ARC-AGI NECROPSY — W1–W17 CAUSE OF DEATH
| WAVE | APPROACH | RESULT | CAUSE OF DEATH | FOSSIL TYPE |
|---|
| W1–4 | Single runner REPL | 6 solves baseline | N/A — established floor | FOUNDATION |
| W10 | 3 runners simultaneous | 14 solves 3.5% | N/A — best run | BREAKTHROUGH |
| W11 | 4-model cascade | 0 solves | Routing overhead > gain. No Protocol Zero entries. | ROUTING_FAIL |
| W12 | V-Demon + ARC-3 | 0 ARC-1, 0.21× ARC-3 | LLM reasoning timeout on ARC-3. Too slow. | TIMEOUT_DEATH |
| W13 | V16-reflect+snap | 0 solves | Reflection step added tokens without adding insight. | BLOAT_FAIL |
| W14 | V18-squeeze | 0 solves | Squeeze fires correctly, but no exact solve — cell match only. | PARTIAL_APPROACH |
| W15 | V22 Normal Force | 0 solves | Physics model correct but boundary conditions incomplete. | PHYSICS_MISS |
| W16 | V23 compact | 0/60 0% | msclo RTX 5090 saturated. Hardware ceiling hit. | HARDWARE_CAP |
| W17 | V17 Grand Unification | TRUNCATED 11/30 | msclo :11435 timeout. W18 needed. V24+V25 unvalidated. | PENDING |
LEAN4 PROOF GAP NECROPSY — 5 HONEST SORRYS
Session: 96 → 5 sorrys. Phases 2–8 kernel-verified. 91 proofs closed. The 5 remaining are Family B: all RH-dependent. Not failures — honest named gaps at the Godel boundary.
| SORRY | FILE | CAUSE | CANON SYM | STATUS |
|---|
| li_lambda | AxiomElimination.lean | Needs ordered zero sum definition | WLD | HONEST GAP |
| lambda_positivity | AxiomElimination.lean | Proved IF RH holds. Conditional. | FEP | RH-DEPENDENT |
| zeta_zeros_orderable | AxiomElimination.lean | Requires Re=½. IS the RH. | FOF | RH BOUNDARY |
| zetaZeroImPart | AxiomElimination.lean | Ordered enum needs Re=½ constraint | H=H† | RH-DEPENDENT |
| zeta_zero_at_gamma_1 | AxiomElimination.lean | ζ(½+14.134725i)=0 needs certified computation | γ₁ | NEEDS CERT |
These 5 are not failures. They are the honest shape of the Riemann Hypothesis in negative space. The machine traced the boundary of what it cannot assert without proof. 96 → 5 is the success.
FLEET ARCHITECTURE NECROPSY — WHAT DIED AND WHY
| SYSTEM | WHAT IT WAS | CAUSE OF DEATH / GAP | LESSON |
|---|
| campfire-heal-engine | Node.js heal loop for fleet events | node:22-alpine wrong entrypoint | Entrypoint must match image OS |
| msclo-gateway v3.2 | msclo portal bridge | Wrong GATEWAY_URL port :18830 vs :18832 | Always verify WS port at recreate time |
| PTTP Redis | Traffic tracking to portal | REDIS_URL not wired to K8s deployments | Need: kubectl set env for pemos-portal |
| ct-builder-gateway | CT cluster openclaw gateway | mkdir /root/.openclaw Permission denied | Need HOME=/data image fix |
| model-gateway | PEMCLAU router | Scaled to 0. Stub dormant. | Need scale-up + PEMCLAU wiring |
| gpu-test-a10 pod | A10 GPU test | Pending 8+ days — no GPU quota | Request quota increase or use msclo RTX |
| forge BIOS 4301 | i9-14900KS performance | CPU clocked 3200MHz not 6200MHz rated | Flash 4301→4505 to unlock full boost |
| lianli01 offline | forge WSL asleep | Windows host up, WSL sleeping | Kay must run `wsl` on Windows terminal |
ARCHITECTURE LESSONS — WHAT THE FOSSILS TEACH
WHAT WORKS
• 3+ runners simultaneous (W10) > 1 runner with better model
• Protocol Zero immunization: name the wall permanently
• Static embed binary: zero runtime deps, instant cold start
• Redis streams (campfire:events) > polling for fleet events
• Python direct write > bash heredoc for large files
• Thistlethwaite descent: restrict operators layer by layer
• One file per sorry: name the gap, don’t hide it
WHAT DIES
• Single large model doing everything (timeout)
• Heredoc for HTML > 200 lines (truncation)
• Subagents for HTML builds (3 min timeout, 0 output)
• Reflection steps that add tokens not insight
• 32b model on FLOOD_FILL_ENCLOSED (capacity wall)
• Hardcoded stats without live API fallback
• Containers without explicit network flags (502 hell)