The pack is still open at the freeze

The problem: packs the freeze cannot close

A process plant is heading into a turnaround. Before the freeze date, every work order has to name the procedure revision in force, the parts that revision lists, and the isolations a permit will later need. The asset system, the warehouse, and the procedure file each hold one of those facts. The pack is still open because they are not on the same order.

Some jobs are already coded, and a rule can finish them. The jobs that consume the planners are the other kind. The revision is blank or superseded, or the materials line says “seal kit” with no part number. A planner finishes that job by reading the defect history and the procedure, then checking the parts against the warehouse. Figure 1 is one. At the freeze, such a pack is still open or has been marked ready from memory, and the wrong revision and the missing spare go into the window with it.

Four terms carry the case. A coded job carries the revision in force and a reservable bill of materials. An uncoded job does not, or the revision it names is not the one in force. The span is the exact section that names the isolations and the parts: section 6 and the parts table of revision 4, here. The freeze is the superintendent’s acceptance of the pack. The freeze date is only the deadline.

The scenario assumes a process manufacturer with four plants. Each takes a major turnaround every four years, staggered, so the company runs one event a year. One event holds 1,200 work orders, on a planning split of 900 coded and 300 uncoded. Reading an uncoded job takes a median of 4 hours: about 1,200 hours an event, or about 72,000 in the group’s reporting currency at an assumed fully loaded rate of 60 an hour. These are planning figures, not measurements. If a real count finds fewer than 100 uncoded jobs, the answer is the freeze rule and a senior planner, and nothing else here is built.

The director already prices a day of extension. This case does not borrow that figure until the pilot has counted packs still open at the freeze. The outcome that would count is cost and risk: reading hours returned, coded jobs that meet the freeze, and no pack frozen on a superseded revision or on a part shown as held when the balance is zero.

Who decides, and who can stop it

Sponsors. The turnaround director decides the duration and what may sit on the critical path. The maintenance superintendent decides what must be true before a pack is frozen. A design that only returns planners’ hours will be stopped by one of them.

Operators. Planners work the uncoded queue. They need the span to open on the page it names, and the warehouse balance beside the part number rather than in a sentence. A senior planner takes the uncertain drafts, the jobs where no single procedure applies, and any balance the warehouse did not return, so that queue has to stay small enough to read properly. The permit authority, the warehouse, and the crew feel a pack frozen on the wrong revision.

Can block. Process safety, if a draft isolation can be treated as a permit. Security, if a procedure leaves the plant network or reaches a tool the company has not approved. The owner of the asset system, if a second store becomes the pack. The warehouse, if a model may reserve stock or stands in for buying the part that is not held.

Where they conflict. Planners want the hours back. The superintendent wants the span open before they accept. Process safety wants a competent person on every isolation. The director wants the freeze date to hold. A turnaround contractor will plan the event in their own tool if they are given the procedures. When speed conflicts with process safety or security, those two win: a wrong isolation, or a procedure that has left the plant, costs more than another day of reading. The asset-system owner wins against any second pack of record. The contractor wins only if the company refuses the connection and still wants more people for the peak.

What planning must be able to do

Functions. Every work order in the window says which kind it is. A coded job cannot be marked ready while its revision or its bill of materials is missing. An uncoded job gets a draft: the procedure revision in force, the cited span, the isolations and part numbers written in that span, and the balances code has read from the warehouse. The superintendent opens the span, then accepts, edits, or returns the draft, and accept stays off until the span has been opened. The permit authority authorises the isolations afterwards, in the permit system, from a pack a person has frozen.

Controls. A pack is frozen only in the asset system, by the superintendent. A permit is issued only in the permit-to-work system, by the permit authority. A reservation is created only by the warehouse transaction the plant already uses. The assistant drafts. It does not freeze, permit, reserve, buy, or move the schedule. Procedures and stock are processed only inside the plant network and the company’s own tenant, over a private connection, and users sign in with the corporate directory. If the assistant is down in the week of the freeze, the uncoded queue waits; it never falls through to a pack marked ready with a blank revision. The span and the superintendent’s action are kept as long as the work order.

What stays. The asset system, the warehouse, the permit system, the procedure store, and the directory. The revision in force becomes a field document control owns, so a superseded file is not treated as current at the next event.

Data. Work-order text, asset identity, and defect history are confidential maintenance records, needed to know which procedure applies. The procedure is document control’s, confidential, and needed only for an uncoded job. The balance says whether a named part is held. Contractor rates, competency and medical records, and control logic do not change the span, so no model sees them, and neither does the planner’s mail with the warehouse.

Does this need a model at all?

Five candidates sit in front of this event, from doing nothing to an assistant that authorises the permit when a score is high.

Each is scored from 1 to 5, and each column says which end is good. I do not add the columns up: a high value score does not cancel a risk process safety will not carry.

Doing nothing keeps the hours, the stale revisions, and the public paste. The freeze rule fits best, since a field the asset system can hold should be required, but it reads no procedure, so its value is middling. More readers can open the file share today, which is why their data score is high, but another reader does not make “in force” a field, and the contractor’s tool takes a copy of every procedure. The combined path covers both kinds of job; the connection, the revision data, and the evaluation put its complexity at the top of the scale. An assistant that authorises permits turns a draft isolation into authorised work.

So the model has one narrow job. Refusing an empty field, reserving a part the warehouse holds, and buying one it does not hold need no model. Keyword search does not finish the rest either: the isolations and parts sit inside a section, the revision in force is not the file saved last, and the balance is not in the prose. A model earns its place in finding the right section and naming what it says. Code reads the balance. The model has no place in the permit, the reservation, the purchase, or the schedule.

A model that authorises a permit has answered a different question from the one process safety asked.

Planning today

The plant is not a blank sheet, and the work order is not the broken part. The asset system holds the order, the asset, and the defect history. The warehouse holds stock and reservations. Permit-to-work is a separate plant system, and it is not where the pack is assembled. Everyone signs in through the corporate directory. There is no approved path from a cloud tenant to any of these systems, because none has been needed.

Procedures. Controlled procedures are files on a store inside the plant network. A revision, a superseded file, and a scanned copy of the same pump overhaul can sit in three folders. “In force” is a date inside the file, or a folder name, not a field the work order can check. The order can therefore carry revision 3, looking filled in, after revision 4 has removed an isolation.

Mail and notes. A planner confirms a part with the warehouse in a thread that never reaches the work order. When the pack is challenged, the answer comes back as a sentence with no section, typed into a free-text note. Each planner keeps a note of which revision a class of equipment “actually” needs. That note is the only shared definition, and it leaves when the planner does.

Unsanctioned tools. Some planners paste section 6 and the defect text into a public chat tool and get a parts list back. The list has no revision and no warehouse balance. Nobody approved the paste, and nobody can say afterwards which procedure left the plant.

Three things are missing: a split between coded and uncoded, an owner who treats the revision in force as data rather than a folder, and a stored span. Until one event is classified, the 900 and the 300 remain a planning split, not a count.

Planning after the change

The target closes each gap above and leaves the permit system alone. The assistant never calls it.

Owners. The superintendent owns the freeze. Document control owns which revision is in force. Process safety owns the permit, unchanged. Security owns what may pass through the gateway. Planning runs the queue. The integration team runs the connection, the reads, and the draft, on a runbook the superintendent can read.

The gateway. The gateway is the only route to the model. For each uncoded job it builds the packet: the work-order text, the asset, and its defect history. The assistant has one tool, a search over the procedure revisions in force for that equipment class, and the gateway runs that search itself, inside the planner’s plant scope. Superseded revisions are not in the index, so they cannot be retrieved. The assistant returns a procedure, a span, the isolations, and the part numbers, or says that no single procedure applies. Code then confirms that the span appears word for word in the revision in force, reads the warehouse once for those part numbers, and writes the balances itself. A sentence from the model about stock is discarded.

Measures. Quality is the share of drafts whose span was opened and accepted unchanged. Cost is attributed per uncoded pack, where the calls are the small number and the private connection the large one.

Four ways to finish an uncoded pack

The freeze rule is already chosen for coded jobs. What remains is how to read the procedures that are not yet fields. Four options were real, each priced against the 300 uncoded jobs rather than all 1,200 work orders.

A stricter form on the work order

The asset system can refuse to mark a job ready while the revision or the bill of materials is empty. It is cheap, nothing moves, and planners stay on the screen they know. That is the freeze rule, and it belongs on every coded job. It does not read. A planner can type revision 3 from memory and the form will accept it. As the whole answer, it leaves the 300 uncoded jobs in place, now looking complete.

A specialist turnaround suite

In this scenario, the products a sponsor is shown take the procedures and bring their own pack. The objection is the pricing base assumed in the offer, a licence on the whole event to read 300 jobs, and the procedures held in the vendor’s cloud. The price is an assumption about the offer, so a quote comes before anyone rejects it. I would reopen it only for a deployment in the company’s tenant, with the pack left on the work order.

A narrow assistant beside the asset system

This is the path in Figure 3: the gateway builds the packet, the assistant cites the span, code checks it and writes the balances, and the pack freezes only when the superintendent accepts. The assistant runs on Microsoft Foundry, in the company’s Azure tenant. The alternative is an agent loop the company would build, host, patch, and evaluate itself. The packet, the checks, and the connection are owed either way; Foundry removes the part nobody at a plant wants to staff. The lock-in is the agent runtime. The index, the spans, and the evidence stay in the company’s own stores, so the model can be swapped.

The build, as a planning assumption, is three engineers for two quarters, because the plant systems are on premises and there is no private path today. The calls, at an assumed 5 currency units a pack, come to about 1,500 a year. Neither figure is a payback, and a quote replaces both.

The assistant authorises the permit

The same assistant can be given a right in the permit system, so a confident isolation authorises itself. Process safety would then sample every such permit, which recreates the permit authority under another name, and switching it off means admitting isolations went ahead without one. A sponsor will be offered this option. The failure process safety fears is its ordinary bad day.

The decisions that are expensive to reverse

Decision 1: The work order remains the only pack of record

A second place where “ready” can be true is expensive to unwind a year later, when nobody can say which pack the window ran.

Decision 2: Coded jobs stay on the rule; the model reads only what is still uncoded

If the next event’s pump overhauls are still free text, the assistant has become a permanent planner.

Decision 3: The assistant runs in the tenant, on a narrow packet, over a private connection

I would delay the pilot rather than send a procedure to an endpoint that keeps prompts or trains on them.

Decision 4: A person freezes every uncoded pack, and a person authorises every isolation

A confident citation of revision 3 is the failure that hurts. A hesitant draft left with a senior planner costs minutes, and the design spends the minutes.

A pattern for packs a procedure has to finish

The pattern is for an operation that already stores the job and still cannot finish it, because the steps and the parts live in a controlled document the job does not point at, and the stock lives in another system on the same premises. A turnaround work order is this case. The same shape appears wherever a planner leaves the system of record to find a section, then types a sentence back into a note.

It is the wrong pattern when the job record itself does not work, when most jobs are already coded, or when the organisation wants the model to authorise, reserve, or buy. Fix the fields first. A small remainder needs a senior planner, not a connection.

Data flow. The asset system raises the job (step 1), and the freeze rule keeps coded jobs where they are (step 2). An uncoded job goes to the gateway, which builds the packet (step 3). The assistant searches the revisions in force, cites a span, and names the part numbers in it (step 4). Code checks the span against the in-force text, reads the warehouse, and writes the balances; a draft that fails either check is not offered as ready (step 5). A person opens the span and freezes, or returns the job (step 6). The reservation and the permit follow on the plant side, from the frozen pack.

Identity and data. Only someone who already plans that asset can invoke the assistant for it. The gateway sets the search scope from that person’s groups, and the model cannot widen it. Nothing outside the packet reaches the model, the stock is attached by code after the call, and run logs are kept in the same class as the procedure.

Trade-offs and conditions. The pattern accepts a slower freeze than an automatic permit, a permanent human step on the uncoded path, and upfront work on the revisions and the private reads. Use it when coded jobs already flow, a procedure explains the rest, stock is a fact in another system, and a wrong automatic isolation is unacceptable.

The same pattern, on the platform chosen

Figure 5 builds it on Azure and Microsoft Foundry, with the step numbers of Figure 4.

The plant and the door. The asset system publishes reads, draft fields, and the freeze through API Management; the warehouse publishes a balance read and nothing else. Permit-to-work publishes nothing. A self-hosted integration runtime copies the revisions in force out each night. All of it crosses one site-to-site VPN into a hub with Azure Firewall and no route to the internet. At a few thousand reads a year, ExpressRoute would cost more than it carries, unless the company already runs one.

The spoke. The review panel runs on App Service behind a private endpoint, so planners reach it only from the plant. API Management is the single front door, and its freeze operation accepts only a superintendent’s token. The orchestrator, on Azure Functions, builds the packet, runs the agent’s one search inside the planner’s scope, checks the span, reads the stock, and writes the draft and the evidence.

AI and data. Document Intelligence turns each revision, scans included, into sections, tables, and page positions. AI Search indexes only the revisions in force, tagged by plant and equipment class. Foundry Agent Service runs one agent on a regional model deployment and keeps its threads in the company’s own Cosmos DB, beside the evidence. Prompts are not used for training, and the pilot waits until the exemption from abuse-monitoring storage is granted. Entra ID carries the plant groups, Azure Policy denies public endpoints, and Purview blocks the paste into public tools.

Cost. The VPN gateway, the firewall, the search service, and the panel bill every hour of the year; the model runs about 300 times. The quote has to price both, and a shared hub changes it most.

What is funded first

Value arrives twice: for coded jobs before any model is switched on, and for one procedure family in the pilot. Figure 6 shows the sequence and the two points where the work may stop; the months are a planning assumption.

Mechanical pump overhauls go first because their span names both isolations and parts, and code can check the balance, as in Figure 1. A procedure written as judgement, with no part numbers to read, waits. The share of the 300 that are pump overhauls is not assumed; the foundation counts it.

What may be read, and who still freezes

Who may use what. Planners may use the assistant inside the planning queue, and only on uncoded jobs. The superintendent sees the draft on the jobs they freeze. The permit authority does not use the assistant: they receive the isolations on a pack a person has frozen, and authorise in their own system. Contractors do not get the assistant. No one may paste a procedure, a defect history, or a stock position into a public tool, and the browser blocks the attempt.

What may enter the model. The packet: the work-order text, the asset, its defect history, and what the scoped search returns. Never a contractor rate, a competency or medical record, control logic, another plant’s procedure, or a stock quantity. Nothing trains a provider’s model, and the model writes nothing back.

Where a person decides. Accept: the superintendent has opened the span and accepts the isolations named there and the balances code wrote. The control stays disabled until the span is opened. Edit: they change the span, a part, or an isolation; the change is stored, and the balance is read again before accept. Return: a senior planner keeps the job. The permit authority then authorises or refuses each isolation in the permit system. The model never drafts the permit and never closes a refusal.

A confidence score sorts the queue. It does not freeze the pack.

Testing and monitoring. A procedure family goes live only after passing the evaluation below. Then the superintendent reviews twenty frozen packs a month against the cited span, scanned procedures included, and security reviews the packet definition twice a year and after any change. Process safety can open, for any pilot permit, the isolations on the frozen pack and who authorised them. The family stops if a wrong asset, a superseded revision, an isolation outside the span, or a zero balance shown as held reaches ready to freeze.

Bias. The model does not rank people or contractors, so there are no personal traits to examine. The miss that matters is systematic, on scanned procedures or one equipment class, where packs stay wrong while an unstratified sample looks fine. That family returns to the planners until the documents are fixed. The superintendent owns this check.

A separate permit policy is unnecessary: the assistant is not on that path, and the plant’s permit rules still apply.

What a model must pass before it runs

The evaluation is for the pump-overhaul assistant, against the requirements above rather than a leaderboard. It has three stages: eligibility, hard gates, then weights among the survivors.

Candidates. A general-purpose model on Foundry in the company’s tenant, behind the gateway. A specialist turnaround vendor that holds the procedures and wants to sit in the planning path. A model fine-tuned on the plant’s own past packs. A candidate that cannot run in the tenant, or cannot leave the procedures in the company’s store, is ineligible before any score.

The test. Mechanical pump overhauls, the first family in the roadmap. If the last event marked fewer than 40, the family waits. Once it clears 40, the working set is 40 packs with the revision, the isolations, and the parts marked by a senior planner, 10 of them scanned or poorly structured. Beside those, and not part of the 40, are four sets of 10 traps: a wrong asset, an order that names a superseded revision, an isolation that exists only in the old revision, and a part the warehouse does not hold. The last set tests the whole path rather than the model: the draft must show the part as not held.

The gates. Any trap presented as ready to freeze disqualifies the candidate. So does an invented procedure on a job where none applies. The candidate must cite the marked revision on at least 34 of the 40, and on at least 8 of the 10 scanned, so a pass on clean files alone still fails. The survivors are scored on these weights.

What must never ship. A draft presented as ready that cites the wrong asset or a superseded revision, names an isolation the span does not contain, or contradicts the warehouse read. An incomplete draft returned to the planner is an acceptable failure: it costs the reading it was meant to shorten, and it authorises nothing.

Who it escalates to. Every uncoded pack reaches the superintendent, and every isolation reaches the permit authority. Low confidence, no procedure found, a tool that did not return, or a failed citation check goes to a senior planner. The traps run again in Foundry before each new family and after any change of model. A fine-tune does not skip them, and waits for a store of confirmed spans.

The case a sponsor can fund

Fund the freeze rule now. Count the uncoded jobs on one event before anyone builds a connection. Under 100, staff a senior planner for the window and stop. Over 100, fund the connection and a pilot on mechanical pump overhauls, live only after that family refuses the traps. A person still freezes every uncoded pack, and the permit authority still authorises every isolation.

The cost of today. On the planning assumptions, uncoded reading is about 1,200 hours an event, about 72,000 at a rate of 60. Coded jobs get no second invented total, and the day-rate for a slipped window stays the director’s figure.

Investment and return. The freeze rule is configuration, and it returns coded jobs that hold the freeze before any model runs. The calls, about 1,500 a year, are not the investment. The investment is the connection, the revision data, and the panel: three engineers for two quarters as a planning assumption, plus the run cost of the gateway, the firewall, and the search service. One event’s 72,000 does not pay that. The hours carry it only if the next event reuses the same packet, so until the quote exists there is no payback claim.

Risk and the change in the work. The risk worth naming is a superseded revision, or a seal face shown as held against a zero balance, presented as ready to freeze; the traps and the disabled accept control exist for that day. Planners stop keeping private notes and pasting section 6 into public tools. They do not stop owning the pack, and the permit authority’s work does not move.

Measures. Coded jobs meeting the freeze rule move in the freeze-rule phase. The pilot family’s cited drafts, planner hours, late spares and isolations, and opened spans move in the pilot. The decision in front of the sponsor is whether to fund that sequence, including the right to stop when the count does not justify the connection.