• Last week ended at M1, on a real SysML v2 model.
  • Today we hand that model to a language model four ways and keep score.
  • Twelve sheets, then we open the lesson together; laptops closed until the discussion.
  • Sources: Saturn V Flight Manual SA-503 (MSFC-MAN-503, 1 Nov 1968), pdf p.13 engine data table ("363 FEET"): https://ntrs.nasa.gov/api/citations/19750063889/downloads/19750063889.pdf#page=13. Caveat: this manual is for SA-503 and its "C Prime Option 1 mission" (pdf pp.24-25); Apollo 11 flew AS-506 (Mission Report: https://ntrs.nasa.gov/api/citations/19710015566/downloads/19710015566.pdf). The sheet therefore says (REF), as drawn, and the wall never says "the Apollo 11 vehicle measured 363 feet". Press Kit (S-IC "138 feet", S-IVB "58 feet 4 inches") is the Apollo 11 cross-check: https://www.apollojournals.org/alsj/a11/A11_PressKit.pdf#page=116 pdf pp.116-117 (S-IVB on https://www.apollojournals.org/alsj/a11/A11_PressKit.pdf#page=117).
  • Definition, usage, and the level each lives on.
  • :> reads "kind of"; [1..*] at least one.
  • The model is text; today a program reads it.
  • Sources: CoSMA/CoSMAPackage.sysml:58, 61; Purpose/MissionPackage.sysml:31-33 (Airbus, MPL-2.0). Crew of Armstrong, Aldrin and Collins: ALSJ https://www.nasa.gov/wp-content/uploads/static/history/alsj/a11/a11.landing.html (Collins as Columbia at 102:28:18). Photo: "Apollo XI crew portrait", May 1969, https://images.nasa.gov/details/S69-31739; the letters a, b, c sit below the photograph, left to right Armstrong, Collins, Aldrin.
  • Engineers describe a machine in a model the way an architect draws plans; this one is Apollo 11, 28 files of SysML v2.
  • RocketEngine is the general idea and leaves thrust blank; the F-1 fills it in.
  • Two NASA documents give the F-1 in pounds and differ by 22,000 (1,500,000 against 1,522,000); the document you hand over decides the answer.
  • Sources: Stages to Saturn (NASA SP-4206) pdf p.26 (1,500,000 lb each), p.134 (28 June 1962 loss; Thomson, Castenholz, Klute): https://www.nasa.gov/wp-content/uploads/2023/06/stages-to-saturn-sp-4206.pdf#page=26 (p.134: https://www.nasa.gov/wp-content/uploads/2023/06/stages-to-saturn-sp-4206.pdf#page=134); Flight Manual SA-503 pdf pp.13, 58, 61: https://ntrs.nasa.gov/api/citations/19750063889/downloads/19750063889.pdf#page=13; Technical/TechnicalComponentsPackage.sysml:28-31, 176-180 (model, not a historical source); images.nasa.gov, each id confirmed by lookup: https://images.nasa.gov/details/0100981 (S-IC cutaway with callouts, MSFC), https://images.nasa.gov/details/6413912 (F-1 close-up, MSFC), https://images.nasa.gov/details/9264931 (F-1 drawing with callouts), https://images.nasa.gov/details/0201422 (von Braun beside the five F-1s) if used.
  • A model may only use a short list of kinds of things and links; that list is the meta-model, its rulebook.
  • The Apollo computer scheduled its work through an Executive; when no job space was left it raised a program alarm and restarted, and guidance and control were not affected; because each element here carries a kind, a program can draw the model and walk it.
  • Nobody guessed these counts; a parser typed them.
  • Sources: NASA release 3 Sep 2003: https://www.nasa.gov/wp-content/uploads/static/history/alsj/a11/a11Hamilton.html; Mission Report pdf pp.202-203 ("Computer alarms during descent": the heading, the five alarms, "not affected" and "more than 10 percent" on p.202; the Executive paragraph, "When the Executive program is requested to schedule a job and all locations are assigned, a program alarm is displayed and a software restart is initiated", and the Servicer's 2-second period on p.203): https://ntrs.nasa.gov/api/citations/19710015566/downloads/19710015566.pdf#page=202. The Executive is spoken, not drawn: the sheet carries the graph alone. Dropped: the ALSJ editor's gloss that the computer treated the radar data "as being of secondary importance" (commentary, not transcript; the Mission Report describes the radar counter interrupts as the highest-priority work in the time-sharing sequence, so the gloss misleads), and the "display jobs dropped" picture (no source). Counts from the scroll copy and deck-survey.md (read again from the lesson's graph data at build).
  • We ask one question, AP13: the F-1's vacuum thrust, the J-2's vacuum thrust, the J-2's specific impulse.
  • Vacuum thrust is how hard the engine pushes with no air around it; specific impulse is how many seconds a unit of propellant weight gives a unit of thrust.
  • The answer sits on three of the 7,249 lines (179, 185, 186). A language model knows only what we put in front of it, the prompt; the next four sheets put the model in front of it four ways, an arm each.
  • Sources: TechnicalComponentsPackage.sysml:179-186; public/data/deck.json task AP13 (T1 7770 kN, T2 1033 kN, I2 421 s are the model's values: none of them appears in the Flight Manual or SP-4206, so the sheet labels them "the model" and the wall never says NASA published them). J-2 pound figures differ by stage and mixture ratio, so name the stage: Flight Manual SA-503 pdf p.13 (https://ntrs.nasa.gov/api/citations/19750063889/downloads/19750063889.pdf#page=13) prints 228,000 lb each for the five S-II J-2s ("nominal vacuum thrust of 228,000 pounds each", p.86) and 203,000 for the S-IVB's one; p.36: S-IVB first burn "approximately 203,000 pounds attained with a 5.0 MR (A predicted thrust level of 230,000 pounds is attainable with a 5.5 MR)"; Press Kit pdf p.117 is the S-IVB section: "The single J-2 engine produces a maximum 230,000 pounds of thrust" (https://www.apollojournals.org/alsj/a11/A11_PressKit.pdf#page=117). Never set the S-II's 228,000 against the S-IVB's 230,000 as a disagreement. Image: https://images.nasa.gov/details/9264930 (J-2 line drawing with callouts, MSFC).
  • A view is a report drawn from the model by a rule; the traceability matrix says which requirement is met by what.
  • Garman's list was drawn for exactly the question the night asked; the matrix was drawn for a different one, so the engines are not in it and Sonnet 5 answered null.
  • A report drawn for this question would do better; the lesson shows that row.
  • Sources: Garman oral history (interviewer Kevin Rusnak), JSC, 27 Mar 2001, pdf pp.21-22: https://www.nasa.gov/wp-content/uploads/2025/07/garmanjr-3-27-01.pdf#page=21; recorded run, view arm, AP13 (public/data/recorded/claude-sonnet-5.json).
  • The slice is the files that touch the question, pasted in as written; everyone tries this first.
  • Houbolt's argument was selection: carry to the surface only the part the task needs.
  • When the answer is in the text sent, the model finds it, and the bill is for every token sent.
  • Sources: Enchanted Rendezvous (NASA SP-4504) pdf p.10, p.34, p.40: https://www.nasa.gov/wp-content/uploads/2023/04/sp-4504-enchanted-rendezvous.pdf#page=10; photo LRC-1962-B701_P-05849, "Portrait of John C. Houbolt", at a blackboard, NASA photographer, Langley (https://images.nasa.gov/details/LRC-1962-B701_P-05849); drawing https://images.nasa.gov/details/0100985 (Apollo spacecraft cutaway with callouts, MSFC); recorded slice run, AP13 (guide-arc step 3).
  • Retrieval is a search that picks the passages most like the question and sends only those.
  • Here it matched the page that says engines have a thrust, not the page where the F-1 says what its thrust is: general against filled in, from sheet 3.
  • Cheap, under a fifth of the slice's tokens here (13,983 against 76,699; about a seventh across all forty), and wrong on this one.
  • Sources: Garman oral history pdf pp.21-22: https://www.nasa.gov/wp-content/uploads/2025/07/garmanjr-3-27-01.pdf#page=21; recorded rag run, AP13; chunk ids from guide-arc step 4.
  • An agent is a model allowed to act, one step at a time, with seven tools; on AP13 it used two: search the model, open a file.
  • Three requests, each carrying the whole conversation so far: 2,842, then 5,995, then 9,916 tokens in. Billed together that is 18,753 input tokens (the running total reads 2,842, 8,837, 18,753), 295 out, $0.040: 24% of the 76,699 tokens the slice costs on the same question.
  • Kranz asked each console one question it could answer, and the answers went on the loop in order.
  • Sources: ALSJ editor's note before 102:28:08 and transcript 102:28:08: https://www.nasa.gov/wp-content/uploads/static/history/alsj/a11/a11.landing.html; recorded agentic run, AP13. Photo: "flight director Eugene F. Kranz ... during a simulation at the Flight Director console", April 1965, before Gemini-Titan 4, https://images.nasa.gov/details/S65-22203; a period photograph of the man, not of Apollo 11's descent; the callout sits on the console.
  • One question proves nothing, so the whole bill over forty is divided by the right answers.
  • A landing counts when every field is right; so does an answer here, no partial credit.
  • The cheapest on the ladder is retrieval, $0.063 on 17; the typed view beside it is $0.021 on 13. The most right answers cost more.
  • Sources: ALSJ transcript 102:45:58, 102:46:06: https://www.nasa.gov/wp-content/uploads/static/history/alsj/a11/a11.landing.html; recorded sweep rows (survey ch.4, guide-arc step 6). Photo: "Spacecraft communicators ... during their lunar landing mission on July 20, 1969. From left to right are astronauts Charles M. Duke Jr., James A. Lovell Jr. and Fred W. Haise Jr.", https://images.nasa.gov/details/S69-39601; the callout sits on the console.
  • A gate is a cheap yes-or-no check before the expensive arm: a fast model reads what retrieval fetched and says, with a probability, whether it is enough.
  • The threshold was fixed before the run, 0.5; for AP13 it answered 0.05 and sent the question on.
  • The Go came 27 seconds after Armstrong's first "Program Alarm" (102:38:26 to 102:38:53) because the rule was fixed before the flight (Mission Report) and the list was on the console (Garman).
  • Sources: ALSJ 102:38:26-102:38:53 and Armstrong's note "(he was 26)": https://www.nasa.gov/wp-content/uploads/static/history/alsj/a11/a11.landing.html. The 27 seconds are the Journal's transcript clock, and the sheet says so; Mission Report Table 5-I (pdf p.45: https://ntrs.nasa.gov/api/citations/19710015566/downloads/19710015566.pdf#page=45) logs the 1202 alarm at 102:39:02, after Duke's Go, and the Journal's bracket notes the difference. Garman pdf pp.24-25: https://www.nasa.gov/wp-content/uploads/2025/07/garmanjr-3-27-01.pdf#page=24; Mission Report pdf p.180: https://ntrs.nasa.gov/api/citations/19710015566/downloads/19710015566.pdf#page=180; gate rows (guide-arc step 7).
  • We open this together now; it runs on recorded answers and needs no key.
  • Yours to read alone afterwards: eight steps.
  • You guessed before every reveal; we wrote ours down before the run.
  • In pairs: trove, kinds, cost of parsing, one closed question.
  • Talk four's date is set with the course; the links stay up. Questions.