docs(dfw): tokenomics session — talking track and slide deck

[bryantharpe]

A 10-minute, discussion-led session for DFW Innovation Day, anchored on this repository as the worked example.

The core message is that AI economics are driven by architecture and workload routing, not token pricing — so the evidence is the repo itself rather than a price-per-million table.

What's here

docs/dfw-innovation-day/

File	What it is
tokenomics-deck.html	12 slides, single self-contained file. Images inlined as data URIs — no build, no network, opens offline.
talking-track.md	The script with timing marks, Q&A backup (including likely pushback and honest answers), and a Monday rehearsal checklist.
README.md	How to run it, and what remains to be filled in.

The deck carries its own presenter tooling: per-slide speaker notes (S), a 10-minute talk clock that warns at 8:30 and goes red past 10:00 (T), an overview grid (G), keyboard navigation, and a print stylesheet so a PDF backup is one Cmd-P away.

The argument, slide by slide

1. The product is real — the v1 web UI localizing a star field in 1.8 ms, 47 catalog stars matched, benchmarked against an independent Python reference.
2. No human wrote the implementation — built by the agent fleet against written specs, humans in the loop at exactly four points.
3. It was built twice — v1-original vs. the spec-only rebuild on main, which is a natural A/B on cost: 33 days → 10, 144 commits → 54, 14,660 lines → 9,929, and v1 had zero CI where the rebuild can't merge without ten required checks.
4. Three decisions moved the bill — route work by size (small → local model, large → cloud relay), keep author and reviewer on different model lineages, and treat the frontier tier as a budget rather than a default.
5. Twelve deterministic gates cost zero tokens — and per the run-7 audit, ten of the twelve were hygiene while three carried the actual weight.

Appendices cover the two expensive failure modes (ps-exp-01's 1,410 commands for zero code; ps-exp-02's byte-for-byte recovery that every gate approved), the provenance chain, and a source table listing every claim against the command or file it came from.

Two deliberate gaps, marked in place rather than guessed

• Token/spend figures render as dashed PENDING tiles. The Grafana LLM consumption data wasn't reachable from this environment. On a slide about cost, a plausible invented number is worse than a visibly empty one — the placeholders are styled so they cannot be mistaken for real figures. Search the HTML for is-pending.
• The palette is brand-adjacent, not sampled. slalom.com is blocked by this session's egress policy (403 on CONNECT), so the colors are a best read rather than measured values. Every color derives from one token block at the top of the file, so correcting it is a single-block edit.

Notes for review

• Docs only — no code, no build files, nothing under crates/ or openspec/ touched.
• Every non-placeholder figure is reproducible from this repo at v1-original and main; slide 12 is the provenance table.
• The two-series chart palette (#EB6834 / #0C62FB) was validated for colorblind separation and contrast rather than picked by eye — worst-pair CVD ΔE 31.6, normal-vision ΔE 40.6, both above 3:1 on the light surface.
• Screenshots reuse the ones already committed under docs/screenshots/, cropped and re-encoded to WebP.

🤖 Generated with Claude Code

https://claude.ai/code/session_01QcgK3D8YTuJtRQLTiortmE

---

Generated by Claude Code

[bryantharpe]

review/judge — APPROVE on dd717dc8835e

VERDICT: APPROVE
BLOCKING: NONE
NITS: NONE

---

Independent U3 review by the judge role (glm-5.2, Zhipu — different lineage
than every author seat). Verdict is on blocking findings, not a score. To
supersede with human sign-off (U4): judge-pr.sh override 85 "<reason>".