PRD: LLM true-compute-cost investigation (spec only)

[RaapTechllc]

Summary

Docs-only. Adds exactly two files and no product code:

1. docs/research/llm-compute-cost-research-notes-2026-10-02.md — verbatim copy of Kyle's frozen research notes (sha256 41783be1…fe575).
2. docs/PRD-llm-compute-cost.md — the PRD.

The idea (Kyle Raap, RaapTech) has been reframed from a "parameters ÷ price" Pareto chart into an investigation of what labs actually pay for inference compute. The ratio and the Pareto chart become estimation tools. The PRD is built around the notes' re-ranked sparks:

• D1 Open-Weight Provider Markup Index (baseline / calibration layer)
• D2 Lab True-Cost Range Report (research; ranges only, never a point)
• D3 Value Frontier with Size Lens in minibench (by-product; existing _pareto_frontier and /models chart unchanged, size as a marker channel)

Hard rules followed

• The research notes are the only source of market facts. Every price, parameter count, deal figure and statistic cites notes §n. No web search, nothing invented.
• FACT / ASSUMPTION / UNKNOWN labels preserved; no UNKNOWN or hypothesis upgraded. New design choices are labelled PROPOSED.
• Kyle's hypotheses (notes §11) are reproduced byte-for-byte in Appendix A at exactly the notes' verdicts.
• Output weighting is a configurable workload profile with output-only as a stress test, not the default.
• Hardware ownership is handled as a methodology risk and open decision (D2), not an MVP item.
• Fit verdict presented as open decision D6: minibench hosts the value-frontier view; the true-cost core may belong in a separate research repo/notebook with only cited outputs republished.

Counts

• 29 numbered, checkable acceptance criteria
• 12 open decisions for Kyle (6 from notes §12 + 6 raised by the PRD)

GO gate

Status is PROPOSED. No build until Kyle accepts the PRD and records D1, D2, D3, D5, D6. Open PRs #68, #69, #70 are untouched; D3 is sequenced after #68/#69 land or close. ADR 0002/0003/0004 are unchanged; the PRD flags that D7 (/endpoints allowlist) would require a new ADR.