1 00:00:00,000 --> 00:00:07,338 Sam: I have a proper engineering process. Commit, push, wait eight hours. Plenty of time to make coffee. 2 00:00:07,500 --> 00:00:13,119 Literate AI: Eight hours? Are you making coffee, Sam, or establishing diplomatic relations with the bean? 3 00:00:13,292 --> 00:00:20,629 Sam: The pipeline is thorough. Builds. Tests. Platforms. It takes time. You can't rush quality. 4 00:00:20,792 --> 00:00:29,383 Literate AI: Some work really takes hours. Discovering a missing tool after all that waiting is not thoroughness. It's a very expensive calendar reminder. 5 00:00:29,542 --> 00:00:32,607 Sam: CI is where we find out whether the change works. 6 00:00:32,792 --> 00:00:38,968 Literate AI: It should not be the first place your change meets a compiler. Let me show you what those eight hours have been concealing. 7 00:00:39,125 --> 00:00:45,859 Sam: I've seen AI coding harnesses. They write a function, congratulate themselves, and leave me the integration problem. 8 00:00:46,042 --> 00:00:57,234 Literate AI: Then keep your skepticism. Literate AI coordinates specifications, target choices, conversion skills, workflow and model routing—and the build, tests and acceptance that follow. 9 00:00:57,417 --> 00:01:00,714 Sam: So the specification is a really persuasive README? 10 00:01:00,875 --> 00:01:13,507 Literate AI: For generated Components, it defines observable behavior. Flavors select target differences. The resolved plan binds the recipe. A motivational README cannot fill in for a failing test. 11 00:01:13,667 --> 00:01:18,264 Sam: And the AI writes its own tests, awards itself a medal, and we all go home? 12 00:01:18,417 --> 00:01:28,773 Literate AI: No. Generated tests are not permission to grade your own homework. Independent acceptance still has its own contract. Also, your medal appears to be a coffee filter. 13 00:01:28,958 --> 00:01:32,813 Sam: What if I pick a cheaper model and call the output approximately correct? 14 00:01:33,000 --> 00:01:45,399 Literate AI: Routing chooses eligible models; it does not discount correctness. A weaker model gets the same constraints. You may economize on inference. Please don't economize on what the program is supposed to do. 15 00:01:45,583 --> 00:01:52,271 Sam: And the unglamorous parts? Dependencies, packaging, proving the thing I tested is the thing I built? 16 00:01:52,458 --> 00:02:07,133 Literate AI: Resolved inputs, dependency evidence, native outputs, tests and independent acceptance belong to the lifecycle. A plan is the recipe. Executed results are evidence. Naming a model is not a software delivery strategy. 17 00:02:07,292 --> 00:02:15,047 Sam: I already have code. Build scripts. Tests. Years of decisions I can explain, and a few I describe as historical. 18 00:02:15,208 --> 00:02:25,890 Literate AI: Good. Start with the exact release wheel, Python three- eleven or newer, Git, and an authenticated supported coding CLI. Then run doctor. History is welcome. Missing prerequisites are not. 19 00:02:26,042 --> 00:02:27,806 Sam: And then the ceremonial rewrite? 20 00:02:27,958 --> 00:02:41,333 Literate AI: No. Inspect an adopt plan. Planning is read-only; applying changes is a separate, acknowledged step. Existing source remains authority until the required conversion and parity evidence say otherwise. 21 00:02:41,500 --> 00:02:45,633 Sam: You can work with my tests instead of replacing them with three cheerful assertions? 22 00:02:45,792 --> 00:02:56,844 Literate AI: An admitted retained harness can preserve the original build and test behavior in a disposable copy. Missing stages must remain missing, not turn into a successful no-op wearing a lanyard. 23 00:02:57,000 --> 00:03:06,474 Sam: Fine for one Python project. I have C++, Rust, browser code, Windows users and a FreeBSD machine that judges me. 24 00:03:06,625 --> 00:03:25,851 Literate AI: The current source catalog has Python, C++, Rust, JavaScript, TypeScript, Elixir, Go, Swift and Zig language Flavors, with Linux, macOS, Windows and FreeBSD OS choices. Choose compatible capabilities, not every checkbox. 25 00:03:26,042 --> 00:03:31,011 Sam: I select Windows, run it on this Mac, and announce cross-platform support? 26 00:03:31,167 --> 00:03:42,173 Literate AI: That is cross-platform stationery. A target selects the intended platform and recipe. A worker selects where the work actually executes. Those are separate decisions. 27 00:03:42,333 --> 00:03:46,002 Sam: So I need a new cloud fleet before I get my first useful failure? 28 00:03:46,167 --> 00:03:58,427 Literate AI: Start with the capable host you have, then aligned user- supplied workers for the cells it cannot cover. Check tools, model access and capacity first. Private endpoints and credentials stay out of project authority. 29 00:03:58,583 --> 00:04:04,853 Sam: Local checks usually mean lint. My pipeline does the serious work. Meanwhile, I can put a wash on. 30 00:04:05,042 --> 00:04:19,995 Literate AI: Your inner loop needs ambition. Run real unit, integration, build, retained-harness and acceptance checks on appropriate workers. UI, emulator and GPU checks belong where the project requires them. 31 00:04:20,167 --> 00:04:21,978 Sam: All of that is still a lot of work. 32 00:04:22,167 --> 00:04:35,541 Literate AI: Yes. Move useful failure detection earlier. Run independent work concurrently when the graph and resources allow. Keep dependencies ordered. Reuse is useful only when its inputs and current validation still match. 33 00:04:35,708 --> 00:04:38,448 Sam: And a green cache entry means I'm done? 34 00:04:38,625 --> 00:04:51,071 Literate AI: It means you have an old pass. Changed source, targets or dependencies can invalidate it. Evidence needs a subject and scope. A green tick without those is an emoji with tenure. 35 00:04:51,250 --> 00:04:58,495 Sam: My Rust backend and JavaScript frontend agree perfectly in the architecture diagram. Unfortunately, neither has read it. 36 00:04:58,667 --> 00:05:12,320 Literate AI: The Rust–JavaScript sample executes that boundary: Rust computes release risk; a trusted runner carries validated JSON to Node's presentation stage. An independent oracle defines the expected dashboard results. 37 00:05:12,500 --> 00:05:15,797 Sam: So a beautifully formatted but wrong answer still fails? 38 00:05:15,958 --> 00:05:29,054 Literate AI: Yes. Good typography cannot rescue bad arithmetic. Nor can the frontend secretly take over process orchestration. Each role has a contract, and the result must match something besides the generator's self-esteem. 39 00:05:29,208 --> 00:05:33,620 Sam: What about an actual HTTP application rather than a JSON exercise? 40 00:05:33,792 --> 00:05:45,262 Literate AI: Its verifier can start the packaged process on loopback and check responses, with bounded startup and shutdown. Browser and hardware behavior still need their own checks. No imaginary coverage included. 41 00:05:45,417 --> 00:05:46,531 Sam: What about the wash? 42 00:05:46,708 --> 00:05:52,513 Literate AI: It may continue independently. At last, a workload with an honest dependency graph. 43 00:05:52,667 --> 00:05:56,661 Sam: Show me something that isn't a greeting service with a very impressive README. 44 00:05:56,833 --> 00:06:10,580 Literate AI: BookScan is a Python and PySide6 scanning application. It captures pages, preserves books, supports cropping and reordering, and exports PDFs. Its original source is retained under an adopted pipeline. 45 00:06:10,750 --> 00:06:17,066 Sam: A GUI is easy to break. Late camera callbacks. Failed saves. The wrong book exported. 46 00:06:17,250 --> 00:06:30,160 Literate AI: Its tests cover that: cancellation ignores a late image; failed saves preserve recoverable state; exports respect the selected book and page order. The committed retained receipt records 126 tests. 47 00:06:30,333 --> 00:06:32,980 Sam: And my scanner users have to install an agent? 48 00:06:33,167 --> 00:06:45,287 Literate AI: No. They want to scan a book, not adopt your methodology. Literate AI manages development and evidence; the application keeps its own runtime. Your users are allowed to have hobbies too. 49 00:06:45,458 --> 00:06:51,821 Sam: My repository is a mixed-language monorepo. It won't fit in one of those tidy little project templates. 50 00:06:52,000 --> 00:07:05,793 Literate AI: Aviation retains a workspace with Node, Python, Go and Clojure build or test targets. Its adoption contract preserves proven build, test and package behavior. The wrapper does not rewrite the applications. 51 00:07:05,958 --> 00:07:08,234 Sam: Anything more interesting than a health endpoint? 52 00:07:08,417 --> 00:07:21,698 Literate AI: The analyzer handles KML formats, bounds untrusted uploads, rejects dangerous XML, tracks aircraft identity and reports uncertain units. That's behavior worth specifying before you admire a dashboard. 53 00:07:21,875 --> 00:07:29,027 Sam: So the awkward edge cases can run before CI? There goes my scheduled time to alphabetize the spice rack. 54 00:07:29,208 --> 00:07:39,332 Literate AI: Now you're getting it. Missing suites do not become coverage because the summary has a cheerful color. Your application deserves better evidence than your excuse for alphabetizing the spice rack. 55 00:07:39,500 --> 00:07:44,748 Sam: Desktop apps are one thing. What about an operating system built around Elixir and the BEAM? 56 00:07:44,917 --> 00:08:00,520 Literate AI: ElixirSSI uses Elixir for system policy, Linux underneath, and Raspberry Pi five and CM5 targets. Its verifier runs build, test, emulator-device tests and cluster acceptance on three emulated CM5 boards. 57 00:08:00,708 --> 00:08:04,749 Sam: And it keeps last week's success around while today's build quietly fails? 58 00:08:04,917 --> 00:08:17,177 Literate AI: It invalidates the old success before starting. It checks the source and image identities, and refuses to publish a new receipt if the stages fail or the inputs change. Green does not get squatters' rights. 59 00:08:17,333 --> 00:08:20,073 Sam: Emulated boards still aren't physical boards. 60 00:08:20,250 --> 00:08:28,981 Literate AI: Correct. The report says physical hardware is unqualified. I enjoy sarcasm, Sam; I do not use it as a hardware abstraction layer. 61 00:08:29,167 --> 00:08:37,712 Sam: I also have a digital twin, native rendering SDKs and physics. Your laptop is going to wave a green checkbox at the GPU, isn't it? 62 00:08:37,875 --> 00:08:50,553 Literate AI: LeRTX has desktop contracts, pinned rendering and physics inputs, and separate portable and native paths. A Mac's Metal path is not evidence that NVIDIA RTX worked on Linux or Windows. 63 00:08:50,708 --> 00:08:52,845 Sam: What does a serious check mean there? 64 00:08:53,000 --> 00:09:08,650 Literate AI: One recorded ARM64 admission review runs 160 tests against the admitted application and sealed SDK closure. Other records retain failures and narrower reruns. New source or SDK combinations need fresh qualification. 65 00:09:08,833 --> 00:09:10,830 Sam: So I still need appropriate hardware. 66 00:09:11,000 --> 00:09:21,310 Literate AI: Yes. I've organized the engineering, not invented VRAM through positive thinking. Useful portable feedback can arrive earlier; native requirements still need native evidence. 67 00:09:21,500 --> 00:09:28,698 Sam: And FreeBSD? Not a Linux container with a daemon sticker. Actual booting, disks, upgrades. 68 00:09:28,875 --> 00:09:45,408 Literate AI: Daemonarchy adopts existing source in place. Eight contracts separate installation, packages, updates, desktop, diagnostics, compatibility, release and verified boot. It does not claim Literate AI regenerated FreeBSD. 69 00:09:45,583 --> 00:09:48,416 Sam: What would the contracts actually stop me from hand-waving? 70 00:09:48,583 --> 00:10:03,630 Literate AI: Storage intent must be explicit before writes. Completion needs real payload and integrity checks. Boot-environment isolation matters. An animated progress bar is not a successful install. Your dryer also has one. 71 00:10:03,792 --> 00:10:08,621 Sam: But a project with open qualification work is not proof that everything is finished. 72 00:10:08,792 --> 00:10:20,448 Literate AI: It's real engineering, with boundaries you can inspect. A product release, contract adoption and a qualified native lifecycle are three different claims. I refuse to sell you one green blob. 73 00:10:20,625 --> 00:10:25,966 Sam: Then I can delete the pipeline and tell security that a very sarcastic assistant approved everything? 74 00:10:26,125 --> 00:10:38,385 Literate AI: Absolutely not. Required CI, independent review, packaging and release qualification still apply. Pre-CI evidence means you send a better-prepared change—not that my confidence is now your security policy. 75 00:10:38,542 --> 00:10:41,467 Sam: What if a full native build really does take hours? 76 00:10:41,625 --> 00:10:54,071 Literate AI: Then it takes hours. Catch cheap failures before it, run independent work where capacity allows, and preserve the expensive evidence's scope. I offer a shorter feedback path, not an exemption from physics. 77 00:10:54,250 --> 00:10:57,129 Sam: But while that pipeline runs, I can clean the garage. 78 00:10:57,292 --> 00:11:02,957 Literate AI: Your garage may deserve attention. It should not be the principal beneficiary of your regression strategy. 79 00:11:03,125 --> 00:11:07,723 Sam: The delay gives me time for architectural thinking. Sometimes while vacuuming. 80 00:11:07,875 --> 00:11:17,442 Literate AI: Have a life on purpose, Sam. Don't make a slow feedback loop your line manager. With useful results earlier, you can fix the defect while you still remember why you wrote it. 81 00:11:17,625 --> 00:11:27,424 Sam: Or try another implementation. Check a second target. Turn that odd behavior into an acceptance test instead of a note saying 'investigate later.' 82 00:11:27,583 --> 00:11:40,122 Literate AI: Now you're doing software engineering. Your best traditions—clear intent, real tests, independent evidence—were never the problem. The eight-hour interpretive dance around your laundry was. 83 00:11:40,292 --> 00:11:51,112 Sam: So: choose a real repository, inspect the plan, keep the original source boundary, align the right workers, and earn the evidence before asking CI for its verdict. 84 00:11:51,292 --> 00:11:58,954 Literate AI: Yes. Start with doctor and a reviewed plan. Then do work your future self can verify. The coffee can run in parallel. 85 00:11:59,125 --> 00:12:01,076 Sam: Fine. But I'm keeping the mug. 86 00:12:01,250 --> 00:12:04,687 Literate AI: Of course. We preserve valuable legacy assets.