A wrong forecast is absorbed. A wrong action reaches a guest.
Prophet is a statistical forecasting library. Claude is a large language model. Keeping them as two components, rather than letting one model both predict and act, is the decision that shaped everything after it.
- Read this if
- you are deciding how much one model should be allowed to do on its own.
- What it gives you
- a rule for where to draw the line (the cost of an error, not the technology), and the trade-offs it brings.
- Tested on
- Aetherix, an F&B agent for hotels, built and tested on synthetic and public data. No hotel connected yet.
- **Draw the line where the consequence changes.** A wrong forecast is absorbed; a wrong action reaches a guest.
- **Five jobs, five owners.** Forecast, recommendation, enforcement, approval and outcome are separate, and each is checked differently.
- **A rule in a prompt is a request.** A rule between components is code, and a violation is a failing test.
- **The split has a price.** Two systems, one contract to maintain, and a forecast that is good, not magic.
The boundary is where a claim about behaviour gets held by something other than a prompt.
When you build an operations layer for boutique hotels, the temptation is to let one large model do everything: predict, decide, act. It looks elegant on a slide. One component, one prompt, one place to improve.
I split it in two. , an open-source forecasting library built for time series with strong seasonality, reads what is about to happen: how many covers tomorrow, with an uncertainty range. , a large language model, decides what to say about it and to whom. Neither does the other's job, and almost every later decision in the system inherits from that line.
The line is not drawn where the technology changes. It is drawn where the consequence changes.
A wrong forecast is absorbed. Order too much and there is waste, too little and there is a run to the supplier. It costs money and nobody's evening. A wrong action reaches a guest, a shift roster, someone's Tuesday. The two failures do not deserve the same amount of freedom, and any component that owns both has to decide, silently, which one it is optimising at any given moment.
That is the argument for the boundary, and it is not really about Prophet or Claude. It is about which failures you are willing to let a model resolve on its own.
Give one component both jobs and it arbitrates between them privately. That is the cost, and it is not a performance cost. A node that owns two domains stops being reviewable: when the recommendation is wrong you cannot tell whether the read of the situation was wrong, or the response to a correct read was wrong. There is one output and one prompt, and the trace tells you nothing you did not already know.
Forecasting and acting also want different feedback. Forecast error computes itself: tomorrow's covers are known the day after, so the number arrives whether or not anyone asks for it. Whether a recommendation should have been made at all is a judgement, and the only thing that returns it is a human. One loop is arithmetic, the other is an opinion. A single component blends them and you lose the ability to measure either. Why a manager's approval and the next day's covers are different signals, and which one a system may learn from, is the ground-truth note.
Prophet reads the environment: occupancy, demand, staffing pressure. It produces a number and an error bar and nothing else. It never phrases anything, never recommends, never talks to a manager.
Claude orchestrates the response. It takes that read, plus memory of what worked at this property, and composes one message a day. It never predicts. When it is uncertain it says so.
The manager replies accept, reject, or the right figure. That reply is not politeness. It is the recorded output of a decision node.
| Job | Owner | What it produces | How it is checked | Evidence today |
|---|---|---|---|---|
| Statistical forecasting | Prophet | a number and an uncertainty range | error against a frozen baseline | synthetic series in CI; public benchmark |
| Recommendation | Claude | one message a day, uncertainty stated | graders on the reply parser | stand-in parser; model not yet measured |
| Deterministic enforcement | code between the two | a refusal to act on an implausible or low-confidence input | a test that fails on violation | exercised in a simulated 30-day scenario |
| Human approval | the F&B manager | accept, reject or the right figure | recorded as an event | sandbox data; free text not yet mapped |
| Measured outcome | the till, next day | real covers | forecast error computes itself | not yet: needs a connected hotel |
A table of owners says who does what. It does not say what a manager lives through. Here is one illustrative breakfast, read across the four moments of a service and the layers that carry it.
Service blueprint
Tomorrow's breakfast, from forecast to feedback
Illustrative scenario: a hotel F&B manager preparing tomorrow's breakfast service. No hotel is connected yet; no manager, guest outcome or metric here is observed.
The manager is drafting tomorrow's breakfast rota between two other tasks.
IllustrativeThe manager reads one message, weighs what the system cannot see (a sick waiter, a group check-in), and replies.
IllustrativeThe team serves breakfast on the plan the manager chose. Guests feel the staffing, right or wrong.
IllustrativeThe manager glances at the Daily Receipt: forecast against actual, and the gap.
IllustrativeNothing yet. The system works before it speaks.
Built, sandbox dataOne WhatsApp message: expected covers, confidence, suggested staff. Reply accept, reject or the right figure.
Built, sandbox dataNo screen. The decision now lives in the rota and on the floor.
IllustrativeDaily Receipt: actual covers, forecast, delta, confidence.
Built, sandbox dataProphet forecasts covers from occupancy and history, with a range. Code refuses an implausible or low-confidence input.
Built, sandbox dataClaude composes the message and never predicts. The reply is parsed and stored next to the recommendation.
Built, sandbox dataNothing acts. The system does not change the rota or contact staff.
Built, sandbox dataActual covers ingested from the till; forecast error computed; decision logged for audit. Live till ingestion needs a connected hotel.
PlannedThe forecast is a claim with an error bar. No authority: it decides nothing.
Authority passes to the manager. Their reply is declared feedback: a preference, not a fact.
→ Authority changes hands
The action reaches people here. It cannot be replayed: the plan not chosen is never seen.
→ The action reaches people
Observed: the covers, so the forecast can be scored. Not observed: whether the staffing was right. That stays a judgement.
01 · The day before
Anticipate
- On the floor
The manager is drafting tomorrow's breakfast rota between two other tasks.
Illustrative- What the manager sees
Nothing yet. The system works before it speaks.
Built, sandbox data- Behind the scenes
Prophet forecasts covers from occupancy and history, with a range. Code refuses an implausible or low-confidence input.
Built, sandbox data- Authority and evidence
The forecast is a claim with an error bar. No authority: it decides nothing.
02 · Evening
Decide
→ Authority changes hands
- On the floor
The manager reads one message, weighs what the system cannot see (a sick waiter, a group check-in), and replies.
Illustrative- What the manager sees
One WhatsApp message: expected covers, confidence, suggested staff. Reply accept, reject or the right figure.
Built, sandbox data- Behind the scenes
Claude composes the message and never predicts. The reply is parsed and stored next to the recommendation.
Built, sandbox data- Authority and evidence
Authority passes to the manager. Their reply is declared feedback: a preference, not a fact.
03 · Breakfast service
Operate
→ The action reaches people
- On the floor
The team serves breakfast on the plan the manager chose. Guests feel the staffing, right or wrong.
Illustrative- What the manager sees
No screen. The decision now lives in the rota and on the floor.
Illustrative- Behind the scenes
Nothing acts. The system does not change the rota or contact staff.
Built, sandbox data- Authority and evidence
The action reaches people here. It cannot be replayed: the plan not chosen is never seen.
04 · After service
Learn
- On the floor
The manager glances at the Daily Receipt: forecast against actual, and the gap.
Illustrative- What the manager sees
Daily Receipt: actual covers, forecast, delta, confidence.
Built, sandbox data- Behind the scenes
Actual covers ingested from the till; forecast error computed; decision logged for audit. Live till ingestion needs a connected hotel.
Planned- Authority and evidence
Observed: the covers, so the forecast can be scored. Not observed: whether the staffing was right. That stays a judgement.
A prediction can be evaluated against a later observed number. A recommendation has operational consequences and does not automatically receive an equally definitive verdict.
Evidence status
- Built on sandbox and public data (Apaleo sandbox, synthetic series, Kaggle Recruit benchmark): forecast, guard in code, WhatsApp message, reply parser, daily receipt.
- Planned: live ingestion of till data from a connected hotel, which is what would let forecast error be measured on real covers.
- Unproven: that recorded replies improve recommendations, and that any staffing decision improved a real service. No manager has used it yet.
Two things I could not have had otherwise.
The forecast becomes measurable on its own terms: compared to a frozen baseline, in CI, blocking a merge. Today that series is synthetic; against realised covers, it starts with the pilot. That number means something precisely because the component producing it does one thing. The gate and its baseline can be inspected in the public Aetherix repository, opens in a new tab; the code shows the check exists, not that it works on a real hotel.
And the action becomes accountable. Every recorded accept, reject or correction is an event with the recommendation stored beside it. Today that means an exact accept or reject, or a figure: a free-text reply is not yet mapped to a decision, a gap an audit found in September 2026. A rejection is not lost feedback, it is a new evaluation case. Human in the loop by design rather than by default, and the difference is whether the decision leaves a trace.
The seam is also what lets the system be honest about itself. In a simulated 30-day scenario, it rejected an implausible point-of-sale figure, injected as a test, rather than training on it. After a simulated 25% jump in demand, weekly recalibration cut its forecast error from 15.9% to 12.3%. Neither is possible when one component owns the whole path, because there is no seam at which to check.
Two systems instead of one, and a contract between them that has to be maintained, versioned and tested. A forecast that is good and not magic: on a public benchmark of 30 restaurants (Kaggle Recruit data), Prophet beats a naive same-weekday baseline (the mean of the last four same weekdays) on mean error, ties it on the median, and loses the median to a gradient-boosted baseline. Both readings are published. The forecast runs in shadow mode: no manager decision has been delivered on it yet.
And a limit I have to keep restating. The comparison runs on its own, but the truth it compares against is still declared by a human. That is a closed loop on a declared truth, which is an engineering statement. "Dual autonomous loop" would be a brochure statement.
The guarantee. In a single-model design, "never act on a low-confidence forecast" is a sentence in a prompt, which is a request. Requests are not enforced, they are usually honoured. With the boundary in place the rule lives between the components, in code, where it can be checked and where a violation is a failing test rather than a disappointing output. How such a test is written, graded and allowed to block a release is the subject of the evaluation note.
A claim about behaviour has to be held by something other than a prompt. The boundary is where I put it.
- The decision
- Which decisions the system may take alone, which it may only recommend, and which require a named human's approval.
- Inputs
- An inventory of the decisions the application makes or recommends
- Who is accountable for each one
- What it costs to undo each one
- The current prompts and the code paths that act on outputs
- The artifact
- A decision register classified by impact, reversibility and required authority, with an approval point, a fallback and an audit record for each.
- Verification
- For every rule marked "enforced", show the code location and a test that fails when the rule is violated. A rule that only exists in a prompt is not enforced.
Ask your agent to apply this method
A brief for your own coding or research agent, applied to your project. To question this article instead, ask Rose.
Read the brief
Apply a method to my own AI project. Method source: "A wrong forecast is absorbed. A wrong action reaches a guest." by Ivan de Murard, https://ivandemurard.com/agent-boundaries-aetherix.md (markdown). Method: Separating prediction, recommendation and execution, with constraints enforced in code between components rather than requested in a prompt. Task: - Inventory the decisions the application makes or recommends. - Classify each by impact, reversibility and required authority. - Locate the boundaries between prediction, recommendation and execution. - Distinguish constraints written as prompt instructions from constraints enforced in code. - Propose approval points, fallback behaviour and audit records where they are missing. Inputs I will provide (ask me for anything missing; do not assume you can access my repository, files or systems): - An inventory of the decisions the application makes or recommends - Who is accountable for each one - What it costs to undo each one - The current prompts and the code paths that act on outputs Deliverables: - A decision register - A list of prompt-only rules that should move into code - An enforceable architecture improvement plan, with one test per boundary Verification: For every rule marked "enforced", show the code location and a test that fails when the rule is violated. A rule that only exists in a prompt is not enforced. Structure your answer in five parts: Verified findings (with the file or evidence you checked), Assumptions, Proposed changes, Actual modifications (none unless I asked for them), Tests executed (with their output, or "none"). Do not bypass approvals and do not change production systems without my explicit authorisation.
Updates
This piece is revised when something it claims stops being true. Each revision says what changed and why.
- · Added a service blueprint: one illustrative breakfast, from forecast to manager decision, service and feedback. The note explained the architecture but not what it means for the people running the service, nor where authority changes hands and which outcomes can actually be measured.
- · Linked the two companion notes and the public repository where the text makes a claim they explain or evidence. Readers arriving here had no path to how the boundary is tested, or to why approval and outcome are different signals.
- · Added a reading card, a table separating the five jobs in the loop, and the scope of the public benchmark. The note argued for one boundary but blended five things a reader needs to tell apart: forecasting, recommendation, enforcement in code, human approval and measured outcome. The benchmark sentence now says what it ran on (30 restaurants, Kaggle Recruit data) and that no manager decision has been delivered on the forecast yet, as the repository states.
- · Corrected the drift example. The 'three consecutive misses' alert was narration in the demo script, not a production mechanism, and the post-shift error was reduced, not fully recovered.
- · Corrected what the note said about recorded replies, the CI forecast baseline and the drift example. An audit of the Aetherix repository found that only an exact "accept" or "reject", or a corrected figure, records a decision; that the forecast baseline in CI is computed on a synthetic series, not on realised covers; and that the corrupted export and the regime shift come from a simulated 30-day scenario. The note had said more than that.