How AI spots CEO spin on earnings calls
When profits drop, executives rarely use sad words—they dodge questions and pivot to upbeat buzzwords. We tested a 4-judge communication pipeline with experience memory and a built-in referee across 35 blind historical calls, 24 skill trials, and 11 live October 2026 earnings calls.
Gemini 3.1 Pro (Antigravity v2.0) · Transcript judges (temperature=0.0): deepseek/deepseek-v4.1-flash, typesafe/jev-router, qwen/qwen3.8-flash, local/structural-linguistic-v1.
1. Blind test on 35 past calls (names and dates hidden)
Comparing positive/negative word counting against blank-slate AI and our memory + referee-equipped agent.
Why it worked: Blank-slate AI missed 6 calls where companies were in government merger reviews (FTC/DOJ) that legally restrict executives from discussing future savings. Experience memory (INT-004) recognized that legal exception, while referee skills (SKILL-001, SKILL-002) cut 30-day forecast error by 60% (3.13% → 1.24%, p = 0.000004).
2. Eleven live October 2026 earnings calls
Select a company to view its plain-English summary, 4-judge scores, and referee check.
3. Test any transcript excerpt
Paste an analyst question and executive answer or load a preset.
4. Exact model versions & full paper
All code, cached multi-model annotations, and datasets are verifiable via SHA-256 checksums.
| Role | Model Version | Platform | Settings |
|---|---|---|---|
| Lead Research Agent | Gemini 3.1 Pro (Oct 2026) |
Google Antigravity for Science (v2.0) | Agentic verification |
| Judge 1 (Dodging) | deepseek/deepseek-v4.1-flash |
OpenRouter API v1 | temperature = 0.0 |
| Judge 2 (Spin Check) | typesafe/jev-router |
OpenRouter API v1 | temperature = 0.0 |
| Judge 3 (Blame & Plan) | qwen/qwen3.8-flash |
OpenRouter API v1 | temperature = 0.0 |
| Judge 4 (Baseline) | local/structural-linguistic-v1 |
Local Python 3.12 | Deterministic |