July 20, 1969. Five minutes into the powered descent, a master alarm sounds and Eagle's computer display shows 1202. Some twenty seconds later, Mission Control calls "GO" — because the people on the loop knew precisely what that code meant, and precisely what the computer would do next. That is what it looks like to truly know your control system.
The Apollo Guidance Computer flew the lunar module on Luminary 099 — about as safety-critical as software gets, written in assembly for a machine with less memory than this page. During the descent it raised five program alarms: four 1202s and one 1201. Each time, it restarted itself, shed everything unimportant, and kept flying the ship.
The source code is public domain now. Which makes it the perfect demo system: everyone knows the story, nobody's NDA is at risk, and every claim below can be checked against the actual lines that flew.
The keypad works about as much as a museum piece should. RSET really does clear the lights.
Pick a question, press Ask Daritas, and the answer appears here — the way a fast-lane answer reads, with citations linking to the real source on GitHub.
Yes. Point any AI assistant at this code and ask about 1202 — you'll get a fluent answer, and much of the time a correct one. The explanation is not the product. What you can act on is, and that takes two things an answer alone doesn't have. Don't take our word for it — the method below is published so you can check the difference yourself, on this same public code.
Every claim in a Daritas model must carry a verbatim quote at the source location it cites — a provenance check run by machine: rejected if the quote isn't there, with the rejections disclosed. A chat answer asserts; a Daritas model makes its evidence — and its limits — inspectable.
In the expert lane, an automation engineer verifies the model against its evidence and signs it — a person, by name, accountable for what it says. Every deliverable is labelled machine-generated or engineer-reviewed, and the label is never blurred. Accountability is the part no copilot ships.
That is how Apollo actually did it. Weeks before the landing, a simulation threw a program alarm and the team called an abort they didn't need. Afterwards, engineer Jack Garman prepared a list of the program alarms and their operational significance — worked out in advance, on paper. When 1202 flashed for real, Houston wasn't consulting an oracle; the answer had already been prepared by an accountable human. That list is the product. The explanation is just where it starts.
Comparisons where a vendor picks the questions and grades its own homework are worth nothing. So here is the whole setup: public source, fixed questions, stated scoring. Run it against any assistant you like and compare.
Apollo 11 Luminary 099 — the lunar module's flight code, public domain and on GitHub. No private corpus, nothing you can't obtain.
The five on the DSKY above, fixed in advance: what 1202 is, what 1201 is, whether they are the only alarms of that kind, why they fired, and why “GO” was still right.
Five dimensions, each checkable by a third party: citation provenance, entailment, coverage, contradictions, and staleness.
| Dimension | What it asks | Daritas, on this package |
|---|---|---|
| Citation provenance | Does every citation quote the file and line it claims — verbatim? | 17 of 17 verified verbatim at their cited lines, by a script that runs in CI. |
| Entailment | Does the cited line actually support the claim made from it? | Every claim is tied to source; a quote that is not at its location is rejected before you see it. |
| Coverage | How much of the package was actually read, and what was not? | 629 functions, 3,792 grounded claims across 62 .agc files — coverage stated, not implied. |
| Contradictions | Does the answer contradict the source, or itself? | Mission history that is not in the code is labelled as context, never blended into what the source shows. |
| Staleness | If the source changes, does the answer know? | Claims are pinned to a baseline; a changed file marks affected claims stale until re-evidenced. |
We have not published scores for other assistants, because we have not run them under this method — and printing a rival's column we guessed at would be exactly the behaviour this page argues against. When we do run it, the model names, versions and dates will be stated and the raw outputs published alongside the scores. Until then the invitation stands: the source is public, the questions are above, and the scoring is written down. Ask your own assistant the same five and check its citations against the lines that flew.
The 1202 alarm is remembered as a scare. It should be remembered as a triumph of knowing your system: the computer exposed a bounded failure mode the team could read, the design bounded what could happen next, and the understanding existed before it was needed. Nobody was reading source code during the landing.
Most industrial control systems are the opposite: the understanding lives in someone's head, walks out at retirement, and gets reconstructed under pressure at three in the morning. Daritas builds the evidence-backed model before the master alarm — every claim tied to the line of code that proves it, reviewed and signed by a named engineer, kept current as the system changes.
One more difference: the AGC's source is public domain. Yours isn't — which is why Daritas runs inside your own network, and your code never has to leave the site.