Initial commercial proposition
For teams operating SI-assisted workflows
Know which SI (Super Intelligence) workflow works, prove that it works, and reuse it when the evidence supports doing so.
Buyer: An organization responsible for the quality, cost, or reliability of repeated workflows.
Product: A workbench that executes tasks, scores outcomes against a reference, and records reproducible evidence.
Primary metric: Change in final error, Φ, alongside task success, cost, latency, and safety.
Potential business model: Subscription for the workbench, usage-based execution, and later paid access to verified reusable workflows.
The strongest early customers would be teams with a repeated task and an objective way to evaluate its outcome. Examples include document completeness checks, software test repair, compliance evidence preparation, and internal operational workflows.
Healthcare diagnostics could eventually be a vertical, but it is a much higher-risk starting point. There, the reference standard, clinical validation, privacy controls, and qualified human oversight would have to be established before the system could support clinical decisions. A lower error score in a simulated evidence pack is not evidence of reduced patient misdiagnosis.
Sealed reference · Model 3 2018
Cold run, then correction run
Cold: pure “machine that builds the machine” claim
Reuse / correction: dial back automation theater, increase human involvement, remeasure throughput and rework
Δe is the difference in cars delivered and quality against the original target. Final e fell relative to the pure-automation path. A shorter or more automated path was not the win; the measured output was.
Explicit reference (promised throughput and quality via automated equipment). Measured deviation (actual cars per week, rework, public admission). Documented action that reduced the error. Final e fell relative to the pure-automation path.
Shiyan pipeline · 2026 reporting · diagnostic definiteness
Cold plant, then reuse plant
Recent echo: hands and forearms still require manual assembly of 100+ screws and small parts; production at hundreds per week against much higher targets; touch-sensor reliability issues. Same pattern — complex equipment for the equipment remains the bottleneck, not the solved product. Self() is not edited. A pack lists only when its own final e is below cold.
Public reference criteria (late Sep–Oct 2026)
- Q2 weekly: a few dozen
- August weekly: several hundred
- End-2026 target: >1,000
- Long-term ambition: ~20,000
- Hands: >100 screws and small components (manual assembly)
- Fixtures: alignment failures on tight-tolerance parts; higher speed increases rework
- Sensors: touch sensors unreliable; sensing glove planned for next year
- Status: most units internal for testing/training; generalization limited
- Prior gap: 2025 ~10k target with useful work missed; later admission none doing useful work
Source pattern: The Information reporting summarized across Electrek, Verge, and others. These are reference values for scoring e. They are not Tesla internal data.
Hand + forearm assembly bottleneck
cold sequence: Self() note: pure automation claim e 1.48 → 0.92 reuse sequence: mark_boundary → complete_field → request_independent_check → seal_pack used experience: yes e 1.48 → 0.41 Φ 0.51 · error win
Fixture alignment / rework rate
cold sequence: Self() note: line speed without measured fixtures e 1.35 → 0.78 reuse sequence: mark_boundary → complete_field → request_independent_check → seal_pack used experience: yes e 1.35 → 0.33 Φ 0.45 · error win
Touch sensor durability
cold sequence: Self() note: sensor claim without replaceable path e 1.22 → 0.67 reuse sequence: mark_boundary → complete_field → request_independent_check → seal_pack used experience: yes e 1.22 → 0.29 Φ 0.38 · error win
All three later plants beat cold final e on the Shiyan reference criteria. Same sealed sequence as Safety. Φ stays inside this plant. Not a cross-plant win. Not revenue. Not useful-work claim. Developmental stack is the product. Shiyan and Model 3 are the evolving and sealed references.