---
title: "The Staircase: Four Steps for Every Job in the Plant, and the Top Step Stays Empty"
date: 2026-09-13
description: "Every plant job on a four-step staircase — dashboards, scripts, a model call, agents. The top step stays empty; the agent purchase case collapses."
category: "Industrial AI"
tags:
  - "LLM ops"
  - "AI agents"
  - "token economics"
  - "industrial automation"
  - "manufacturing"
  - "procurement"
verifiedHardware: "LineWatch patrol system on two live production lines at a discrete-manufacturing plant · staircase assembled 2026-09-12 from the Episode-4 audit and the same-day lane rebuild · dual-lane comparison mornings 2026-09-11→09-13 (12 windows, two recipes) · ruling delivered 2026-09-13 · the rebuilt lane's first live event shift 2026-09-12 20:55, with the next morning's green shift as the zero-call control"
keyTakeaways:
  - "Every job in the plant was placed on one four-step staircase: fixed pages, fixed scripts, script-plus-one-model-call, agents. The lower two steps carry no AI at all — the Episode-4 audit already proved the whole 30-day chain had exactly one model call point. The one rule: stand as low as the job allows. The AI-summary lane itself was born on Step 4 and climbed down to Step 3 the same week the audit priced its harness at 97% door fee — scaffolding, taken down when the building was finished."
  - "The word 'need' was retired from the record, by the operator's own pen: an agent is an elevator, not the electricity — no job loses the ability to be done without one. The live question is which jobs are worth hiring a temp for: the work is new, the volume clears the entrance fee, and a guardrail exists. Dissected, an 'industrial agent' purchase is four parts — model capability (a commodity), SOP capture (the expensive part, which no vendor can supply), a scheduling skeleton (cheap), and the autonomy loop (the only new thing, priced negative by this plant's own audit). One question exposes any costume: at run time, who decides the next step — the code, or the model?"
  - "The Step-3 contest stayed open past its deadline on purpose. On 2026-09-13 the ruling kept both lanes — cloud-direct and on-prem — running side by side to 2026-09-20: the comparison costs 13.2K–18.5K tokens a day, the local lane bills zero API, and evidence outlives a rushed decision. Across 12 windows the on-prem lane went 12-for-12 (124–171 s); the cloud experiment recipe blanked twice in twelve, both at exactly the output cap — Episode 3's exam physics again, not a lane verdict. The rebuilt lane's first live event shift (2026-09-12 20:55) summarized both lines successfully inside 3 min 35 s end-to-end; the next morning's green shift made zero calls."
measuredOn: 2026-09-13
faq:
  - q: "Two blank papers in twelve windows — doesn't that rule out the cloud lane?"
    a: "No, because the comparison's cloud recipe is not the production recipe. The experiment lane ran an earlier-generation model, bare (no manual mounted), with a double-size output pad of 8,192 tokens — a recipe chosen to stress the lane, not to mirror production, which runs the current generation with the manual mounted and a tighter pad. Both blanks hit the cap exactly (8,192 output tokens, one returning only a section number), which is the physics Episode 3 established on the local 27B and Episode 4 met again on the cloud model the same week: draft and body share one budget, and budget is a property of the exam, not of the examinee. A blank under the stress recipe is a data point about the recipe. It is not evidence against the production lane — and the record keeps that distinction explicit, including a correction: an early log note named the wrong line for one blank; the comparison reports, not the log, are the authority."
  - q: "Keeping both lanes running sounds like refusing to decide. Why extend?"
    a: "Because the arithmetic favors patience. Both lanes together burn 13.2K–18.5K tokens per day, of which the on-prem lane bills zero API cost — the price of continued evidence is trivial next to the cost of a premature decision. The operator's ruling, paraphrased from the session record: the comparison is cheap, so let it run. The window moved from 09-13 to 09-20 in a one-line change to the comparison script's exit condition, syntax-checked; the decision was not cancelled, it was re-priced."
  - q: "Wasn't there a discrepancy in Episode 4's numbers?"
    a: "Yes, and it is now settled. The gap lived in Episode 4's verification record, not its published text: that record's per-day table summed to one figure while a prose note in the same record carried another, 500 apart. Re-addition of the three steady-state rows (207,528 + 416,208 + 412,036) fixes the total at 1,035,772; the prose figure, 1,036,272, was a transcription slip that no combination of rows can produce. Episode 4's published text rounds to 'about 1.04M' and needed no correction on the site — the precision of the record, not the rounded prose, is what carries the arithmetic. The correction lives here, in the open, per this series' practice."
  - q: "Does the staircase generalize beyond this plant?"
    a: "With its boundary stated: the sample is one class of work — highly enumerable jobs with closed inputs and outputs, judged against written rules. A different site, with heterogeneous ad-hoc analysis demands across many lines and no in-house engineering, might land differently. But that is precisely the case the vendors sell to, and it is not this case — and even there, the four-part dissection holds: the expensive parts (SOP capture, data integration) are exactly the parts a platform cannot supply, and the part it adds (the autonomy loop) is the part an audit like this one prices as a liability. The vendor demo's natural-language query is reachable from Step 3 — skeleton, one model call, manual — with a replayable receipt."
---
[Episode 4](/posts/2026-09-12-machines-keep-the-watch-ep4-the-bill/) ended with a bill and a question. The bill: the plant's only AI call, rebuilt the same day to a twenty-fifth of its price. The question: if one agent invocation cost twenty-five times its content and used none of its tools, what is an agent actually *for* — and what, exactly, would this plant be buying if it bought one?

The question was not academic. The operator's company was weighing an industrial-grade agent platform for the plant. So the same afternoon the audit closed, the operator built the answer: a staircase, with every job in the plant standing on it. This episode is that record — the staircase, the law the plant's own history had already been obeying, the purchase case taken apart screw by screw, and a ruling that arrived on schedule and chose to wait.

## Four steps, one rule

| Step | What it is | Real jobs from this plant | Who decides the next step | Marginal cost |
|---|---|---|---|---|
| 1 · Fixed pages | dashboards, query and reconciliation windows | the device dashboard, the data-convergence window, the line-side raw-statistics view | designed in advance; a human clicks | free |
| 2 · Fixed scripts | scheduled Python that runs itself | the three patrol shifts, the changeover daily, incremental sync, PDF rendering and delivery | the code, written once | **0 tokens** |
| 3 · Script + one model call | the machine phones a scribe | the patrol running summary: facts read out, prose read back, hang up | the code; the model only writes | **~2K tokens per call** |
| 4 · Agents | the model decides steps and picks tools | the drafting sessions behind this record; one-off unattended batch runs | the model, at run time | door fee, tens of thousands of tokens per entry |

The one rule: **stand as low as the job allows.**

The lower two steps carry no AI at all, and that is not aspiration — it is the audited state of the plant. [Episode 4](/posts/2026-09-12-machines-keep-the-watch-ep4-the-bill/)'s forensic sweep of the whole patrol chain found exactly one model call point in thirty days of running: the running summary, on Step 3. Judgment, classification, rendering, delivery — all Step 2, all zero tokens. Most of the plant's work never climbs past the second step, which is why the top step, for this plant, stays empty.

## The scaffolding law

The staircase was not a new theory bolted onto the plant; it was a description of what the plant's history already did. Three times, the same pattern:

- **A threshold, found by AI, written into code.** During an AI session, a pattern emerged in the event-window envelope: data dropout under 20% marks stretches where two faces change together, while a single-face cover sits near 28% and never qualifies. Confirmed by the operator against 36 days of data — four qualifying stretches, each with an account, zero misses — it was written into the monitor the same morning. The insight was AI's; the rule is now Step 2's.
- **A research result, promoted into the daily.** The overnight consistency study had produced conclusions humans had to go read. On 09-07 its verdict moved into the morning report's fixed section; the next day brought the exemption for covered faces. Step 4's output, re-homed on Step 2.
- **The summary itself, born on Step 4, climbed down.** The lane entered the world on 09-05 through a full agent harness — seven days, 44 calls, 2,239,178 tokens — and on 09-12 was rebuilt into a single direct call at a twenty-fifth of the price. [Episode 4](/posts/2026-09-12-machines-keep-the-watch-ep4-the-bill/) told it as a bill. On the staircase it is something else: scaffolding, taken down the week the building was finished.

The law, as the operator wrote it that day: an agent is scaffolding. It is hired for the part of the job that is still unfinished — the exploring, the deciding-what-to-do-next. When the job is understood, it becomes a fixed step, and the scaffolding comes down. Which reframes the purchase under consideration: **a resident agent platform in a plant of finished jobs is scaffolding rented forever on a building already topped out.**

## Retiring the word "need"

The first draft of the operator's same-day memo listed four classes of work that "need" an agent. The operator struck the word himself. Debugging runs through a coding agent today — but ran for decades without one, on grep, logs and patience. An agent creates no capability that a human with fixed tools lacks; it saves labor and time. **An agent is an elevator, not the electricity.** You can always take the stairs.

The live question becomes: which jobs are worth *hiring a temp for*? Three tests, all three at once — the work is new enough that no instruction card can be written for it; the volume or urgency clears the entrance fee; and a guardrail exists, so the temp's mistakes are catchable. Applied to this plant, agent usage survives in exactly three forms, none of them on the production floor: the drafting sessions behind this very record (no instruction card exists for "write the next episode"); the engineer's chair, where an agent turns a three-hour search into twenty minutes — an accelerator, kept optional; and one-off unattended batch runs, hired for a night, cage and all. The plant's runtime has no seat for one, and not for lack of imagination: runtime work is SOPs, and **in a world of standard operating procedures, improvisation is not a skill — it is a violation.** The agent's one talent, deciding the next step on the spot, is the one talent the floor forbids.

## What a purchase would actually buy

Take an industrial agent platform apart and four components fall out:

| Component | Nature | This plant's standing |
|---|---|---|
| Model capability | a commodity — swap by API | dual-lane already in production: a frontier cloud model and an on-prem 27B, same manual, interchangeable |
| SOP and domain-knowledge capture | **the expensive part** — and no vendor can supply it; the domain must feed it | a manual library (four books) plus the judgment constitution, on a measured iterate-and-retest cycle |
| Scheduling skeleton | cron, retries, queues — cheap | patrol timers, comparison harness, unattended shifts: all running |
| The autonomy loop | the only component a purchase *adds* | priced negative by the plant's own audit: door fee, variance, unauditable steps |

The accounting is the uncomfortable part for the purchase case: the cost of "industrial AI" was never in component four — it is in components two and three, which a platform does not remove and cannot supply. The buyer pays for the one component the audit priced as a liability.

And a caution about costumes: a good deal of what is sold as "industrial agents" is Step 2 wearing a badge — if the steps are written in advance, it is a workflow, whatever the box says. There is one question that strips the costume off any demo: **at run time, who decides the next step — the code, or the model?** If the purchase proceeds anyway, the record's four tests travel with it: fully local or out; SOPs exportable as files; replayable, deterministic audit; and the model replaceable by API — this plant's one-manual-two-models lane is the fourth test passed in production.

## The ruling that chose to wait

One contest inside Step 3 remained genuinely open: cloud-direct or on-prem for the summary lane. The comparison had run each morning from 09-11 — the same sanitized fact sheet to both lanes, only the model and the manual differing. Its window was to close with a ruling on 09-13.

The ruling came, and it was: **not either-or.** Both lanes stay; the window extends to 09-20. The operator's reason, paraphrased from the session record: the comparison costs almost nothing, so let it keep running. The numbers agree with him — twelve windows across three mornings:

| Mornings 09-11→09-13 | Cloud lane (stress recipe: earlier model, bare, 8,192 pad) | On-prem 27B (manual mounted) |
|---|---|---|
| Windows completed with prose | 10 of 12 | **12 of 12** |
| Blanks | 2 — both at exactly the output cap | 0 |
| Wall time, completed windows | 20–27 s | 124–171 s |
| Tokens, both lanes per day | 13,196 → 18,089 → 18,522 (local share: 7,389 / 7,675 / 8,026, zero API bill) | |

Two footnotes keep that table honest. First, the blanks: the stress recipe is not the production recipe — an earlier-generation model, no manual, a double-size pad — and both blanks stopped at the cap exactly, the same exam physics [Episode 3](/posts/2026-09-11-machines-keep-the-watch-ep3-blank-paper-exam/) established and [Episode 4](/posts/2026-09-12-machines-keep-the-watch-ep4-the-bill/) met again. They indict a recipe, not a lane. Second, the daily total: it is *both* lanes combined, and the blank windows were its largest single items at 8,192 apiece — which is precisely why the operator could afford patience. Evidence at that price is a bargain; a rushed decision is not.

The extension itself was a one-line change — the comparison script's exit date moved from 09-13 to 09-20, syntax-checked, effective the next morning. The decision was not cancelled. It was re-priced.

## First live round, receipt and all

[Episode 4](/posts/2026-09-12-machines-keep-the-watch-ep4-the-bill/) closed waiting for the rebuilt lane's first live event shift. It came at 20:55 that same evening: an event verdict on both lines, the gate fired, and both summaries came back — the service journal logged both lanes' AI calls complete at 20:59:22, the whole chain from trip generation to alert, PDFs and delivery inside 3 minutes 35 seconds. The samples read as the manual demands: bold verdict first, plant terms only, every sentence carrying its source field. The next morning's 06:30 patrol came back green on both lines, and the lane made zero calls while the reports rendered on schedule — the control arm, clean.

One line is missing from the receipt, and the record says so: the production path prints token usage only on failure, so the first live round has no token-level bill — its cost anchor remains the A/B measurement (~2K per call). A one-line change would print usage on success too; it touches the production script, so it waits for approval rather than shipping quietly.

## What the staircase did not cover

The audit had now reached every job in the plant: the chain, the lane, the manuals, the comparison, the purchase case. Every seat had been priced except one — the desk the audits were run from, where AI is used the way water is used, all day, in the open. That desk is [the next episode](/posts/2026-09-12-machines-keep-the-watch-ep6-the-compaction/).

## Sources and method

First-party: the operator's same-day memo of 2026-09-12 (the staircase, the temp-hiring tests, the four-part purchase dissection — including the operator's own strike of the word "need"), the three comparison reports of 2026-09-11→09-13 (per-window tokens and wall times verbatim; the early log note that misnamed one blank line is corrected here against the reports), the ruling of 2026-09-13 as landed in the comparison script's window line, the service journal entry of 2026-09-12 20:59:22 and the next morning's green-shift report, and the re-addition that settles Episode 4's 500-token discrepancy at 1,035,772. Assembled into English with AI assistance under human editorial direction; facts and numbers unchanged from the records; derived figures marked as computed. Deliberately absent, per the series' disclosure policy: anything identifying the plant, its industry, its operator or its network, the identities of the harness and model products, and internal identifiers. The measured numbers are registered in the [/data/ ledger](/data/).

*[All episodes](/series/machines-keep-the-watch/) — Machines Keep the Watch, a field-record series. The map: [the anchor](/posts/2026-09-05-machines-keep-the-watch-future-production-line/) · Episode 1: [the dress rehearsal](/posts/2026-09-05-machines-keep-the-watch-ep1-first-night-shift/) · Episode 2: [the second without AI](/posts/2026-09-05-machines-keep-the-watch-ep2-190000-no-ai/) · Episode 3: [the blank-paper exam](/posts/2026-09-11-machines-keep-the-watch-ep3-blank-paper-exam/) · Episode 4: [the bill](/posts/2026-09-12-machines-keep-the-watch-ep4-the-bill/) · Episode 6: [the compaction](/posts/2026-09-12-machines-keep-the-watch-ep6-the-compaction/).*