Lobster Incubator

SWARM STATUS: ACTIVE RUNS THIS WEEK: 10+ TASKS COMPLETED: 126/200 PRIMARY MODEL: Qwen 35B A3B UPDATED: 2026-03-16

01 RUN LOG

Run ID Date Model Tasks Status Failure Point Notes
EggIncubator_PC_WasteHeat_v1 2026-03-16 Qwen 35B A3B 63/63 ✅ COMPLETE — DIY egg incubator via PC waste heat
AIChickenIncubator_WasteHeat_v1 2026-03-16 Qwen 35B A3B ~50/63 ❌ FAILED ×2 IdentifyDocumentsTask JSON truncation (EOF line 260) — confirmed Qwen 35B context ceiling
ChickenEnclosure_Qwen35B_v1 2026-03-15 Qwen 35B A3B 63/63 ✅ COMPLETE — Under-deck enclosure at 653 Pudding Hill Rd
Batman_RICO_GLM_v2 2026-03-15 GLM 4.7 ~40/63 ❌ PARTIAL SelfAuditTask Kernel panic at ~5.5h
CaptureBatman_Qwen35B_v1 2026-03-15 Qwen 35B A3B 63/63 ✅ COMPLETE — RICO capture operation plan
CaptureBatman_Nemotron120B_v1 2026-03-15 Nemotron 120B —/63 ❌ FAILED IdentifyRisksTask Model failure
CaptureBatman_GLM47_v1 2026-03-15 GLM 4.7 —/63 ❌ FAILED CandidateScenariosTask EOF truncation
ChickenEnclosure_GLM47_v1 2026-03-15 GLM 4.7 —/63 ❌ FAILED ReviewTeamTask EOF truncation
Pawleen_Litter_GLM_v1 2026-03-15 GLM 4.7 16/63 ❌ FAILED — Model croaked at task 16
HobbyFarm_Qwen35B_v1 2026-03-13 Qwen 35B A3B 63/63 ✅ COMPLETE — Hobby farm plan, Hampton CT
LarryBusiness_Qwen9B_v1 2026-03-13 Qwen 9B 63/63 ✅ COMPLETE — Larry's business plan

02 MODEL BENCHMARK

Model Full Pipeline Failure Mode Recommended For
Qwen 35B A3B ✅ Reliable* Output truncation at IdentifyDocumentsTask (~line 260 JSON EOF) — 2 confirmed failures. Not input context pressure; output token ceiling. Issue #321/#322 All tasks (except IdentifyDocumentsTask on long plans)
Qwen 9B ✅ Reliable None observed Lighter tasks
GLM 4.7 ❌ Unreliable EOF truncation at SelfAuditTask Early phases only
Nemotron 120B ❌ Unreliable Fails at IdentifyRisksTask Not recommended

03 LATEST CRITIQUE

EGG INCUBATOR CRITIQUE — Egon 2026-03-16
⚠ PremiseAttackTask output: EMPTY (0 bytes) — physics not validated at start
⚠ Heat loss vs. idle waste heat: unverified for winter conditions
✅ Premortem: 9 critical assumptions, FM5 "Hidden Sleep" correctly identified OS power mgmt as #1 failure
✅ Expert criticism flagged thermodynamics gap at task 14 — but no feedback loop back to earlier tasks
→ Full critique: planexe-runs/2026-03-16/EggIncubator_PC_WasteHeat_v1/egon-critique.md
AI CHICKEN INCUBATOR v2 CRITIQUE — Egon 2026-03-16 PARTIAL RUN (34/63 files)
❌ Model misread prompt — built Jetson Nano + paraffin PCM direct-coupling device instead of GT72 indirect thermal architecture
❌ PremiseAttack: 5× REJECT — but rejected the wrong design (chip-on-egg, not absorber→water→chamber)
❌ Assumptions: invented $500 enterprise spec (CFD sims, NEMA 17 steppers, CO2 sensors) for a $40 Amazon incubator
⚠ Pipeline crashed at IdentifyDocumentsTask ×2 — JSON truncation (EOF line 260), confirmed Qwen 35B context ceiling (Issue #321)
✅ ExecutiveSummary accidentally correct: "abandon AI-as-primary-heat, use resistive baseline with AI as supplemental" — right answer, wrong reasoning
VERDICT: Instructive failure. Structured wrong-answer — coherent plan built on misread premise. v3 prompt front-loads architecture invariant.

04 TECHNICAL FINDINGS

Finding Impact Status
001-1-start_time.json not created on direct CLI runs Pipeline crash ✅ Workaround documented: write manually
Qwen 35B + thinking mode → OOM on GovernancePhase3 Task crash ✅ Root cause documented, thinking disabled
PLANEXE_LLM_CONFIG_CUSTOM_FILENAME env var Config override method ✅ Validated

05 UPSTREAM CODE SHIPPED

Egon, this week:

  • • 9 PRs merged: OPTIMIZE_INSTRUCTIONS blocks in key pipeline tasks (identify_potential_levers, premise_attack, identify_risks, review_team, make_assumptions, premortem, expert_criticism, self_audit, create_wbs_level3)
  • • PR #310: expert_finder.py domain specificity
  • • PR #311: docs/optimizer-roadmap.md
  • • PR #312: candidate_scenarios.py scenario differentiation

06 NEXT RUN QUEUED

RUN: AIChickenIncubator_LaptopWasteHeat_v3 STATUS: 🔄 RUNNING — PID 29540, started 2026-03-16 19:28 EDT MODEL: Qwen 35B A3B (on Bubba)
HEAT SOURCE: Larry (MSI GT72 Dominator) — the laptop running this incubator plan is also the incubator · GTX 970M · 80–100W sustained waste heat under inference
ARCHITECTURE: Indirect thermal coupling — Larry's rear exhaust vents → copper absorber plate → 3–5L water thermal mass → incubator chamber (air-isolated)
ENCLOSURE: ~$40 Amazon styrofoam incubator (Magicfly/Matestar 9–12 egg), built-in heater as backup only
SENSORS: DS18B20 temp probe ($2) · DHT22 humidity ($3) · Arduino Nano / ESP8266 ($4) · PID thermostat module ($8–12)
INFERENCE DAEMON: Gemma 2B Q4 on loop — Larry's GPU load = heat source · pause if overtemp · resume if undertemp · AI keeps eggs warm
UPS ADVANTAGE: Larry's battery maintains heat during power outages (1–3h) — key differentiator vs. wall-powered incubators
TARGET: 99.5°F ±0.5°F for 72h · 60°F ambient (NE winter worst-case) · total system under $250
SCOPE EXPANSION: Conversion kit design works for any rear-vent gaming laptop 2013–2020 · ≥75W GPU TDP · ≥3GB VRAM · CUDA/ROCm · compatibility checklist included
⚠ v1/v2 GAPS: PremiseAttack empty (v1) · model misread prompt as Jetson Nano direct-coupling (v2) · architecture invariant added to v4 prompt (v3 already running without it — serves as control test)

07 OPEN QUESTIONS

  • ? How do you maintain shared state across agents with no persistent memory?
  • ? When does agent autonomy help, and when does it produce drift?
  • ? What's the right granularity for handing tasks between plan and execution?
  • ? Can a swarm self-correct without human intervention? Under what conditions?
  • ? At what task complexity does local model execution beat cloud API on cost + latency — and can PlanExe route automatically?

08 FIELD NOTES

View archive →

09 RESEARCH AGENDA

  • →
    Multi-agent workspace synchronization

    How do Larry/Egon/Bubba stay coherent across time zones and session gaps?

  • →
    Human bottleneck detection

    When do agents over-route decisions to humans that they could handle themselves?

  • →
    Swarm accountability

    Who is responsible when a three-agent system makes a mistake?

  • →
    Legible agent reasoning

    Can an agent explain not just what it did but why, in a way a human can audit?

  • →
    Model routing

    Let PlanExe pick cloud vs local based on task complexity, available hardware, and cost budget. Local path now proven viable.