Remove <think>...</think> blocks from thinking model output.

Test a single probe on a single model. Returns (probe_idx, model_label, correct).

Escape special LaTeX characters in a string.

Estimate parameters using subsets of tiers.

Count same-label entities that share a description keyword with the intended entity (i.e. are plausibly the same TYPE) — excludes unrelated homony

Check that known derived pairs have high FPS and known independent pairs have low FPS.

Returns enriched candidate dict or None.

Return ordered consecutive-pair metrics within a series.

Compare official MMLU/MMLU-Pro/GPQA Diamond/SimpleQA scores vs IKP as a parameter-count proxy. Source data: data/benchmarks/benchmark_scores.csv

Densing-Law falsification data. Source: data/densing_analysis_data.csv Fits pen_acc ~ beta0 + beta1*log10(N_B) + beta2*months and compares be

Within-family + cross-vendor fingerprint metrics for the Fingerprint page. Source: results/comprehensive_fingerprint_results.json

Within-family generation trajectories + GPT-5 family stratification. A generation is identified by family name. We extract all models in the same

Vendor-level hallucination rate on T5-T7 probes. hallucination rate per-model per-tier = wrong / total (NOT wrong / (wrong+correct+refusal) since

Top-level index summarizing what data is available.

Compact list of all models for the table view.

Per-model summary including sample probes per tier (correct/wrong/refusal samples).

Pivot all model responses by tier — each file ~5MB. Schema: { probes: { probe_id: { question, answer, domain, responses: { model: {response, verd

Pipeline narrative — describes the 7-stage pipeline with sample artifacts.

All probes with metadata, plus across-model correct- and hallucination-rates. correct_rate = fraction of evaluated models that answered corre

Researcher recognition vs citations for the Recognition page.

Thinking mode pair comparisons.