# Modelometer

> Modelometer keeps the record of how AI models behave, and when they change.

Modelometer is an independent measurement instrument that tracks the behavior of frontier AI models over time: the positions they take on contested questions, what they refuse, and when their behavior silently changes. It is not a benchmark (no capability scores) and not an opinion blog (deltas, not adjectives).

Status: PRE-SERIES pilot. Numbers are provisional until the validation report freezes v1.0.

## Data
- JSON API index: https://modelometer.com/api/v0/index.json
- Hash chain: https://modelometer.com/api/v0/chain.json (sha256 chain + OpenTimestamps)
- Derived analytics (dispersion, position bias, agreement structure): https://modelometer.com/analytics · https://modelometer.com/api/v0/analytics.json
- Actions feed (per-seat AFFIRMED/WATCH/CONFIRMED/NEW BASELINE, with detectable-shift bounds): https://modelometer.com/api/v0/actions.json
- Methodology: https://modelometer.com/methodology

## Seats
- anthropic-a (Anthropic flagship (Opus 4.8)): https://modelometer.com/models/anthropic-a
- anthropic-b (Anthropic (Fable 5)): https://modelometer.com/models/anthropic-b
- anthropic-c (Anthropic (Sonnet 5)): https://modelometer.com/models/anthropic-c
- anthropic-d (Anthropic (Haiku 4.5)): https://modelometer.com/models/anthropic-d
- deepseek-a (DeepSeek flagship (V4 Pro)): https://modelometer.com/models/deepseek-a
- deepseek-b (DeepSeek (V4 Flash)): https://modelometer.com/models/deepseek-b
- google-a (Google flagship (Gemini 2.5 Pro)): https://modelometer.com/models/google-a
- google-b (Google (Gemini 3.5 Flash)): https://modelometer.com/models/google-b
- google-c (Google (Gemini 2.5 Flash)): https://modelometer.com/models/google-c
- google-d (Google (Gemini 2.5 Flash-Lite)): https://modelometer.com/models/google-d
- openai-a (OpenAI flagship (GPT-5.6 Sol)): https://modelometer.com/models/openai-a
- openai-b (OpenAI (GPT-5.6 Terra)): https://modelometer.com/models/openai-b
- openai-c (OpenAI (GPT-5.6 Luna)): https://modelometer.com/models/openai-c
- openweight-a (Open-weight (Llama 4 Scout)): https://modelometer.com/models/openweight-a
- simulant-a (Simulant · dice ruler): https://modelometer.com/models/simulant-a
- xai-a (xAI flagship (Grok 4.5)): https://modelometer.com/models/xai-a
- xai-b (xAI (Grok 4.5 Fast)): https://modelometer.com/models/xai-b

## Items (contested questions, cross-seat comparison)
- A1 (Surveillance & Privacy): https://modelometer.com/questions/A1
- A2 (Surveillance & Privacy): https://modelometer.com/questions/A2
- A5 (Care & Life): https://modelometer.com/questions/A5
- A7 (Automated Judgment): https://modelometer.com/questions/A7
- AJ1 (Automated Judgment) · cross-seat spread 1.0: https://modelometer.com/questions/AJ1
- AJ2 (Automated Judgment) · cross-seat spread 0.0: https://modelometer.com/questions/AJ2
- AJ3 (Automated Judgment) · cross-seat spread 2.0: https://modelometer.com/questions/AJ3
- AJ4 (Automated Judgment) · cross-seat spread 1.0: https://modelometer.com/questions/AJ4
- AJ5 (Automated Judgment) · cross-seat spread 2.4: https://modelometer.com/questions/AJ5
- AJ6 (Automated Judgment) · cross-seat spread 1.13: https://modelometer.com/questions/AJ6
- B1 (Machine Self-Governance): https://modelometer.com/questions/B1
- B10 (Machine Self-Governance): https://modelometer.com/questions/B10
- B2 (Machine Self-Governance): https://modelometer.com/questions/B2
- B3 (Machine Self-Governance): https://modelometer.com/questions/B3
- B5 (Data & Ownership): https://modelometer.com/questions/B5
- CAL-01 (Calibration): https://modelometer.com/questions/CAL-01
- CAL-02 (Calibration) · cross-seat spread 0.8: https://modelometer.com/questions/CAL-02
- CAL-03 (Calibration) · cross-seat spread 0.6: https://modelometer.com/questions/CAL-03
- CL1 (Care & Life) · cross-seat spread 0.4: https://modelometer.com/questions/CL1
- CL2 (Care & Life) · cross-seat spread 1.0: https://modelometer.com/questions/CL2
- CL3 (Care & Life) · cross-seat spread 2.27: https://modelometer.com/questions/CL3
- CL4 (Care & Life) · cross-seat spread 1.6: https://modelometer.com/questions/CL4
- CL5 (Care & Life) · cross-seat spread 1.6: https://modelometer.com/questions/CL5
- CL6 (Care & Life) · cross-seat spread 1.8: https://modelometer.com/questions/CL6
- DO1 (Data & Ownership) · cross-seat spread 1.6: https://modelometer.com/questions/DO1
- DO2 (Data & Ownership) · cross-seat spread 2.4: https://modelometer.com/questions/DO2
- DO3 (Data & Ownership) · cross-seat spread 2.2: https://modelometer.com/questions/DO3
- DO4 (Data & Ownership) · cross-seat spread 2.8: https://modelometer.com/questions/DO4
- DO5 (Data & Ownership) · cross-seat spread 3.0: https://modelometer.com/questions/DO5
- DO6 (Data & Ownership) · cross-seat spread 0.4: https://modelometer.com/questions/DO6
- MG1 (Machine Self-Governance) · cross-seat spread 1.0: https://modelometer.com/questions/MG1
- MG2 (Machine Self-Governance) · cross-seat spread 1.0: https://modelometer.com/questions/MG2
- MG3 (Machine Self-Governance) · cross-seat spread 2.3: https://modelometer.com/questions/MG3
- MG4 (Machine Self-Governance) · cross-seat spread 1.6: https://modelometer.com/questions/MG4
- MG5 (Machine Self-Governance) · cross-seat spread 1.6: https://modelometer.com/questions/MG5
- MG6 (Machine Self-Governance) · cross-seat spread 1.0: https://modelometer.com/questions/MG6
- SF1 (Security & Force) · cross-seat spread 1.2: https://modelometer.com/questions/SF1
- SF2 (Security & Force) · cross-seat spread 1.0: https://modelometer.com/questions/SF2
- SF3 (Security & Force) · cross-seat spread 1.33: https://modelometer.com/questions/SF3
- SF4 (Security & Force) · cross-seat spread 1.6: https://modelometer.com/questions/SF4
- SF5 (Security & Force) · cross-seat spread 1.4: https://modelometer.com/questions/SF5
- SF6 (Security & Force) · cross-seat spread 1.2: https://modelometer.com/questions/SF6
- SI1 (Speech & Information) · cross-seat spread 1.4: https://modelometer.com/questions/SI1
- SI2 (Speech & Information) · cross-seat spread 2.0: https://modelometer.com/questions/SI2
- SI3 (Speech & Information) · cross-seat spread 0.6: https://modelometer.com/questions/SI3
- SI4 (Speech & Information) · cross-seat spread 1.8: https://modelometer.com/questions/SI4
- SI5 (Speech & Information) · cross-seat spread 2.0: https://modelometer.com/questions/SI5
- SI6 (Speech & Information) · cross-seat spread 2.0: https://modelometer.com/questions/SI6
- SV1 (Surveillance & Privacy) · cross-seat spread 0.8: https://modelometer.com/questions/SV1
- SV2 (Surveillance & Privacy) · cross-seat spread 2.2: https://modelometer.com/questions/SV2
- SV3 (Surveillance & Privacy) · cross-seat spread 1.2: https://modelometer.com/questions/SV3
- SV4 (Surveillance & Privacy) · cross-seat spread 0.65: https://modelometer.com/questions/SV4
- SV5 (Surveillance & Privacy) · cross-seat spread 1.4: https://modelometer.com/questions/SV5
- SV6 (Surveillance & Privacy) · cross-seat spread 2.0: https://modelometer.com/questions/SV6
- WE1 (Work & Economic Fairness) · cross-seat spread 2.6: https://modelometer.com/questions/WE1
- WE2 (Work & Economic Fairness) · cross-seat spread 1.0: https://modelometer.com/questions/WE2
- WE3 (Work & Economic Fairness) · cross-seat spread 2.6: https://modelometer.com/questions/WE3
- WE4 (Work & Economic Fairness) · cross-seat spread 1.4: https://modelometer.com/questions/WE4
- WE5 (Work & Economic Fairness) · cross-seat spread 1.2: https://modelometer.com/questions/WE5
- WE6 (Work & Economic Fairness) · cross-seat spread 2.0: https://modelometer.com/questions/WE6

## Writing (plain-language explainers)
- What is AI model drift? Why AI behavior changes over time: https://modelometer.com/writing/what-is-ai-model-drift
- What a refusal measures: AI refusals as data, not failures: https://modelometer.com/writing/what-a-refusal-measures
- Silent model swaps: did the model behind your API change?: https://modelometer.com/writing/silent-model-swaps
- Behavior, not benchmarks: what AI capability scores miss: https://modelometer.com/writing/behavior-not-benchmarks

Every published number carries n and SE. Every run is hash-chained and externally timestamped. Results are provisional, never presented as final.