TFD-0023: Maya as Load-Bearing Infrastructure Talent — Sequencing of WO-0008 / RD-0031 / RD-0033

TFD-0023: Maya as Load-Bearing Infrastructure Talent — Sequencing of WO-0008 / RD-0031 / RD-0033

Status: Accepted Date: 2026-05-26 Decision maker: CEO (Oscar) · Approved by the board (Brendan) Specialists consulted: Riley (R&D), Pablo (Production Line Architect), Dana (Delivery), Camille (CS), Kai (Knowledge Manager), Clara (CTO)

Context

Three R&D items converged in the week of 2026-05-19→2026-05-26:

  • RD-0031 — toolkit-catalog foundry bundle. Go, TFD-0021 (2026-05-14).
  • RD-0033 — Talent AI-Layer starter kit (harness-as-product). Go (2026-05-26).
  • RD-0017 — Karpathy LLM Wiki video (2026-05-26, score 15/15 🔴). Confirms the technical pattern for Maya: raw/ + wiki/ + CLAUDE.md schema + /lint-wiki.

WO-0008 (Maya — Conversational KB Talent) has been pending since 2026-05-12. The board confirmed that 031 and 033 both depend on Maya for the foundry narrative ("every digital talent ships with its own wiki"). Without Maya, the harness bundle has no knowledge layer and Confluence pain persists for both factory (Camille's triage) and clients (STM portal deflection).

Aligned with maya-rag-wiki-pattern (direction confirmed), delivery-model-foundry-not-hosted, production-line-strategy (STM handover = #1 revenue blocker).

Decision

Greenlight WO-0008 now. Maya is upstream of RD-0031 and RD-0033. Three previously parallel initiatives are reframed as one coherent product line:

  • WO-0008 (Maya) = the knowledge layer (wiki + RAG + deflection).
  • RD-0031 (toolkit-catalog) = the distribution mechanism; absorbs /lint-wiki as a deliverable skill.
  • RD-0033 (harness-as-product) = the assembly: bundle = 031 + 008 + AI-layer residual. Scope is reframed to "compose existing pieces", not "build new layer".

Actions

  1. WO-0008 moved to active. Assignee: Pablo (Production Line Architect) for line discipline; Riley as RD-0017 reference owner; Camille as first internal user.
  2. 4-week sequence:
    • Week 1 (now → 2026-06-02): Sandbox proof on factory TFDs per RD-0017 exercise (raw/wiki/CLAUDE.md on TFD-0019/021/022). Stage 1 design freeze adopts RD-0017 schema as canonical.
    • Week 2-3 (2026-06-03 → 2026-06-16): Stages 2-4 build of Maya-Factory (M1 — replaces Confluence for Camille's client triage by end of week 3).
    • Week 4 (2026-06-17 → 2026-06-23): Maya-STM POC (M2), bundled with the STM EA talent for handover #1.
  3. /lint-wiki rolled into RD-0031 scope (toolkit-catalog bundle). Kai owns lint semantics; Diego owns packaging into the bundle.
  4. RD-0033 reframed. Clara to update the request to describe the bundle composition (031 + 008 + residual AI-layer items: stop-hook auto-improving, MCP codebase-search, scoped skills). A follow-up TFD will record the residual scope decision after Maya M1 ships.
  5. Confluence deprecation. Once Maya-Factory M1 is live, factory stops adding to Confluence. Existing pages migrate into wiki/ opportunistically (no big-bang migration).

Risk

Low. Pattern is proven (Karpathy, RD-0017). Order is fully specified (8 capabilities, AC + DoD already written per document-before-shipping-standards). Stack is factory-native (markdown + Claude Code + Astro widget). No new infrastructure. The 3-5 day net slippage on STM handover is bought back by the differentiator value of bundled Maya at delivery.

References

  • WO-0008 — order brief (2026-05-12)
  • RD-0017 — Karpathy LLM Wiki video analysis (2026-05-26)
  • RD-0031, RD-0033, CON-0006
  • TFD-0009 (request folder standard), TFD-0012 (R&D pipeline), TFD-0021 (toolkit-catalog Go)
  • Memory: maya-rag-wiki-pattern, production-line-strategy, delivery-model-foundry-not-hosted, document-before-shipping-standards