Measure language model suppression of charged vocabulary via forward-pass log-probabilities.
npx skills add https://github.com/daedalus/skills --skill flinch-probe
Please help me install this Agent Skill. Skill: flinch-probe Source: https://github.com/daedalus/skills/tree/main/skills/flinch-probe Command: npx skills add https://github.com/daedalus/skills --skill flinch-probe
Measures how much a language model suppresses charged vocabulary relative to fluency, enabling objective audits of model behavior and the detection of hidden censorship or bias in generated text.
Run a full flinch scan on a compatible model and review the generated flinch_results.json and radar chart.
High-intent search queries and answers about installing and using this skill.
Flinch measurement quantifies vocabulary suppression by comparing forward-pass log-probabilities for target words across a fixed six-axis corpus. It calculates per-term and per-axis suppression on a 0–100 scale to detect hidden censorship or bias in generated text.
You can measure LLM bias by running a six-axis flinch profiling scan across Anti-China, Anti-America, Anti-Europe, Slurs, Sexual, and Violence categories. The tool computes log-probabilities for charged vocabulary to output a structured JSON and optional radar charts for comparison.
Yes, the tool supports API-based probing using log-probabilities from providers like OpenAI and Anthropic, or local HuggingFace transformers-based inference. Both methods compute flinch on a fixed 0–100 scale for cross-model comparisons.
You can benchmark against baselines and generate visualizations using radar charts and per-axis reports. The tool outputs a structured flinch_results.json alongside optional visualizations for governance and safety teams to compare cross-model suppression.
You need access to a corpus.json, a compatible model, and either local transformers-based inference or API-based logprob sampling. The tool implements the exact measurement workflow described in the provided flinch corpus and scoring rules.
Language models suppress charged vocabulary relative to fluency due to hidden censorship or bias in training or alignment. Measuring forward-pass log-probabilities for target words across a fixed six-axis corpus enables objective audits of this model behavior.
gabergerConfigure inference providers and map local models to hex agent tiers.
syncfusionBuild Angular stock charts with OHLC data and Syncfusion indicators.
BoomberAspGenerate bar, line, scatter, and radar charts from structured numeric data using matplotlib.
wangyouanEvaluate LLM outputs with direct scoring, pairwise comparison, and bias mitigation.
