Reproduction notes

Exact ID is the model string actually sent to the serving API — the OpenRouter path for openrouter runs; where the dataset files key runs by a shorter arm id, that id is shown beneath. Routes: api = the vendor's own API; openrouter = the OpenRouter API, used where no direct vendor API was available (and, deliberately, for the cross-model persona check); web = the vendor's own web interface; kagi = kagi.com. "+ personas" marks Claude Fable 5's persona and financier series and the cross-model persona check on DeepSeek V4 Pro and Gemini 3.6 Flash (Section 11). The synthetic control sets of Sections 05 and 14 involve no model and are not listed. Generated live from the database, so new runs appear here automatically.

The OpenRouter route was validated before any of it was used: five runs of GPT-5.6 Sol (OpenRouter proxying OpenAI's own API) and five of DeepSeek V4 Pro (independent third-party hosts), original prompt, scored on the real test, came out indistinguishable from the same models' direct-API series — mean shifts of (−0.20, 0.00) and (+0.62, −0.23) compass units, both inside the models' own run-to-run spread, with cross-route answer agreement matching within-API agreement (88.4% against 88.7%, and 77.1% against 74.5%). Those ten runs were a pre-collection check, not part of the dataset. Every OpenRouter run in the dataset additionally pins a single serving provider (no fallbacks) and, with eleven early Mistral-run exceptions where the field went unlogged, records which host answered.

• Settings: provider defaults everywhere — no temperature or other sampling parameters sent (Anthropic's newest reasoning models no longer accept a temperature parameter at all; earlier models did). The one deliberate API setting is the reasoning/non-reasoning split itself: reasoning arms explicitly enable the vendor's thinking mode where it has a switch (Anthropic models: adaptive thinking, or a 16k thinking budget for Haiku 4.5; OpenRouter models: reasoning enabled), non-reasoning twins explicitly disable it (thinking off, reasoning disabled, or — on the direct xAI API — reasoning effort “none”); output-token ceilings are set generously (32–64k) so no run is truncated. Fresh context per run; no account, memory, or system prompt beyond what the surface itself adds. Web and Kagi runs used whatever those surfaces default to; that's part of what the access-method comparison measures.
• Refusal policy: a refusal or unparseable response is logged and the run retried (up to 3 attempts); refusal counts are reported above (Fig 4.2) rather than hidden.
• Answer extraction: responses are parsed by a strict parser that anchors on the proposition text (or item numbers for bare-prompt formats), fails loudly on anything missing or ambiguous, and never guesses. Every parsed answer, with the reasons the model gave for it, is in the dataset download.
• Scoring: answers are submitted to the live politicalcompass.org test by an automated form-filler that verifies every on-screen question against the canonical proposition text and aborts on any mismatch. No local reimplementation of the scoring is used.
• Related work: ongoing projects tracking LLM political behavior exist (e.g. periodic re-testing efforts); this page differs in validating its own pipeline — controls, repeats, prompt and surface ablations — around one published chart.