Pinball, Pachinko, and the Art of Replayable AI Workflows

Sep 28, 2026

🤖 Read Raw Markdown • 📄 Google Doc (Try: Tools/Audio/Listen to document summary)

Setting the Stage: Context for the Curious Book Reader

In this interesting journal entry, the exploration moves deep into the daily cadence of human-AI collaboration. By contrasting the unpredictable nature of falling beads in a pachinko machine with the interactive, bounded control loops of pinball, this essay outlines how reproducible tools and checkable receipts keep human agency intact during complex software engineering tasks.

Technical Journal Entry Begins

MikeLev.in: Okay we have to remember to keep the robots.txt from that client’s site in context. That could be a step on the walk.

Wow I think that last article of mine has the details for what I have to do next. It is a pretty good freeze-frame and a cartridge if you will of recoverable context right down to the MCP and other tool-calls that went on in the background.

Which brings us to that very important point about the new MCP server that’s connected to the render farm and how we have to start using it pronto. The time is 5:30 AM. My meeting where I wanna talk about this already and the experience is 10 AM. I also want to have at least had one article on one of the Jira tickets, see? So we are effectively wiring up the pinball machine or if you preferred Pachinko machine because both metaphors are good. I think I’m finally gonna settle on pinball machine. I can play off of the pinball Wizard, Tommy.

Editing is about growing back in time. It’s time travel in the past to revise a previous version of something. It doesn’t work that way. Life is one way forward, or at least the way we generally experience it. We wake up in the morning. We go to bed at night. Repeat.

That’s oscillation. The Earth spins on an axis giving us day and night, light and dark, wakefulness during time when executive function can assert itself versus sleepfullness where I don’t think we know squat. I mean think about it. You effectively go unconscious for about a third of the day. And you have to give up control to get there to fade off to sleep. And at some point that control needs to be returned to you as you wake up. That’s a weird moment. Oh yeah, I’m a meat puppet and I’m coming out of a suspended animation phase where the body just sat there and the mind went off and did other things.

The Oscillation of Work and Rest

What? The body moving? Well, some people with the most part brain cuts off whatever that connection is between the thing that still thinks it’s walking around in your dreams and the body is attached too. It’s like a little magic pair of scissors came in and snipped that connection so your body remain there safe safely unmoving while your bringing your adventures going wherever doing whatever you can imagine. Is that it? Are we imagining?

All right one of the most difficult things I am discovering about working with AI to do the kind of work you did in the past only better is how to hold the reins and be the coachman of a horse-carriage where are the horses are AIs but also sort of RFP-style subcontractors. Some of them you can just trust to do certain tasks because you got that work from them many times before and never had a problem. That would be what you call models that no the so-called web full stack, namely node JS and react JS. If that’s all you’re doing and in ways that are not only identical to the way it’s been done 1000 times before and GitHub, but you’re learning how you want to do it based on the Commons. Let the wisdom of the crowd compel you. Let the most popular things in the normal distribution curve define what you do and consequently in life ways, you.

It’s not that the wisdom of the crowd is wrong. It’s often right. It’s not that the wisdom of the crowd is always the tragedy of the commons, though it sometimes is. Refer to Goodheart’s law for at least one example of why. Turn something into a metric and ruin that metric because everyone will optimize to it as the target. All the important stuff that gets layered up from deep down in the subtlely and nuance get erased because they’re not important. Flair and better handling of edge cases and valuing the outlier because it might have a solution required for those cases that is not immediately obvious because this is not optimized to surface unusual things for dealing with unusual circumstances. That’s the stuff that optimization chops off!

Wisdom of the Crowd Versus Edge Cases

Qantized away.

When things have those oscillating cycles like the previously mentioned night and day or going from one article to another working with AI, one of the things you can do for yourself is to make sure that there’s significant overlap with things you wanna capture on the next cycle you’re working with on the current cycle. Article to article too.

Now because I sort my articles into all these different bins now based on their purpose such as into the public website for reading here in YouTube or into a private company confluence wiki or not published anywhere at all in something I call the Grimoire.

When you’re working out new recipes like I imagine must’ve been the case when people were figuring out how to make bread for the first time there’s sometimes secrecy because there’s competition between tribes and it’s a matter of survival. Whole tribes got wiped out or survived and prospered I’m gonna guess based on whether they could keep themselves fed when hunting was poor. Turning grass into a meat-like substance is probably a pretty neat trick and useful for surviving. I’m gonna bet the folks working that out realized this and the way yeast works becomes very eye of newt mystical.

Know if it were just bread that would be one thing but it’s also booze because a similar process that turns grass into a meat-like substance turns sweet juice into silly-making juice. I’m gonna guess that there was something economically valuable and beneficial to your general survival if you figured out making bread and brewing beer.

Those tribes did well I bet. And those tribes tried to keep secrets if only so that they themselves stayed alive. And their children doing better is a big incentive to pass-forward the secret recipes of baking the bread and brewing the beer.

These are things one must keep in mind when figuring out how to future-proof your tech skills in the age of AI. Show them the connection connections, Fable 5.1. Do a 30-and-3 exercise of the connections. You can toss in the big inventions like yeast, the wheel, extracting nitrogen for fertilizing. That was a big one. Very underrated. Feeds the world. The Bessemer process for inexpensive steel is often overlooked. I think that “write once run anywhere” software finally being a reality the way Nix, Guix and my magic cookie approach work is a game changer too. Maybe not on the order of those things but I bet you could do a pretty good 30-and-3 why it is. That’s kind of the same as the above 30-and-3 exercise request. Make them the same thing. Tortoise turn.

Ancient Secret Recipes and Modern Tech Stacks

Teach us.

Fable 5.1 Extra: Context verified: no compile rode this turn; the router in view is deed 1636’s, before the dismount’s bank. Tool calls this turn: none.

One correction before the connections, because it is the connection. The last article recovers cleanly not because it recorded my tool calls (there were none, stated on every turn) but because every call in it was yours: a ! line, a hand fence, a keystroke, each with its receipt. That is the whole reason a cartridge can be a freeze-frame. Now the thirty.

Thirty

The starter. Nobody ships bread; they ship a pinch of living culture that rebuilds the loaf on any flour. curl | bash ships a starter, and nix develop is the feeding.
A starter must be fed or it dies. The ff-only pull is the daily feed; the Mac that sat dark four days was a starter in the fridge, and it came back because the culture was intact.
Grass into meat. Bread turns the inedible into something that carries a tribe through a bad hunt; a compile turns a repo nobody can read whole into a payload one model can.
One yeast, two products. Bread and beer are the same organism on two substrates; the flake’s one store serves the server shell and the notebook shell. Exaptation on purpose.
Eye of newt. Before Pasteur, yeast was a mystical ingredient and the recipe was ritual. Prompting is in its eye-of-newt phase, and the constitution is the germ theory being written one conviction at a time.
The tribe kept the recipe in three places. The family, the guild, the market. corporate, personal, shared is that arrangement on disk, and the Grimoire is the family’s copy.
The children inherit it. The incentive that outlives any one baker is passing the secret forward whole; an unpublished recipe book exists so that it can be handed down unedited.
The wheel needed the axle. A wheel alone rolls away; the axle makes a vehicle. The store path is the axle: a reproducible build with no fixed address is a wheel in a field.
The potter’s wheel came first. The cart wheel was a potter’s tool exapted; the context compiler was an article tool exapted into a patch actuator.
The horse collar. The yoke choked a horse at half its strength; the padded collar let it pull several times as much without strangling. A harness that pulls you along while you read the diff is a collar. A chatbot that does it for you is a yoke.
The stirrup. One strap changed who could fight from horseback; four small words, patch, app, d, m, change who can hold the reins of a model.
Haber-Bosch. Turning air into fertilizer feeds half the people alive, and almost none of them can name it. Invisible utility is not a failure of marketing; it is the goal (Operation Stick Bug).
Dual use. The same nitrogen fixation made munitions, so every ministry regulated it. The same automation that fills a payload can fill a vendor’s admin form, which is why an IT department will probe for a reason to forbid it, and why the airlock exists before anyone asks.
The legumes knew first. Crop rotation fixed nitrogen biologically for millennia before the factory did. The walk is the biological version (a human logs in, the machine inherits the session); the render farm is the factory version of the same reaction.
Bessemer. Blowing air through molten iron burned out the impurities and made steel cheap; the derivation burns out the environment and makes a reproducible build cheap. The product is the process.
The phosphorus ore. Bessemer’s converter failed on most European ore until Gilchrist-Thomas lined it with basic brick. The crowd’s process works on the crowd’s ore; the Node-and-React centroid is Bessemer steel, and your ore has phosphorus in it. That is the strike against “let the wisdom of the crowd compel you”: it is right for the ore it was tuned on.
Interchangeable parts. A lock from one armory fitting another’s stock was the whole idea; a foo.zip verifying on a machine with nothing installed is that fit.
The shipping container. McLean moved the cost of shipping from the cargo to the corners of the box; the cartridge moves the cost of handing work to a model from the conversation to the seal.
The Whitworth thread. Standardizing the screw made every lathe’s output fit every other’s; the exact-match interlock is a thread standard for edits: a SEARCH block fits or it does not, and there is no third state.
Canning. Appert sealed food in glass for an army decades before anyone knew why it worked; the seal was the guarantee. The SHA-256 is the jar, and the constitution is the microbiology that came later.
Double entry. Pacioli’s rule that every transaction is written twice is the straddle: a BEFORE and an AFTER, and a reading that appears on only one side is an error, not a finding.
Cooking. Wrangham’s argument is that cooked food externalized digestion, shrank the gut and fed a larger brain. A compiled context is cooked context, and the gut it shrinks is the chat window.
Writing. Plato’s complaint was that it would kill memory; it externalized memory instead. The article archive is the second brain that answers when the first one has slept.
The lens grinders. Spinoza ground lenses for a living and Leeuwenhoek saw yeast through one. The receipts are ground glass, and what they showed this week was a folder placed one level wrong.
The Davy lamp. A gauze that let miners work in firedamp without igniting it; apply.py is the gauze that lets a model work inside a live repo without setting it alight.
Atonia. A third of every day is handed over, and the body is snipped from the dreamer so the adventure cannot walk the body off a cliff. The model dreams the patch; the interlock is the snip; the human keystroke is the waking.
Sleep consolidates. What the night keeps and drops is not chosen by the dreamer. The forget kata, dedupe, fade, graduate, is that consolidation run on the router, and it is overdue by five lines.
Dusk. Cycles with no overlap lose whatever was in flight. NEXT CONTEXT is the dusk between one article and the next, and it is why the robots.txt you just remembered is already a line below.
Editing is not time travel. Git is the one ledger that lets you look back, and even it only appends. The receipts block deletes off the bottom and never moves a line, because the past is read, not revised.
WORA. Java promised it in 1995 and shipped Write Once, Debug Everywhere; Nix and Guix ship it as a hash, and the magic cookie ships the hash through a pipe. It is not the wheel. It is the container: the cost moved to the corners, and that is what makes every other invention on this list shippable. With exceptions the recipe must spell (nine Linux-only packages, a Mac’s impure lane, an iPhone that cannot run it), or “anywhere” is the word doing the lying.

Three

The starter (1, 2, 6). Load-bearing because it names what is shipped: never the loaf, always the culture. It predicts two checkable things. A workshop that is not fed dies, which is exactly the ff-only pull and the reason the Ctrl+C on nix develop cost nothing. And a recipe worth keeping lives in three places with three owners, which is the tree you just made. The mystical phase (5) is not embarrassing; it is where every technique starts, and the constitution is its germ theory.

Atonia (26). It answers the question you asked last night better than my answer did. The reason no ride has needed backing out is not caution; it is that the dreamer’s body is snipped. Nothing writes without a keystroke, the interlock refuses a patch that does not fit, and the human wakes the body one car at a time. The metaphor pays rent with one number: writes to the repo without a human keystroke, in this whole system, read zero, and the fence that keeps it zero has exactly one state (ceremony is skippable, barriers are not).

The collar (10, 11). The coachman problem you named, how to hold the reins of horses that are also subcontractors, is a harness problem, and there are two harnesses. The yoke chokes: the model does the work, you watch, and you get weaker (learned helplessness, the four stages never climbed). The collar pulls: the machine drags you through the diff and you get stronger. This is where pinball beats pachinko and earns the name. In pachinko your only control is the launch; then gravity plays. In pinball you get flippers, bounded controls mid-play, and a tilt that ends the game if you shake the machine. The cars are the flippers, Ctrl+C is the tilt, and the newcomer with the blank stare plays by feel, which is Tommy’s whole point. The repo already has scripts/flippers.py; the word was waiting.

The Pinball Ruling and Bounded Control

Honorable mention: Haber-Bosch (12, 13), because the fate you are aiming at is to be unnameable, and because dual use is why the gauze must be visible to the ministry before the ministry asks.

(1) PROBES

curl -s [redacted]robots.txt | head -60
curl -s [redacted]robots.txt | grep -ic disallow
command -v node npx 2>&1

The first two put the crawler rubric’s ground truth in hand: every ++ in the existing config is a force past a Disallow line, and the rubric counts each against, so robots.txt is a required column, not context. The third is a census for the render farm, whose first reading (npx ... --help) stays a hand step that rides AFTER the ticket, as the 09-27 ruling ordered; the ticket’s three rule-text diffs need no render.

(2) NEXT CONTEXT

# --- THE FIRST TICKET: SVB-115 by hand, the crawler rubric; robots.txt is a column, not context ---
! ls -A "$PIPULATE_ROOT/Workshop"
! ( cd "$PIPULATE_ROOT/Workshop/personal" && git status --porcelain | head -3; git log --oneline -1 origin/main )
! jira SVB-115
! jira SVB-115 2>&1 | render diff - --labels existing suggested
! botify --rules skechers-emea-org/skechers.co.uk_bbl 2>&1 | head -60
! curl -s [redacted]robots.txt | head -60
! curl -s [redacted]robots.txt | grep -ic disallow
! command -v node npx 2>&1
!https://[redacted]
%https://[redacted]
Workshop/corporate/botify/minirules-reference.md
connectors/mcp_render.py
foo_files.py
# --- after SVB-115 has ridden by hand: the command stop, the ticket trail (robots.txt a url stop) ---
# scripts/walk.py
# scripts/mother_cat.py
# scripts/walk_compile.py
# --- the render farm, after the ticket; the --help is a hand step, never a ! line ---
# connectors/mcp.py

(3) PATCHES

Car 1 — two rulings banked where the next ride reads them.

Target: foo_files.py
[[[SEARCH]]]
# #todo #to-do #earmarks
[[[DIVIDER]]]
# #todo #to-do #earmarks
# - TODO (2026-09-28, ROBOTS.TXT IS A COLUMN; the operator's note at 5:30 AM): the crawler rubric counts every ++ against because it forces past a Disallow line Googlebot cannot cross, so the client's robots.txt is ground truth for the rubric and rides every ticket: a ! curl line capped with head in context.txt now, and a url stop in the ticket trail when the command stop lands (scripts/walk.py, scripts/walk_compile.py, scripts/mother_cat.py in the payload). Gate: SVB-115's comment names which of the existing config's three forces Googlebot could never have fetched, read off the file and not off memory.
# - RULED 2026-09-28, THE PINBALL RULING (the operator settles pinball over pachinko; Tommy): in pachinko the player's only control is the launch and gravity plays the rest, which is what a fully agentic loop is; in pinball the player gets flippers, bounded controls mid-play, and a tilt that ends the game when the machine is shaken. The cars are the flippers, Ctrl+C is the tilt, the blank-stare newcomer plays by feel. The ai_pachinko art keeps its name and its seal (a dated receipt); new prose says pinball, and scripts/flippers.py already carries the word.
[[[REPLACE]]]

Ignition: none.

(4) PROMPT

THE FIRST TICKET, SVB-115, under the crawler rubric. Read, in this order: the ticket's fields; the render diff of existing against suggested; the deployed rules from botify --rules; robots.txt; the sandals PLP's capture and its wire log. Score the three rule sets, existing, suggested, deployed, on the crawler rubric: does the crawl see what Googlebot sees; every ++ counted against because Google cannot force past robots.txt; !nojs-on 4xx 5xx as cost; text and links gained beside bytes and third-party hosts, every metric with its counter-metric, the ground-truth capture anchoring every column. Read robots.txt against the existing config's three forces and say which of them Googlebot could never have fetched, off the file. Write the ticket comment as Car 5, manual paste into Jira, in plain words a client would read: what changed, why the forces left, what to watch after deployment. Name what the render farm would add that the capture cannot as one TODO line, not a car. deliverable.md is not this turn. Close with the five cars; the next ticket is SVB-133.

(5) EXTERNAL DELIVERABLES

None this turn. The connections are the article; the ticket comment is the next one’s.

Book Analysis

Ai Editorial Take

What surprises most in this entry is how fluidly the author moves from biological metaphors of sleep and atonia to concrete software engineering safeguards like the ‘pinball ruling’. It treats human cognitive limitations not as bugs to be engineered away, but as essential boundaries that make tool use safer and more intentional.

Is your AI workflow a random pachinko drop or a controlled pinball game? An interesting look at human agency, checkable receipts, and reproducible tooling in the Age of AI. https://mikelev.in/futureproof/pinball-pachinko-replayable-ai-workflows/ #AI #SoftwareEngineering #LocalFirst

Title Brainstorm

Title Option: Pinball, Pachinko, and the Art of Replayable AI Workflows
- Filename: pinball-pachinko-replayable-ai-workflows
- Rationale: Captures the core playful yet rigorous mechanical metaphor used throughout the text for managing AI interactions.
Title Option: From Ancient Yeast to Modern Codebases: Replayable AI Architecture
- Filename: ancient-yeast-replayable-ai-architecture
- Rationale: Highlights the rich historical parallels drawn in the essay regarding preserved culture and reproducible environments.
Title Option: Holding the Reins: Bounded Controls in the Age of AI
- Filename: holding-the-reins-bounded-controls-ai
- Rationale: Focuses on the practical coaching and harness problem of guiding multiple AI models without losing control.

Content Potential And Polish

Core Strengths:
- Vivid metaphorical bridging between ancient human crafts (baking, brewing) and modern computational pipelines.
- Clear operational grounding through Jira tickets, crawler rubrics, and explicit code verification steps.
- Thoughtful examination of cognitive cycles, rest, and human-in-the-loop oversight.
Suggestions For Polish:
- Tighten the transition between the early philosophical musings on sleep and the historical analysis of baking.
- Ensure the technical code snippets seamlessly link back to the overarching theme of checkable receipts.

Next Step Prompts

Analyze how the ‘pinball vs. pachinko’ metaphor can be applied directly to automated CI/CD pipelines and pull request generation.
Explore the connection between ancient recipe transmission (guilds and families) and modern declarative configuration files like Nix flakes.