# The $100 AI challenge — protocol

Commission: Mike Hoydich accepted five assignments and authorized publication in chat on September 29, 2026. Codex performs the work and checks its own output. This is one instrumented session, not a blinded comparison of vendors or independent human evaluation.

Budget target: $100 API-equivalent text-model usage. Existing Codex access is used; no new plan purchased. Actual subscription allocation, tool costs, hosting, machine electricity and human time are unknown or excluded. No claim of an actual $100 invoice or all-in cost under $100 is permitted. Show observed token counters, cached vs uncached input, output, price source and cutoff. Reasoning output is already included in output and must not be billed twice. Tokens before the commissioning message are excluded. Existing conversation context sent during this run is included. The accounting cutoff precedes publication, so deploy and subsequent chat overhead are excluded.

## Fixed assignments and acceptance criteria
1. Brand: create a fictional “One Good Hour” offline activity-card concept, a responsive landing page and three campaign messages. Pass if inspectable, coherent, clearly fictional, no fabricated testimonials or live checkout. This is a launch kit, not a launched business.
2. App: make a bill splitter for 2–8 people using amount, tax percentage and tip percentage on the pretax amount. Pass if shares sum exactly to total, rounded cents are allocated deterministically, invalid inputs are rejected and mobile UI works. No backend, accounts or payments.
3. Document: generate a labeled synthetic eight-clause room-rental memo and summarize it with clause references, a reconciled cost scenario and unresolved contradictions. Pass if every factual finding maps to the source and uncertain terms remain uncertain. This is reading comprehension, not legal advice or legal benchmarking.
4. Weekend: plan a low-pressure Los Angeles weekend for two adults under a separate $100 outing allowance. Use official current museum information, explicit assumptions and unbooked reservations. Pass if arithmetic fits and source-dependent requirements are linked; no claim of attendance or available tickets.
5. Short: create a finished 20-second silent kinetic-type short for the fictional brand. Pass if playable, duration and dimensions verified, downloadable, and accurately described as AI-authored code rendered into video rather than a text-to-video model test.

No success ratings, saved-hours claims or invented human fixes. Publish observed failures and agent revisions. Human editing time was not measured. Source files and artifacts will be available for inspection.
