-collect
7.5	arc-03-03-novelty-gap-gate	arc-03-01-synthesis
9	arc-04-01-experiment-design	arc-03-02-hypothesis-gen
9.5	arc-04-04-reproducibility-design-gate	arc-04-01-experiment-design
15.5	arc-06-03-result-claim-gate	arc-06-02-research-decision
17.5	arc-07-05-writing-compliance-gate	arc-07-02-paper-draft
18.5	arc-08-05-bibliography-quality-gate	arc-07-02-paper-draft
20	arc-08-01-quality-gate	arc-07-01-paper-outline
21.5	arc-08-06-reproducibility-bundle-gate	arc-08-01-quality-gate
24.5	arc-09-03-academic-integrity-gate	arc-09-01-paper-review-loop
26	arc-10-01-claim-evidence-trace-gate	arc-07-04-paper-revision
27	arc-10-04-numeric-truth-gate (pre-gate for stage 26)	arc-07-04-paper-revision
27.5	arc-10-02-figure-quality-gate	arc-08-03-export-publish (then re-run stages 23–25)
28	arc-10-03-submission-format-gate	arc-08-03-export-publish (then re-run stages 23–25)
28.5	arc-10-05-final-acceptance-gate	arc-10-03-submission-format-gate

---

Noncritical Stages

Stage 0 (arc-00-04-environment-probe) is CRITICAL and BLOCKING — failure here blocks the entire pipeline. Checks:

• Execution backend (GPU/CPU/SSH) — required for experiments
• LaTeX compilation (pdflatex + bibtex) — required for paper export
• Network connectivity (doi.org + arxiv.org) — required for literature verification

Stage 4 (arc-02-02-literature-collect) is CRITICAL and BLOCKING on hallucinations — any unverifiable DOI/arXiv ID triggers E04_HALLUCINATION_DETECTED.

The intended paper-quality floor is also explicit: references should normally reach at least 30 verified items, with at least 20% from the last 5 years, unless the topic genuinely lacks recent literature. This is a quality target for writing/review loops, not a license to fabricate citations.

Stage 20 (arc-08-01), Stage 20.5 (arc-00-06), Stage 21 (arc-08-02), and Stage 24 (arc-09-01) are noncritical for iteration bookkeeping, but only Stage 20.5, 21 and 24 may truly be skipped. Stage 20 must still be approved before Stage 22 export or any later stage can claim valid late-stage handoff or final acceptance; failures or degraded-local review outcomes must be logged honestly.

Stage 25 (arc-09-02-paper-polish) is always critical — De-AI patterns and structural failures MUST block export.

Stage 23 (arc-08-04) blocks on hallucinated citations, unverifiable citations, or unjustified citation-threshold failures (E23, E23_CITATION_THRESHOLD) — this check runs before late-stage gates to catch fake or insufficient references early.

New Critical Gates (Stages 7.5, 9.5, 15.5, 17.5, 18.5, 21.5, 24.5, 28.5):

• Stage 7.5 (arc-03-03): Blocks non-novel research gaps from entering hypothesis generation (E07B)
• Stage 9.5 (arc-04-04): Blocks coding until reproducibility contract is explicit (E09B)
• Stage 15.5 (arc-06-03): Blocks writing until claim-scope is frozen and evidence-linked (E15B)
• Stage 17.5 (arc-07-05): Blocks peer review until structure/marker/anti-slop compliance (E17B)
• Stage 18.5 (arc-08-05): Blocks downstream propagation of bibliography defects (E18B)
• Stage 21.5 (arc-08-06): Blocks export if reproducibility bundle is incomplete (E21B)
• Stage 24.5 (arc-09-03): Blocks camera-ready if integrity/disclosure issues exist (E24B)
• Stage 28.5 (arc-10-05): Final aggregate gate — blocks "success" if any critical upstream gate failed (E28B)

Stages 26–28 are always critical late-stage gates:

• Stage 26 (arc-10-01): block on unsupported, untraceable, or direction-conflicted claims (E26, E26_DIRECTION_CONFLICT), and on numeric pre-gate failures (E26_NUMERIC_*) from arc-10-04
• Stage 27 (arc-10-02): block on figure quality or authenticity failures against the polished package (E27)
• Stage 28 (arc-10-03): block on numeric-truth, submission format, citation-threshold, figure-authenticity, or reproducibility noncompliance (E28, E28_NUMERIC_PRECHECK_FAIL)

---

Decision Loop

• proceed → next stage
• refine → arc-05-02 (counts against cap)
• pivot → arc-03-02 (counts against cap)
• Cap: 2 total pivot/refine cycles, then forced proceed with quality_warning.txt

---

Phase I: Review & Polish (Stages 24–25)

arc-09-01-paper-review-loop

External adversarial multi-round review. Breaks the self-review trap. It reviews the stable stage-22 package, records remediation handoff items for downstream execution, and does not mutate the canonical package in place. If no independent reviewer is available, the stage downgrades to review_mode: degraded_local and records that downgrade explicitly instead of silently pretending independence. Degraded-local review may inform downstream fixes, but it cannot waive blocking concerns for citation authenticity, numeric truth, claim traceability, figure authenticity, or submission-format compliance.

Constant	Value
MAX_ROUNDS	40
POSITIVE_THRESHOLD	score ≥ 6/10 AND verdict "ready" or "almost"
REVIEWER_ROLE	external_adversarial_reviewer
REVIEWER_MODEL	gpt-5.4
REVIEWER_PROVIDER	codex_mcp
HUMAN_CHECKPOINT	false
REVIEWER_DIFFICULTY	nightmare

Reviewer mode:

• nightmare: Codex/GPT-5.4 MCP independently verifies paper, code, citations, numeric claims, and figure provenance; if MCP is unavailable, the stage downgrades honestly to degraded_local and logs that downgrade explicitly

arc-09-02-paper-polish

Dedicated De-AI Polish Pass + Reverse Outline Test + final structural verification.

Key checks:

• No forbidden AI words (delve, pivotal, landscape, etc.)
• Reverse outline forms coherent narrative
• Section word counts within target ranges
• All claims from experiment_summary covered
• No stale section files
• Regenerates the final polished package used by stages 26–28

Phase J: Late-Stage Gates (Stages 26–28)

arc-10-01-claim-evidence-trace-gate

Builds a machine-checkable claim-evidence matrix and blocks unsupported/untraceable claims. Stage 26 execution requires pre-gate arc-10-04, which writes stage-26-pre/numeric_truth_report.json.

arc-10-04-numeric-truth-gate (pre-gate for Stage 26)

Validates quantitative claim consistency against experiment artifacts before claim-evidence gating.

arc-10-02-figure-quality-gate

Validates venue-aware hard figure-count thresholds, readability, style consistency, file quality, statistical consistency, data traceability, and figure authenticity against the canonical polished package (stage-25/deliverables/charts/ plus stage-25/main.tex).

arc-10-03-submission-format-gate

Final submission gate: enforces numeric-truth pre-gate integrity, compile cleanliness, schema-valid template/report contracts, page/anonymity/front-matter checks, figure-gate consistency including authenticity, citation-threshold checks, and complete reproducibility-ready bundle.

Late-stage canonical handoff

• Template semantics are resolved once at artifacts/<run_id>/template/template_manifest.json and reused by stages 16–28
• Stage 22 produces the canonical review/export bundle at artifacts/<run_id>/stage-22/ (paper_final.md, main.tex, main.pdf, references.bib, compile_report.json, manifest.json, reproducibility_report.json, template_manifest.json, deliverables/)
• Stage 23 verifies citations against artifacts/<run_id>/stage-22/, emits artifacts/<run_id>/stage-23/references_verified.bib, and does not mutate the stage-22 bundle in place
• Stage 24 performs adversarial review against the same stage-22 bundle and records review_state.json plus AUTO_REVIEW.md under artifacts/<run_id>/stage-24/
• Stage 25 regenerates the canonical polished bundle at artifacts/<run_id>/stage-25/ (paper_polished.md, main.tex, main.pdf, references.bib, compile_report.json, manifest.json, reproducibility_report.json, template_manifest.json, deliverables/) from stage-22 content plus the stage-23 verified bibliography
• Stage 27 writes artifacts/<run_id>/stage-27/figure_quality_report.json, which must match the polished package and resolved venue-specific figure rules
• Stages 26–28 validate only the stage-25 polished bundle; Stage 26/28 also require intact stage-26-pre/numeric_truth_report.json, and Stage 28 writes gate artifacts under artifacts/<run_id>/stage-28/
• If stage 27 or 28 rejects, rollback returns to stage 22 so stages 23–25 can be re-run against a refreshed export before the remaining late-stage gates resume; the refreshed export must include a real repaired change, not a no-op loop through the same artifacts

Why this arrangement

• Review and citation verification happen against a stable exported package
• Polish creates a new canonical package instead of mutating the review package in place
• Trust gates validate the final polished artifacts, not an older draft/export snapshot
• Late rollback preserves the verification chain by forcing stages 23–25 to run again after stage-22 regeneration

Late-stage acceptance matrix

Acceptance dimension	Primary stage	Must prove	Blocking rule
Environment reality	Stage 0 (arc-00-04)	Real execution, local LaTeX, and network are available	Missing critical capability blocks pipeline start
Citation authenticity	Stage 23 (arc-08-04)	Every citation is verified or reused from verified corpus	Any hallucinated/unverifiable citation blocks
Citation sufficiency	Stage 23 / 25 / 28	Bibliography reaches ≥30 verified refs and ≥20% from last 5 years, or has explicit topic justification	Unjustified failure blocks late-stage handoff
Review observability	Stage 24 (arc-09-01)	Review artifacts are recorded honestly, including degraded-local fallback when needed, and any discovered blocking concerns remain visible downstream	Missing/false review logging is not acceptable, though stage 24 itself remains noncritical
Claim traceability	Stage 26 (arc-10-01)	Every important claim maps to evidence, and directional/hypothesis-support wording matches that evidence	Unsupported or direction-conflicted claims block
Numeric truth	Stage 26 pre-gate (arc-10-04)	Quantitative statements match experiment artifacts, and the pre-gate report remains present whenever Stage 26 progress is claimed	Numeric mismatch or missing pre-gate evidence blocks
Figure quality	Stage 27 (arc-10-02)	Figures satisfy count, readability, style, and file-quality rules	Critical figure failure blocks
Figure authenticity	Stage 27 / 28	Figures are evidence-bearing and traceable to real experiment outputs	Decorative/fabricated figure blocks
Reproducibility semantics	Stage 22 / 25 / 28	Commands, seeds, environment, and provenance are preserved	Missing or semantically broken reproducibility artifact blocks
Submission-format compliance	Stage 28 (arc-10-03)	Compile clean, schema-valid, venue-valid final package	Format/package noncompliance blocks

This matrix is the final acceptance contract for the whole chain: a run is not "successful" unless all blocking dimensions are satisfied.

---

Canonical Error Codes (E00–E28)

Code	Stage	Retry Cap	Description
E00_CRITICAL_MISSING	ENVIRONMENT_PROBE	0	Critical capability missing (execution/LaTeX/network) — BLOCKING
E00_NO_EXECUTION_ENV	ENVIRONMENT_PROBE	0	No execution backend
E00_NO_LATEX	ENVIRONMENT_PROBE	0	LaTeX compilation environment missing
E00_NETWORK_UNREACHABLE	ENVIRONMENT_PROBE	1	doi.org or arxiv.org unreachable
E00_PROBE_FAILED	ENVIRONMENT_PROBE	1	Probe failed
E00_CONDA_NOT_FOUND	ENVIRONMENT_PROBE	0	Conda not installed
E00_CONDA_ENV_FAILED	ENVIRONMENT_PROBE	1	Conda environment creation failed
E00_SSH_UNREACHABLE	ENVIRONMENT_PROBE	0	SSH remote configured but unreachable
E01	TOPIC_INIT	0	—
E02	PROBLEM_DECOMPOSE	1	—
E03	SEARCH_STRATEGY	1	—
E04_INSUFFICIENT_VERIFIED	LITERATURE_COLLECT	2	< 50 verified candidates
E04_HALLUCINATION_DETECTED	LITERATURE_COLLECT	0	Unverifiable DOI/arXiv detected — BLOCKING
E04_NETWORK_FAILURE	LITERATURE_COLLECT	2	Cannot reach doi.org/arxiv.org for verification
E04_NO_DOI_OR_ARXIV	LITERATURE_COLLECT	0	Candidate lacks both DOI and arXiv ID
E05	LITERATURE_SCREEN	0 (→rollback)	—
E06	KNOWLEDGE_EXTRACT	1	—
E07	SYNTHESIS	1	—
E07B	NOVELTY_GAP_GATE	0 (→rollback)	Novelty gap report missing or blocked core gap
E07B_GAP_NOT_NOVEL	NOVELTY_GAP_GATE	0	Core gap is already addressed by prior work
E07B_EXTERNAL_REVIEW_MISSING	NOVELTY_GAP_GATE	0	External adversarial review path unavailable
E08	HYPOTHESIS_GEN	1	—
E09	EXPERIMENT_DESIGN	0 (→rollback)	—
E09B	REPRODUCIBILITY_DESIGN_GATE	0 (→rollback)	Reproducibility design report missing required fields
E09B_SEED_POLICY_MISSING	REPRODUCIBILITY_DESIGN_GATE	0	Seed policy missing or num_seeds < 3
E09B_ENV_LOCK_MISSING	REPRODUCIBILITY_DESIGN_GATE	0	No dependency/environment lock artifact declared
E09B_STATS_METHOD_MISSING	REPRODUCIBILITY_DESIGN_GATE	0	Statistical comparison method missing
E10	CODE_GENERATION	2	—
E11	RESOURCE_PLANNING	1	—
E12	EXPERIMENT_RUN	2	>50% runs failed
E12_NO_EXECUTION_ENV	EXPERIMENT_RUN	0	No execution backend
E12_FABRICATION	EXPERIMENT_RUN	0	Fabricated data
E13	ITERATIVE_REFINE	2	—
E13_LOW_COMPLETION	ITERATIVE_REFINE	1	Low completion rate
E13_REPAIR_EXHAUSTED	ITERATIVE_REFINE	0	Max repairs reached
E14	RESULT_ANALYSIS	1	—
E14_DATA_INVALID	RESULT_ANALYSIS	0	Data validation failed
E15	RESEARCH_DECISION	1	—
E15B	RESULT_CLAIM_GATE	0 (→rollback)	Claim-scope report missing or not fully classified
E15B_UNSUPPORTED_CORE_CLAIM	RESULT_CLAIM_GATE	0	Core contribution claim is unsupported
E15B_MISSING_CAVEAT	RESULT_CLAIM_GATE	0	allowed_with_caveat lacks explicit wording
E16	PAPER_OUTLINE	1	—
E17	PAPER_DRAFT	1	—
E17B	WRITING_COMPLIANCE_GATE	0 (→rollback)	Writing compliance report failed
E17B_MISSING_SECTION	WRITING_COMPLIANCE_GATE	0	Required sections missing
E17B_LIMITATIONS_MISSING	WRITING_COMPLIANCE_GATE	0	Limitations section missing or vague
E17B_MARKERS_REMAIN	WRITING_COMPLIANCE_GATE	0	TODO/FIXME/VERIFY markers remain
E17B_AI_PATTERNS	WRITING_COMPLIANCE_GATE	0	Forbidden writing patterns remain
E18	PEER_REVIEW	1	—
E18B	BIBLIOGRAPHY_QUALITY_GATE	0 (→rollback)	Bibliography quality report failed
E18B_MISSING_ENTRIES	BIBLIOGRAPHY_QUALITY_GATE	0	In-text citation keys have no entry
E18B_UNRESOLVED_ENTRY	BIBLIOGRAPHY_QUALITY_GATE	0	Unresolved verification marker remains
E18B_DUPLICATE_KEY_CONFLICT	BIBLIOGRAPHY_QUALITY_GATE	0	Duplicate citation key conflict
E18B_MISSING_FIELDS	BIBLIOGRAPHY_QUALITY_GATE	0	Required metadata fields missing
E19	PAPER_REVISION	1	—
E20	QUALITY_GATE	0 (→rollback)	—
E21	KNOWLEDGE_ARCHIVE	1	—
E21B	REPRODUCIBILITY_BUNDLE_GATE	0 (→rollback)	Reproducibility bundle report failed
E21B_ENTRYPOINT_MISSING	REPRODUCIBILITY_BUNDLE_GATE	0	No runnable reproduction entrypoint
E21B_ENV_LOCK_MISSING	REPRODUCIBILITY_BUNDLE_GATE	0	No dependency/environment lock artifact
E21B_RUNTIME_RECORD_MISSING	REPRODUCIBILITY_BUNDLE_GATE	0	Runtime environment recording missing
E21B_CLAIM_LINK_MISSING	REPRODUCIBILITY_BUNDLE_GATE	0	Core claims cannot be traced to bundled artifacts
E22	EXPORT_PUBLISH	2	Stage-22 export package failed local LaTeX compilation, reporting, or packaging
E23	CITATION_VERIFY	1 (blocks late-stage handoff)	Stage-23 citation verification found hallucinated, unverifiable, or unjustifiably insufficient citations
E24	PAPER_REVIEW_LOOP	1	Stage-24 review loop state failure, or reviewer downgrade/failure explicitly and honestly logged in review artifacts
E24B	ACADEMIC_INTEGRITY_GATE	0 (→rollback)	Academic integrity report failed
E24B_ANONYMITY_LEAK	ACADEMIC_INTEGRITY_GATE	0	Anonymous submission leaks identity
E24B_DISCLOSURE_MISSING	ACADEMIC_INTEGRITY_GATE	0	Required disclosure missing
E24B_SELF_CITATION_POLICY	ACADEMIC_INTEGRITY_GATE	0	Self-citation policy violation
E24B_MARKERS_REMAIN	ACADEMIC_INTEGRITY_GATE	0	Integrity/disclosure markers remain
E25	PAPER_POLISH	1 (blocks late-stage handoff)	Stage-25 polish, citation-stability/thresholds, reproducibility, or package regeneration failure
E26	CLAIM_EVIDENCE_TRACE_GATE	0 (→rollback)	Unsupported/untraceable claims
E26_NUMERIC_MISMATCH	NUMERIC_TRUTH_GATE (pre-gate)	0	Quantitative claim mismatches experiment data
E26_NUMERIC_UNVERIFIABLE	NUMERIC_TRUTH_GATE (pre-gate)	0	Quantitative claim not traceable
E27	FIGURE_QUALITY_GATE	1 (→rollback)	Figure quality, traceability, or authenticity failed
E28	SUBMISSION_FORMAT_GATE	2 (→rollback)	Compile, format, package, citation-threshold, figure-authenticity, or reproducibility noncompliance
E28B	FINAL_ACCEPTANCE_GATE	0 (→rollback)	Final acceptance report failed
E28B_PACKAGE_INCOMPLETE	FINAL_ACCEPTANCE_GATE	0	Final paper package incomplete
E28B_CRITICAL_GATE_FAILED	FINAL_ACCEPTANCE_GATE	0	One or more critical upstream gates failed
E28B_CITATION_VERIFY_MISSING	FINAL_ACCEPTANCE_GATE	0	Citation verification missing or failed
E28B_INTEGRITY_MISSING	FINAL_ACCEPTANCE_GATE	0	Academic integrity output missing or failed

---

Prerequisites (Detailed Setup)

1. Execution Backend

For GPU experiments:

# Check CUDA availability
python -c "import torch; print(torch.cuda.is_available())"

# Expected output: True

For CPU experiments (slower):

# Verify Python environment
python --version  # >= 3.9
python -c "import torch; print(torch.__version__)"

2. Local LaTeX Installation

macOS:

brew install --cask mactex-no-gui
# Or minimal install:
brew install --cask basictex
sudo tlmgr update --self
sudo tlmgr install latexmk

Ubuntu/Debian:

sudo apt-get update
sudo apt-get install texlive-full texlive-science texlive-publishers
# Or minimal install:
# sudo apt-get install texlive-base texlive-latex-extra texlive-science texlive-publishers

Verify:

pdflatex --version  # Should show version
bibtex --version     # Should show version

3. Network Connectivity

# Required for citation verification
curl -s -o /dev/null -w "%{http_code}" https://doi.org/10.1000/182
# Expected: 200, 301, or 303

curl -s -o /dev/null -w "%{http_code}" https://arxiv.org/abs/2301.00001
# Expected: 200

4. Python Environment

# Create conda environment (recommended)
conda create -n arc-research python=3.10
conda activate arc-research

# Install PyTorch (adjust for your system)
conda install pytorch torchvision torchaudio -c pytorch

# Additional packages
pip install numpy scipy matplotlib pandas

---

Troubleshooting

E00_CRITICAL_MISSING (Stage 0 blocked)

Symptom: Pipeline stops immediately with "Critical capabilities missing"

Fix:

# Check what's missing
cat environment_error.md

# Install missing components
# - For LaTeX: see Prerequisites section above
# - For execution: set up conda environment
# - For network: check internet connection

E04_HALLUCINATION_DETECTED (Stage 4 blocked)

Symptom: "Unverifiable DOI/arXiv detected"

Cause: Some papers in search results lack valid identifiers

Fix: Pipeline automatically filters these. If all candidates rejected:

• Check network connectivity
• Try broader search terms
• Verify doi.org and arxiv.org are reachable

E22 (Stage 22 LaTeX compile failed)

Symptom: "Stage-22 export package failed local LaTeX compilation"

Common causes:

1. Missing template style file:

   # For ACL/CVPR: download official style files
   # See "Quick Start with Templates" section above

2. LaTeX package not installed:

   # Install missing packages (TeX Live)
   sudo tlmgr install <package-name>
   
   # Or (MiKTeX)
   mpm --install <package-name>

3. Compilation errors:

   # Check compile log

cat artifacts/<run_id>/stage-22/compile_report.json

### E26_NUMERIC_MISMATCH (Pre-gate for Stage 26)

**Symptom:** "Quantitative claim mismatches experiment data"

**Cause:** Paper claims numbers not found in `experiment_summary.json`

**Fix:**
- Verify `experiment_summary.json` contains all reported metrics
- Check metric keys match exactly (case-sensitive)
- Re-run experiments if data is stale

### E24_CODEX_MCP_REQUIRED (Stage 24 BLOCKED)

**Symptom:** "Codex MCP external review is REQUIRED but unavailable"

**Cause:** Codex MCP not configured or unreachable

**Impact:** Pipeline **BLOCKS** — cannot proceed to late-stage gates without genuine external review

**Fix (REQUIRED):**
```bash
# 1. Ensure Codex is installed and configured
# See: https://github.com/codex/codex-cli

# 2. Verify MCP server is running
codex mcp status

# 3. Confirm GPT-5.4 model access
# Requires appropriate API key and permissions

# 4. Retry pipeline once MCP is available
/arc-00-03-resume

Why this is mandatory:

• Self-review violates scientific independence
• Local models share context with the executor, creating bias
• The adversarial guarantee requires a genuinely independent reviewer
• No degraded fallback is permitted — fix the MCP connection or abort

---

FAQ

Q: Can I run without GPU? A: Yes, set execution.backend: local_cpu. Experiments will be slower.

Q: Can I use Overleaf instead of local LaTeX? A: No. The pipeline requires local compilation for Stage 22/28. Overleaf is not supported.

Q: What if I only have some templates? A: The pipeline works with available templates. For missing official templates (ACL, CVPR), it will block at Stage 22 with clear instructions on what to download.

Q: How long does a full run take? A: Depends on experiment complexity:

• Simple idea → paper: 2-6 hours
• With real experiments: 4-12 hours (mostly experiment time)
• With iteration loops: add 1-3 hours per pivot/refine cycle

Q: Can I resume from a checkpoint? A: Yes, use /arc-00-03-resume or specify from parameter in entrypoint.

Q: Where are outputs saved? A: All artifacts in artifacts/<run_id>/stage-XX/ directories. Final package at artifacts/<run_id>/stage-28/.

---

Development

Adding New Templates

See templates/README.md for template contribution guidelines.

Testing Skills

# Run environment probe
/arc-00-04-environment-probe

# Check specific skill
/arc-00-02-status
{
  "run_id": "your-run-id"
}

Debugging

Set auto_approve_gates: false to manually review each gate decision.

---

License

This project is part of the AutoResearchClaw ecosystem. See LICENSE file for details.

Citation

If you use this pipeline in your research, please cite:

@software{arc_skills,
  title = {ARC-Skills: Automated Research Chain},
  year = {2026},
  url = {https://github.com/researchclaw/arc-skills}
}