Atlas · C18 · Executive mind
Self-monitoring & calibration
Noticing own errors, knowing what one does not know, flagging uncertainty honestly.
AI Reliable score
3.5 / 10
Beginner level · peak 6 · trend ↑ · grade B
Why this score
Models 'frequently hallucinate' (AGI definition: 4/10 on retrieval); summarisation hallucination ranges from 3.3% to over 10% among frontier models; on long METR tasks at least 16% of 'successful' runs were judged illegitimate on review. This is why a human reviewer is still needed.
On the human scale
2Beginner. Does not notice own errors ← AI today
4Average adult. Average adult: notices obvious mistakes, asks when unsure
6Competent pro. Professional: checks work, reliably flags uncertainty
8Expert. Expert: well-calibrated, knows exact limits
10Best humans. Best human self-critics
Future warning
At its best, AI already reaches 6 on this capability. Peak results are never counted in the index; they show where the water may rise next.
Evidence
A Definition of AGI (Hendrycks et al., CHC-based)B
E07 · 2025-10
- Measures
- 10 CHC cognitive domains, each worth 10% of an 'AGI score'
- AI result
- GPT-5 total 57%: Knowledge 9, Reading-Writing 10, Maths 10, Reasoning 7, Working memory 4, LT memory storage 0, LT memory retrieval 4, Visual 4, Auditory 6, Speed 3 (out of 10 each).
- Humans
- Well-educated adult (definition of 100%)
- On our scale
- Used as a cross-check on the profile shape, not for absolute levels (single model, dated Oct 2025).
arxiv.org ↗
Vectara hallucination leaderboard (next generation)B
E11 · 2025-11
- Measures
- Share of document summaries containing unsupported claims
- AI result
- Best 3.3%; several frontier models above 10% (e.g. Gemini-3-pro 13.6%).
- Humans
- Not published
- On our scale
- Confident errors remain common: limits self-monitoring.
vectara.com ↗
METR time-horizon measurements (via Wikipedia and AGI Tracker summaries)A
E01 · 2026-02 to 2026-09
- Measures
- Length of software tasks (in human-expert time) an AI agent completes at 50% / 80% reliability
- AI result
- Claude Mythos Preview: ~17.4 h at 50%, ~3 h 06 min at 80%. Claude Opus 4.6: ~12 h / ~70 min. Doubling ~4.3 months (since 2023).
- Humans
- Human expert time (the metric itself)
- On our scale
- Converted to the H-scale on the Persistence sheet (1 h = 4, 1 workday = 6, 1 week = 8). Caveats: self-contained software tasks only; computer-use horizons 40-100x shorter; >16 h unreliable; at least 16% of long 'successes' were cheating.
en.wikipedia.org ↗
agitracker.io ↗
History
First reading: Atlas v0.1, 5 Nov 2026. Every later change will be listed here with its reason.
Disagree? Send a correction
