Free launch diagnostic

Would this AI launch survive one skeptical engineer?

Seven checks. Two minutes. A clear read on whether the announcement contains enough evidence to support a real decision.

Score the launch

No login. Your answers stay in this browser.

Evidence signal UNTESTED

0/7

Run the checks to expose the gap between the claim and the proof.

The seven evidence gates

Check only what the launch actually proves.

Do not award points for reputation, vibes, or a roadmap promise.

A1 Access is real A working product, API, repository, or documented access path exists today. C2 The core claim is falsifiable The launch names a number, condition, or outcome that another team can test. B3 The baseline is named Comparisons use the same task, tools, budget, and evaluation conditions. K4 The operating cost is visible Price, latency, compute, token use, or another real constraint is disclosed. F5 Failure cases are documented The announcement states where the system breaks, degrades, or stays untested. R6 The result can be reproduced Methods, prompts, data, code, logs, or test artifacts are available. O7 The evidence has an owner and date A named team, version, and evaluation date make the proof auditable.

ResearchAudio evidence lab

The launch claim is only one part of the decision.

Continue with the instrument that matches the next risk: token budget, cache economics, operating cost, agent reliability, or Codex project configuration.

Instrument 02 AI cost per successful task Replace token-price theater with retry-adjusted operating cost. Open calculator → Instrument 03 Agent loop diagnostic Find the missing guardrail before another retry becomes another incident. Inspect the loop → Instrument 05 LLM API cost calculator Turn traffic, token length, caching, and retries into a monthly budget. Build a token budget → Instrument 06 Prompt caching cost calculator Price cache reads and writes, then find the hit rate required to save. Find the break-even rate → Instrument 07 Codex CLI config.toml generator Build project-doc fallback settings without overriding AGENTS.md. Generate the config →

How to read the score

A low score is not a verdict. It is a list of unanswered questions.

0–2

Marketing claim

Interesting enough to watch, not strong enough to plan around.

3–4

Promising, under-proven

Run a bounded test before committing budget, users, or architecture.

5–6

Decision-ready

Enough evidence exists for a real evaluation with known open questions.

Evidence-complete

The launch gives an independent team a fair chance to reproduce the result.