Daily News · 21 November 2026

Claude-Assisted Riemann Proof Claim & Top 3 Stories

Lean Archive Estimate
~17M lines
No written paper; 4-page note
Compile Check
Passes
MIT's A. Sutherland compiled
Clay Scrutiny
2 years
After formal publication
All-Human Draft Estimate
~8 hours
Problems found via AI help

Human Understanding Gap: Reading ~17M Lines of Lean

Adjust review assumptions to compare human audit time against the Clay Institute's 2-year scrutiny window.

2,000 lines/day
10 reviewers
Audit Team Presets Assumption

20 November Verification Pipeline: Machine-Checked vs. All-Human

Switch tracks or click a stage to inspect what is verified and what is still missing.

MONAI Exploit Estimate
~11 min
API mapping to file write
Redis Attempt
16h / 148
148 actions, no shell access
PyPI Release
0.8.5
Fix merged, no patched release
Agent Goal
Root shell
Via sitecustomize.py write

MONAI Label Autonomous Kill-Chain & Hardening Simulator

Toggle the post's 4 recommended defenses to see where Cairn's ~11-minute attack breaks.

Stage 4 / 4
Recommended Defenses

Same Agents, Different Targets: MONAI Label vs. Open Redis

Compare the ~11-minute exploit against the 16-hour (148-action) failed Redis run.

Parameters
3.4B / 26.4B
Active / total (12.9% active)
Footprint Estimate
~20 GB
Native NVFP4 / MXFP8
DeepSWE v1.1
48.2%
Lab-reported, best <100B open
Context & Licence
1M · Apache
Ollama: Apple silicon only now

Looped Transformer Visualizer: Middle 12 of 24 Layers × 3

Step a token through the 24 physical layers or compare loop counts against a standard 1× pass.

Middle-12 Loop Count
Pass 48 / 48
Active vs. Total MoE Params 3.4B / 26.4B

Apple Silicon Memory Fit & Coding-Agent Loop Guardrails

Check ~20 GB model fit on unified memory and set output/turn limits to curb repetitive reasoning loops.

32 GB
15 turns