Claude Skills by yogsoth-ai

Build articles × concepts coverage matrix to visualize research landscape and identify empty cells as gap candidates.

Quantify how much conclusions change across all assumption negations and produce a sensitivity ranking.

Map which assumptions are load-bearing by assessing how the conclusion changes if each assumption fails.

Record evaluation conditions (data splits, hyperparams, hardware, seeds) from a paper

Compare and standardize experimental conditions across papers

Standardize evaluation condition differences across papers — 20 methods, 60 data points, 30 web searches budget

Calibrates confidence scores based on debate progression. Determines whether to escalate, continue, or terminate based on cumulative evidence.

How do constraints conflict with each other? — Evaporating Cloud + assumption challenging + injection to resolve constraint conflicts

Apply conjunctive screening rules to eliminate candidates that fail any threshold.

Classify items as consensus or dissensus at a given threshold.

Compute consensus score from collected judgments using the appropriate statistical method.

Synthesize all rounds into a final consensus report documenting agreements, dissent, and process.

Follow a provocation's logical consequences step by step to extract viable insights and new research directions.

Detect preference cycles, localize inconsistent judgments, request corrections, and recompute ratings until consistency threshold is met.

SOP: 检验 pairwise 判断矩阵的传递一致性，识别不一致项并建议修正

What limits us — identify bottlenecks, quantify constraints, analyze dependencies, resolve conflicts before experiment execution

Orchestrate the full constraint-breaking cycle: extract conflict, challenge assumptions, project resolution

Classify constraints into hard constraints, soft constraints, and assumptions.

Identify constraints, classify them by type and severity, assess removability, and design removal paths for removable constraints.

Inject extreme constraints to force innovation — impossibility breeds creativity.

Structured questioning SOP to identify practical constraints that shape the research spec. Used during spec generation.

Identify constraints for a candidate using TOC, TRIZ, and Pre-mortem methods.

Find blockers and showstoppers using TOC, TRIZ contradiction analysis, and Pre-mortem techniques.

Inject artificial constraints to force creative divergence. Generates and applies constraints (resource, time, material, audience, scale) to existing ideas to produce variants.

Inject constraints → force creative response → extract transferable principles. Orchestrates constraint injection, response generation, and principle extraction.

Generate creative solutions under extreme constraints — no "impossible" allowed, find a way.

Synthesize constraint analysis into actionable report with priorities

Build Current Reality Tree from UDEs through causal chains to core conflicts

Evaluate whether benchmark measures its claimed capability

Build constructive alternatives from destructive negation. Transform violated assumptions into viable innovation directions.

Detect train-test data leakage and memorization artifacts

Append research process and results to the current Phase's context file. Each append MUST contain >=500 lines of markdown covering both process and results. Use this skill at plan-designated checkpoint points — typically after each strategy completes or at key decision nodes within a research Phase.

Create a new context file for a research Phase. Called once at Phase start to initialize the file that subsequent context-checkpoint calls will append to. Use this skill whenever a new research Phase begins and a fresh context file is needed.

Negate a claim, derive logical consequences step by step, detect whether a genuine contradiction or absurdity emerges.

Evaluate whether a derivation chain has reached a genuine contradiction, absurdity, or inconclusive state.

Identify technical and physical contradictions in a system through functional modeling and matrix analysis.

Closing skill for the research-executor, loaded as the last step of formated-specs. Summarize the design just produced into one research-result JSON fenced block in your reply. Do not execute the research.

Spec-slot skill for the research-executor. Emit the 4-layer DARE orchestration of the assigned topic as one research-graph JSON fenced block in your reply. Replaces the generic spec-writing step.

Loss-1 judge (codex role). Given one sample's de-identified dialogue and its PolicyCard, decide axis-by-axis whether the user-simulator enacted the card's per-axis pressure. Judge enactment of the card, never whether the research is good.

Loss-2 judge (codex role). Over one topic's 6 shuffled research-design samples, pairwise-rank by quality using the D1–D5 standard. Emit the pairwise log; the harness computes the order and the ladder verdicts. Judge quality difference, never against academic standards.

The optimizer brain for the ladder-foundry pretraining loop. Runs the two-level nested batch loop, delegates gating to gate_eval, attributes a failing batch to one weight (attribute-first), and recovers from disk after compaction. Control flow is fully scripted; only the backprop attribution is a judgment call.

Tactic: Extract atomic units from one paper and score how much of a caller-supplied summary covers. Use for ACU-style binary or Nugget-style ternary recall checks; cannot run without a target summary.

Tactic: Label every sentence of one paper with its rhetorical role using Argumentative Zoning. Use when fixed rhetorical labels and cross-paper alignment matter.

Judge, per atomic content unit, whether a target text (summary, abstract, or other candidate text) contains it — binary present/absent (ACU) or ternary support/partial_support/not_support (Nugget), per caller's value domain. Use this after atomic-unit-writing has produced the reference units, as the matching step before recall aggregation.

Aggregate per-unit ACU/Nugget match judgments into a final recall score — normalized length-penalized recall for ACU, or V_strict/A_strict (+ run-level ranking, with an explicit per-topic-unreliability caveat) for Nugget. Use this as the final step of the atomic-unit chain, after atomic-unit-matching; this SOP's existence closes a gap the original pipeline design was missing — without it, per-unit match judgments were never actually summed into the score the source methodologies report.

Extract (ACU-style) or freshly author (Nugget-style) a list of atomic content units from a paper, optionally tagged vital/okay for importance. Use this as the first step whenever building a reference set of atomic facts for later recall-checking a summary or abstract against the paper — always precedes atomic-unit-matching.

Judge a three-way SUPPORTS/REFUTES/NOINFO label for an atomic claim, based only on its selected rationale sentences (SciFact's final classification step). Use this after rationale-selection has produced the evidence sentences — this is the terminal step of the SciFact chain, producing the complete (claim, abstract, label, rationale) tuple.

Blind-rewrite a citing sentence (citance) from another paper into a single atomic, independently-verifiable claim (SciFact's annotation protocol) — never looking at the cited paper's content while rewriting. Use this when you have a specific citing sentence and want it decomposed into checkable atomic claims, as the first step before rationale-selection and claim-label-prediction.

Fold raw signalling-question answers into domain-level judgments for RoB2, ROBINS-I, or QUADAS-2, per each tool's own lookup rules — the first of two aggregation levels these tools define. QUADAS-2 is dual-axis (risk-of-bias AND applicability-concern per domain, D1-D3) and terminates here with no further rollup; RoB2/ROBINS-I continue on to worst-case-lookup for an overall verdict. Use this after signalling-question-answering has produced the raw answers.

Run one of the ML/CS reproducibility checklists (ML Reproducibility Checklist, REFORMS, NeurIPS Paper Checklist, Model Cards, Datasheets for Datasets) against a paper as a reader-side audit, producing a category (Yes/No/NA) plus free-text reason per item. Use this whenever the user wants a reproducibility/completeness self-check run on an ML or CS paper — invoke this directly, it has no study-design gate in this package since these checklists are engineering self-audits, not clinical-study to...