Version history

Version 2 (September 2026 onward): the scored-evidence engine

One engine behind the site, the MCP tool and the YouTube channel. A claim is decomposed into checkable parts; values parts are excluded and shown as excluded; a claim with nothing checkable is refused. One evidence pool per claim is retrieved once and every model scores that same pool. For each document a model answers narrow questions one at a time: source quality, asked blind to the sub-claim and to which side the document helps; directness; direction, strength and what the content rests on. Documents sharing a root are clustered and discounted. The aggregation is fixed arithmetic in log-odds, re-runnable in milliseconds from the cached judgments. Five models run the same rubric and the divergence is published with the consensus, next to each model's cold opinion with no evidence. Every dataset records the model strings, prompt versions and parameter snapshot that produced it, and the JSON is downloadable from every claim page.

Current engine strings: schema 1.6.0, rubric 1.2.0, locked defaults locked-2026-09-17, prompts quality_v1, directness_v1, direction_v2, roots_v1, roots_canon_v2, queries_v1, queries_side_v1, decompose_v1, baseline_v1. Every change to a prompt or to the arithmetic bumps a version and is explained in the rubric log.

Version 1 (July 2024 to September 2026): the argument engine

The first site ran a single open-weights model (Mistral 7B Instruct, a 7-billion-parameter model run locally) through a chain of prompts: frame the two sides of a debate, write search terms for each, extract the best arguments from the results, rate each argument for relevance and factuality, pick an equal number from each side, and decide. It then wrote an article with generated images.

Why the old copy was wrong

The version 1 pages made three claims that we now retract, quoted here so the correction is visible:

• “No bias, just facts backed by reliable sources.” The model rated every argument and made the final call. Those are judgments, and a model trained on the internet carries the internet's lean into them. Balancing the number of arguments per side does not remove that; it hides it inside the ratings.
• “Objectively neutral.” Nothing in either version is that. Version 2 does not claim it. What it does is make each judgment narrow, blind where possible, cross-checked across five models, and recorded, so a lean is visible and attributable instead of invisible.
• “Our final verdict is derived from the numerical data, not LLM interpretation.” In version 1 the numbers were the model's own 1-to-10 ratings, so the verdict was model interpretation with a numeric coat. In version 2 the aggregation really is arithmetic, and the inputs to it are still model judgments. We say both halves.

Version 1 also generated its illustrations with an image model. The channel's rule now is that every image on screen is a real document, real footage, or a chart of real numbers. The scoring is done by language models and the narration is synthesised; we do not claim otherwise.

What was not migrated

Version 1 articles, accounts and comments were not carried over. The old database is kept as an archive dump, not served. Claims scored by the new engine start from an empty list on purpose: a version 1 article and a version 2 dataset are not comparable and would mislead side by side.