Entire

You are working in the Trueup repository (this folder). Foreground only. Do not commit, push or post. Never print or write a key. No private paths. Plain words only (no: span, trace as a noun for people, lead, scenario, harness, adapter, instrumentation, OTel). Finish within 12 minutes. Edit README.md only.

Goal: make README.md lead with the links a judge needs to open, then high-quality, copy-paste code. Keep it under about 450 words. Keep the existing sections that still matter, but restructure:

# Trueup Witness + the one-line description (keep).
For judges, start here — a short list of links, each with a few plain words saying what it proves:
- Live demo (break a real AI bot yourself): https://trueup.38-49-216-59.sslip.io/demo
- Product site: https://trueup.38-49-216-59.sslip.io
- neatlogs records, before and after the fix: https://app.neatlogs.com/traces/5322259c536eae9651b3f0819bfade85 and https://app.neatlogs.com/traces/7304699be40a6194bafc2a45478bcd4f (the page shows the same records step by step, for people without project access)
- How neatlogs caught and fixed Trueup's own mistakes: docs/built-with.md#neatlogs
- neatlogs' daily email said 6.2%, the truth was 11.1%: docs/neatlogs-email.md
- Fixes sent upstream: Arize https://github.com/Arize-ai/openinference/pull/4024 (reviewed by Arize's head of open source), neatlogs https://github.com/neatlogs/neatlogs/pull/207, Traceloop https://github.com/traceloop/openllmetry/pull/4610
- Entire: the fixer session https://entire.io/gh/Jennycruzy/trueup/commit/79210a45418246ceb95e2990d1c003b5fbc7b3d9 (checkpoint 01M4M8M1GSQ2BM1RX2JKDJHPD6), full build history https://entire.io/gh/Jennycruzy/trueup, how each fix was made: fixes/openinference-run_status-openai-agents-top-level-span-status-ok/PROVENANCE.md
- The mistake we made with Entire, and what we did: one line ("20 updates were made in sessions Entire did not record; we said so, faked nothing, and moved all work back in") linking hackathon.md section "What Entire recorded, and what it did not" (use the right anchor).
- Business plan in cfo.ai: https://cfo.ai/s/kVKIburWzn09Baoy (and docs/business-plan.md)
- Fixes gallery: https://trueup.38-49-216-59.sslip.io/fixes · Build log: hackathon.md
The result table (2 → 0, 40 normal calls, 0 false alarms) — keep, short.
Install and Quick start: keep the existing working code block; check it against src/trueup/witness/core.py and make sure it is exactly right and copy-pastes cleanly. Add a second short block showing entire trueup --help (Trueup runs as an Entire command) and trueup audit examples/agent.py --library neatlogs (check your own script against the practice AI server). Verify both commands run (uv run trueup audit --help is enough for the second if a full run is slow).
How it works (4 lines), Docs list, License — keep, short. Run uv run pytest -q (some tests read README). Fix only README wording if a test needs it; never weaken a number check. Print the final README word count and the test result.

I rewrote README.md. It's now 405 words, and uv run pytest -q: 447 passed, 19 deselected. I changed no other file and didn't commit or push.

New order: the title and one-line description, then "For judges, start here" with every link you listed, each with a few words on what it proves. After that come the result table (2 → 0, 40 normal calls, 0 false alarms), Install, Quick start, the command-line block, How it works in 4 lines, Docs and License. The link to our mistake with Entire goes to hackathon.md#what-entire-recorded-and-what-it-did-not-12-oct-observed. That is how GitHub turns that heading into an anchor; I didn't click it on GitHub itself.

Quick start: I checked it against src/trueup/witness/core.py. watch() and run(name) match the code, run.disagreements is filled in when the with block ends, and TRUEUP_WITNESS=0 turns it off. To make it copy-paste cleanly I added the import os and neatlogs.init(...) lines, taken from docs/getting-started.md. I checked those against neatlogs' real init. One limit: agent.handle(...) is still a stand-in for the reader's own agent, marked # your own agent call. The block won't run as pasted until they put their own call there.

Command-line block:

trueup audit --help runs correctly. I didn't do a full audit run.
entire trueup --help only works when Trueup's entire-trueup command is on the PATH. Inside the project (uv run entire trueup --help) it shows Trueup's help. From a plain shell here, it showed Entire's own help instead, because Trueup's environment wasn't active. After pip install ./trueup into the active Python it should work; I didn't test that case. If judges might run it from a fresh clone, the README could say uv run entire trueup --help instead.

Removed to stay short: the "Why" line, the "Really: Crashed" column of the table, the old "Fixes sent upstream" paragraph and the "Built with" section. Their links are all in the judges' list now. I also took the email and build-log links out of the Docs line, since the list above already has them.