nullius

Correctness and legibility checklist for data/results figures in any field: charts, parameter scans, spectra, distributions, fit comparisons, constraint contours, heatmaps, and low-dimensional projection scatters. Use when an agent plots computed or measured data, revises a results figure for a manuscript, or audits whether a data figure is publication-ready. Covers data fidelity (excluded rows never enter plotted summaries, connected-series evaluator fingerprints are homogeneous, a rendered continuous field is continuity-scanned before delivery, claim-titles true against every axis category, one canonical value per claim), label economy floor and ceiling, colour threading with a CVD-safe palette, role-mapped typography, chart choice by data shape, a render-then-verify QA loop (bbox overlap check plus per-panel perceptual crops), a figure-reproduction provenance bundle, and a fail-closed display-acceptance gate that refuses durable or outward-facing use until every plotted quantity is bound to a verification-

Use when writing Julia numerical or scientific code, to apply always-on performance guardrails, gate an accelerated implementation as an identity against its verbatim reference over the full domain of intended use, and escalate to reproducible benchmark gating for speedup or regression claims. Works standalone by default and can optionally emit ecosystem artifacts.

Convergence and reliability gate for NUMERICAL results in any field, including fits, optimizations, integrals, eigenvalues, roots, poles, zeros, ODE/PDE solutions, Monte-Carlo estimates, and downstream feature extraction. Use before trusting, comparing, publishing, or folding a computed number into durable research artifacts. Requires resolution convergence, independent-method checks, regression anchors, post-correction downstream recomputation, method-precondition checks, configuration-threading audits, gate-discrimination (negative-control) audits of purpose-built validation chains, accelerated/heuristic fast-path scoping (no false guarantee, unconditional escape hatch, production-validated precondition), and honest uncertainty/reporting. Emits an auditable reliability matrix. Sibling to derivation-verify for symbolic claims and julia-perf for speed claims.

AI failure-mode checklist (M1-M8) for research agents. Generic across domains. Walk M1-M7 at the moments work becomes durable — you just finished a derivation or computation and are about to write conclusions, fold a number or claim into the plan or contract, check off a task-board item, make a milestone closeout commit, hand off, assemble submission material, or claim a result is final — and, in projects using the engine's approval flow, before requesting an A1-A5 approval gate. Walk M8 before starting each heavy or production computation or hand-building a reusable computational method.

Use when a research milestone (theory or computation) needs executed, reproducible work whose results are independently reviewed before being trusted. Milestone-based research-team workflow for theory+computation projects with reproducible artifacts, independent parallel workstreams (default: host-native subagents; configurable), and a strict convergence gate.

Use when a result, manuscript, derivation, or diff needs independent review. Runs clean-room multi-agent review loops across Claude/Gemini/Codex/OpenCode/Kimi with strict review-contract checks, fallback policy, and convergence gates.

Turn a research idea into an auditable posterior: run an admission gate before any graph is built, decompose the idea into five source-grounded sub-criteria, encode them as a Gaia argument graph (gaia-lang, pinned 0.5.0a4), run exact inference, and write the posterior back to the idea store. Enforces a parameter honesty discipline — three fixed likelihood grades on the Jeffreys evidence scale, a mandatory anchor note on every number, MaxEnt fallback when an anchor fails review — plus pairwise expansion for mutual exclusivity, tournament-result absorption, revival by appending evidence, and review that audits anchors rather than scores. Use when admitting an idea into a tracked portfolio, when new evidence (literature, trial computation, tournament result) should move an idea's posterior, or when auditing why a posterior is what it is.

Turn a shallow, metadata-only literature pull into a DEEP review — multi-hop discovery via the existing literature-workflows recipes, per-paper deep-read notes that fill the research-team KB note template from the actual source (with verbatim quotes + locators), an optional double-reader mode for load-bearing papers (two independent readers from different model families plus a moderator that keeps disagreements visible), cross-paper synthesis (consensus / tensions / gaps), correct Markdown math rendering, and a checkable literature_survey_v1 artifact. Run when a survey feels thin, before promoting an idea, or before writing a related-work / introduction section.

Build, validate, and QA auditable literature graphs from deep-read notes and survey artifacts. Use when an agent needs to turn a literature review into an interactive graph with paper/method/topic nodes, source-backed edges, embedded notes, portable relative links, rendered math, verified figures, and browser-tested interactions.

General-purpose HEP calculation reproduction/audit runner that orchestrates Mathematica (FeynCalc/FeynArts/FormCalc/LoopTools/FeynRules) and/or Julia (LoopTools.jl). Supports: compute-only runs, LaTeX value auditing, auto-generation of Feynman diagrams + one-loop (unrenormalized) amplitudes (FeynRules→FeynArts→(optional)FormCalc), optional LaTeX-driven model_build scaffolding (agent-provided rewrite rules), and auditable out_dir + optional research-team sync.

Cross-index triangulation of citation metadata. For a citation entry under test, fetch canonical metadata for the same work from two or more independent bibliographic indexes, write one JSON block per provider, and run a deterministic offline comparator that emits a per-field agree/disagree/missing matrix with a fail-closed verdict (consistent / conflicted / insufficient_sources). Catches wrong-version, wrong-author, spliced, and hallucinated citation entries whose metadata cannot agree across indexes. Run before freezing a bibliography, during referee spot-checks, or when admitting papers into a literature-review core set.

Active statement-support grounding for cited claims. For each claim that leans on a citation, fetch the cited source and verify its CONTENT actually substantiates the claim — recording a span-backed verdict in a claim_grounding_report_v1 artifact. Generic across domains; routes HEP claims through INSPIRE and everything else through arXiv/OpenAlex. Run right after writing citation-backed prose (an introduction, related-work, or discussion section), before folding cited claims into durable artifacts, and before final conclusions.

Use when working inside an external research project that has or may need nullius state, research_plan.md, research_contract.md, artifacts/runs, team/runs, Codex/Claude Code continuation, recovery, verification, approval, export, handoff, compute environment/tool readiness validation (import + seeded witness + agent-follows-doc), surviving long-running / kill-prone compute jobs (checkpoint + heartbeat + deadline + resume), an opt-in independent reproduction check (fresh-checkout rerun compared against declared expected values), or a production launch authorization preflight (frozen plan hash + hash-bound review verdicts + exact environment fingerprint before a large production run may start).

Use the moment you have just derived a formula, closed form, identity, estimator, bound, or sign/branch choice that later work will rely on — before folding it into durable artifacts. Convergence-gate harness for DERIVATION-heavy results in ANY field (closed forms, integrals/sums, algebraic/operator/logical identities, probabilities/estimators, asymptotic/complexity bounds, sign/branch/boundary choices — domain carried only by the caller's context). For each atomic claim it runs >=2 INDEPENDENT blind re-derivations, an adversarial comparator that clusters them by MATHEMATICAL equivalence, and a tie-break loop that adds fresh independent derivations until >=2 agree — emitting an auditable verification matrix. Backend-agnostic contract with TWO executors: a Claude/Workflow-native one (same-model, fast) and a CLI multi-backend one (Claude/Codex/Gemini/OpenCode/Kimi) that enforces cross-model-FAMILY agreement + an adjudicator veto for the strongest verification. Sibling to `review-swarm` (which REVIEWS an artifac

Check and repair deterministic Markdown hygiene issues in research notes, especially Markdown math escaping, TOC-generated LaTeX escapes, rendered math/link readiness, and portable cleanup before research-harness/research-team/research-writer/research-integrity handoff; also trigger on rendering symptoms such as math showing as raw $...$ source on GitHub or a report page, formulas that render in one viewer but not on the target surface, LaTeX escapes that look doubled, or paper references that are not clickable.

Judged pairwise comparison between two research ideas in a campaign. Commit the comparison criteria to disk first (sha256 over the canonicalized list), generate symmetric anchored advocacy statements, collect independent votes from a judge panel drawn from the third-party agent roster (agents.json; at least three distinct model families for a cross-family match, with an honest single-family native subagent degradation when the roster cannot field that floor), and write a pairwise_match_v1 artifact whose outcome enters the belief layer as one observation. Arguments without a literature or computation anchor are discarded and counted, never weighed. Use when a campaign needs a relative-merit signal between two idea nodes beyond their individual posteriors, when choosing which of two competing ideas to advance, or when a newly admitted idea needs a calibrated first comparison.

Browse and install Nullius skills. This directory is the skills-market catalog (package metadata + installers), not a runnable task skill; use it to discover available skills and install specific ones into your skills home.

Run the local `codex` CLI (OpenAI Codex) in non-interactive mode (`codex exec`) for arbitrary prompts; supports file-based system+user prompts, retries with exponential backoff, and writing output to a target file.

Run the local `claude` CLI in non-interactive mode (with optional retries/backoff) for arbitrary prompts; supports loading system+user prompts from files and writing output to a target file.

Run the local `gemini` CLI in one-shot mode for arbitrary prompts; supports model selection, file-based prompt input, and writing output to a target file with safe fallback if the model alias differs.

Run the local `kimi` CLI (Kimi Code) in non-interactive prompt mode for arbitrary prompts; supports file-based system/user prompt inputs, isolated default working directory, optional workspace exposure, retries, model fallback, and writing parsed assistant text to a target file.

Run the local `opencode` CLI in one-shot mode for arbitrary prompts; supports file-based system/user prompts, JSON event parsing into plain text output, and safe fallback to the CLI default model when a model alias is unavailable.

Use when a manuscript or draft needs an independent referee-style assessment before submission, revision, or handoff. Generates a clean-room, offline (no network) referee-review report with Markdown + strict JSON output (generic profile).

Decision layer for allocating research effort across a portfolio of ideas: Thompson-sampling allocation from per-idea posteriors (a Beta construction with Laplace pseudo-counts), slot-based cuts into deep investment vs reconnaissance vs hold, cold-start handling for ideas without a posterior, and an activation monitor for ideas parked on recorded conditions (external activation conditions and missing admission evidence) with ready-to-paste lifecycle-transition commands. Use before each investment round, after new evidence has updated the belief graph posteriors, or when checking whether waiting ideas can enter play. Keeps utility, cost, and budget OUT of the belief layer; slot counts are caller-supplied (no built-in budget); an explicit random seed makes every allocation reproducible.

Generate derived (non-seed) idea nodes for an idea campaign from research-progress evidence deltas — V0 operator: LiteratureMining over a deep literature survey's tensions and re-anchored gaps (FailureRouting when a failed-approach ledger exists) — and import them through the engine's node.import_generated RPC as an auditable generation pack. Retrieval-first by construction (no retrieval receipt, no evidence URI), one candidate per structurally distinct anchor, mechanical dedup against every campaign node, novelty stated as a falsifiable closest-prior delta claim (never an LLM novelty score), zero generator-side scoring: nodes are born with posterior null and face the same idea-posterior admission gate as seeds. Use when a new or updated literature survey lands, when failed-approach entries accumulate, or when the operator deliberately runs a manual generation burst on a campaign.

Use proactively when a research task would benefit from the Nullius hep-mcp MCP stack for literature search, arXiv source retrieval, broad scholarly metadata/citation context through OpenAlex, INSPIRE/HEPData/PDG/Zotero provider access, evidence collection, citation mapping, writing/export support, or source-backed prior-art checks; explicit user mention is not required. Also trigger when a user asks to use hep-mcp, INSPIRE, arXiv source, arxiv-mcp, OpenAlex, openalex-mcp, HEPData, PDG, Zotero, or local literature/evidence MCP tools. Keep HEP-specific assumptions scoped to HEP workflows and providers.

Build the initial framework of a software package FROM LITERATURE — a 7-phase pipeline (survey, extraction, skeleton, reimplementation, reference-check, composite-gates, closeout), each phase closed by a deterministic fail-closed gate. Orchestrates deep-literature-review, claim-grounding, citation-triangulation, derivation-verify, numerical-reliability-gate, julia-perf, and review-swarm; adds the literature-to-code bridge they do not cover: span-anchored extraction manifests, a traceability ledger, clean-room independent reimplementation, and reference-value cross-validation. Use when turning one or more papers into a new package skeleton whose every equation, convention, constant, and algorithm must be traceable and independently verified.

Content-first revision for academic papers written in LaTeX (read-through -> line edit -> clean + diff + tracked delivery contract -> audit + verification requests).

Scaffold or validate an arXiv-ready RevTeX4-2 (12pt, onecolumn) paper from a `research-team` project, with provenance wiring, BibTeX hygiene, and deterministic Markdown/LaTeX checks (optional multi-backend section drafting).

Create, revise, and verify publication-grade physics diagrams for LaTeX papers: Feynman/process diagrams, tree/loop/exchange diagrams, Bethe-Salpeter/Faddeev/integral-equation schematics, colored recursion blocks, multi-panel process figures, and geometry/angle/coordinate diagrams. Use when your agent needs to draw or repair diagrams with mathematical labels that must not overlap lines, produce editable source plus vector PDF for manuscript insertion, or audit whether a physics figure is publication-ready. Default to explicit-coordinate TikZ; use tikz-feynhand for particle-line idioms in Feynman/process diagrams; use Matplotlib/Python only for geometry/coordinate/angle schematics or styling already-computed data where it is a better fit. Do not use this skill to compute external numerical results for plots.

Teaches the two-step zotero_add then zotero_confirm pipeline for importing papers into Zotero, including collection selection, file attachment, dedup handling, and non-HEP DOI support.

Teaches the PDG (Particle Data Group) tool chain for looking up particle properties, measurements, decays, and references from the local PDG SQLite database.