Muir Robotics

A 17-agent pass generated 161 candidates across ten segments, screened 159 against the nine gates cheapest-kill-first, and adversarially verified every one that cleared. All of them died. That is the second consecutive pass to kill every top-scoring candidate — 5 of 5 last time, 5 of 5 this time.

Which gate does the killing

The histogram is itself a finding — it says where the space is actually thin.

The frontier — the five that cleared screening, and what killed each

With nothing surviving, this is the useful list: how far the search actually got. Every kill carries a primary source.

1. Hand-lettered baseballs C088 Incumbents publish $7.81/ball at 96–239 units with names at +$1.80 — the $22 price was off 2.3×. A fixed CO2 head already does panel-bounded lettering. $490 → $182/hr
2. Domed pet-urn lids C140 The wholesale channel is owned end-to-end by a supplier that already bundles engraving, photo printing and white-label, and sells the whole urn. $228 → $81/hr
3. Beach-stone memorial pendants C137 Drilling stone is force-regulated — "let the drill do the work," ease off at breakthrough. That is the force-feel instant kill, and no configuration escapes it. $165 → $21/hr
4. River-stone memorial mementos C036 The incumbent's order form takes three fields and engraves one message across every stone. Variation is per-order, not per-unit — the amortized-setup case. $162 → $87/hr
5. Airbrushed custom crankbaits C096 $11 was the retail ceiling, not wholesale; paint-only labor clears at $5/lure. Weihai OEMs do custom colors under $2.50 at MOQ 100. $81 → $30/hr

Three kills that attacked the framework, not the candidate

Worth more than the candidate list, and they change how the next round runs.

1. Cube CourierPick cube to bin — first imitation-learned policynext
2. Say-It-SortLanguage-conditioned sortingqueued
3. Tic-Tac-Toe TitanStructured play, multi-step planningqueued
4. Washcloth ValetFirst deformable attemptqueued
5. Claude-at-the-ControlsMCP orchestrator over the policy layerqueued

• Measure the rig on arrival. Neither Seeed nor TheRobotStudio publishes reach, payload, or repeatability. Every threshold in the screen rests on an unmeasured assumption. 30-replay repeatability at full extension, backlash, thermal drift over 60 minutes, payload to failure.
• Establish MTBI. Mean cycles between interventions on a 100-cycle unattended run. Nobody has published this for hobby hardware — it gates every form factor and is the first publishable artifact.
• Accounts. Hugging Face and RunPod, both free tier.
• Scaffold cube-courier repo. LeRobot v0.6 environment, session scripts, eval harness.
• Verification backlog. Eight open rows in the ledger — guest counts, sold prices, blank costs, and whether Arabic and Urdu script handling is genuinely underserved. Guest-count structure is now confirmed qualitatively; the numbers are still unsourced.
• Revise the screen to v2. Three of the five frontier kills attacked the framework: fold in the belly lemma, narrow adapt-to-workpiece to interiors and non-flat blanks, and split Gate 2 into per-unit versus per-order variation.
• Round 2 generation, narrowed. Search radius that changes along the toolpath, interiors and undercuts, non-solids-of-revolution, and genuinely per-unit variation — and check the incumbent's published volume price ladder at Gate 1, not Gate 9. Three of five kills were economic and every one had a competitor ladder the estimate had missed by 2–3×.