The email landed at 4:47 on a Friday, and it had the shape every project manager learns to dread: a real client, a real deadline, and a folder of files that did not all belong in the same box. A pharmaceutical client needed forty-two documents localized into six languages by Tuesday. Most of the batch was exactly what their MT-first pipeline ate for breakfast: press releases, internal training slides, a corporate newsletter, a conference brochure, FAQ pages for the patient-support website. But buried in the folder, with the same naming convention and the same Tuesday deadline as everything else, sat three files that were not like the others: a patient information leaflet, the dosing section of a summary of product characteristics, and a clinical-trial consent amendment. The PM, who knew the pipeline well and the content less well, did the human thing under deadline pressure: she looked at the folder, saw a familiar client and a familiar volume, quoted the whole batch at machine-translation post-editing rates, and routed all forty-two files into the standard light-post-editing lane. The newsletter and the brochure came through fine. The leaflet did not. A contraindication came back fluent, lawyerly, and inverted, telling patients to do precisely the thing the source forbade, and it would have shipped if a post-editor three segments in had not stopped and refused to continue. This lesson is about the rule that should have caught those three files before the quote was ever sent: high-liability content routing, the policy that makes the high-stakes path automatic, by rule rather than by the guess of a tired human looking at a mixed folder on a Friday afternoon.

What Routing by Rule Actually Means

Start with the vocabulary, because the lesson turns on three precise terms and the relationships between them. Routing is the act of sending a file, or a part of a file, down a specific workflow lane the moment it arrives, before anyone opens a computer-assisted translation (CAT) tool. A pipeline has lanes, and routing decides which lane each piece of content enters. Full post-editing (full PE) is the heavier of the two post-editing efforts under the ISO 18587 spectrum: a qualified linguist edits the machine output against the source segment, the approved termbase, and the style guide until the result is accurate, terminologically correct, and indistinguishable in quality from a human translation, with every claim verified against the source rather than read for flow. Its lighter cousin, light post-editing, fixes only what blocks comprehension and leaves merely-awkward prose alone, which is cheap and fast and exactly wrong for high-stakes content. And MT-forbidden content is content where the machine must never produce the trusted draft at all: the correct first move is not to post-edit machine output but to refuse it and route the work to full human translation by a qualified specialist, or in the most absolute cases to a sworn or certified translator whose signature carries legal weight. Machine translation (MT) here means any system, neural engine or large language model (LLM), that converts source to target with no human writing the words.

Now the claim that the whole lesson defends. High-liability content (medical, legal, financial, life-safety) should be sent to full human translation or full post-editing by rule, not by guess. The difference between those two words is the difference between a pipeline that is safe by design and a pipeline that is safe only when everyone is awake, careful, and unhurried, which is to say a pipeline that is safe never, because real pipelines run on Fridays under deadline. A rule is a written policy that fires automatically on a property of the content: if a file is a drug label, it goes to full human translation, period, regardless of who quoted it or what the deadline says. A guess is a human at intake looking at a file and deciding, in the moment, with incomplete information and a quota to hit, whether this particular file feels risky enough to deserve the expensive lane. The previous lesson taught you to match post-editing effort to risk tier. This lesson is narrower and harder: it is about making the highest-risk match happen automatically, so that the match no longer depends on the judgment of whoever happens to be holding the file.

Routing by rule means a property of the content, not the mood of a human at intake, decides the workflow. If the file is a drug label, the rule sends it to full human translation before anyone has a chance to guess otherwise.

Why High-Liability Routing Is Its Own Skill

You might reasonably ask why this needs a separate lesson at all. You already know which content is high-liability, you know it warrants full PE or full human translation, so surely you just apply that knowledge file by file. The answer is that knowing the right tier and reliably reaching it are two different competencies, and the gap between them is where catastrophes live. A linguist can know with perfect clarity that a contraindication needs full human translation and still watch it ship through light PE, because the decision about that file was made by someone else, earlier, under different pressure, looking at a folder rather than at the contraindication. The knowledge was never the problem. The routing was. High-liability content routing is the engineering discipline of converting a thing you know into a thing the pipeline does, so that the right outcome no longer requires the right person to be paying attention at the right moment. That conversion, from individual judgment to automatic rule, is the entire substance of this lesson, and it is exactly the substance generic AI training never teaches, because generic training assumes the hard part is the judgment when in a real shop the hard part is making the judgment fire every single time.

Why a Defaults-to-Safe Rule Beats Per-File Judgment

The heart of this lesson is a single design principle, and it is worth slowing all the way down for, because it runs against the instinct most professionals bring to their work. The instinct says: a skilled person should look at each file and decide what it needs, because skilled judgment is better than a blunt rule. That instinct is correct in a calm room with one file and no clock. It is dangerously wrong in a pipeline. In a pipeline, per-file judgment fails not because the judges are unskilled but because of three forces that grind down even excellent judgment under production conditions, and the defaults-to-safe rule is the engineered answer to all three.

The Three Forces That Break Per-File Judgment

The first force is deadline pressure. Every guess at intake happens under a clock, and the clock pushes every borderline call in the same direction: toward the cheaper, faster lane. When a PM is unsure whether a file is high-liability, the deadline does not whisper "take the safe path." It whispers "you are behind, this probably fine, keep moving." Per-file judgment under deadline is not neutral judgment; it is judgment with a systematic bias toward the dangerous answer, because the cost of over-routing (a file that gets expensive treatment it maybe did not need) is felt immediately as a missed quota, while the cost of under-routing (a shipped contraindication error) is felt later, by someone else, or never, until the arbitration. Humans optimize for the cost they can feel.

The second force is volume and sameness. The Friday folder had forty-two files that looked alike: same client, same naming convention, same deadline, same icon in the file browser. Three of them were lethal and thirty-nine were harmless, and nothing in the folder's surface distinguished them. Per-file judgment assumes the judge will inspect each file deeply enough to catch the three, but volume defeats inspection. A human scanning forty-two files does not read forty-two files; they pattern-match on the folder, conclude "this is a standard client batch," and quote the whole thing as one. The three dangerous files inherit the workflow of the thirty-nine harmless ones not through any decision about them specifically but through the absence of a decision, which is the most common way high-liability content ends up in the wrong lane.

The third force is turnover and inconsistency. The person who routes intake today is not the person who routed it last month, and neither has read the same things or fears the same risks. One PM is an ex-translator who has seen a dosing error and routes anything medical to full translation reflexively; the next is a new coordinator who knows the rate card cold and the consequence catalogue not at all. Per-file judgment makes the safety of your pipeline a function of who happens to be at the desk, which means it is not a property of your pipeline at all. A client cannot audit "we route carefully when the right person is looking." They can audit a rule.

Per-file judgment fails under three production forces: a deadline that biases every borderline call toward the cheap lane, a volume that defeats inspection and lets dangerous files inherit the workflow of harmless ones, and a turnover that makes safety depend on who is at the desk. A rule is immune to all three.

What Defaults-to-Safe Actually Encodes

A defaults-to-safe rule inverts the burden of proof, and that inversion is the whole trick. Under per-file judgment, content is innocent until proven dangerous: it flows into the cheap lane unless someone affirmatively flags it as risky, which means every miss, every distraction, every borderline shrug ends in the cheap lane. Under a defaults-to-safe rule, content matching a high-liability signal is dangerous until proven safe: it is routed to full human translation or full PE automatically, and the only way it reaches a cheaper lane is for a qualified person to affirmatively certify that it does not actually carry the consequence the signal suggested. The default catches the silent case, the rushed case, the nobody-looked case. You are not relying on someone to notice the danger; you are relying on someone to notice the rare absence of danger, which is a much safer thing to rely on, because a missed safe-downgrade costs money and a missed danger-upgrade costs a life or a lawsuit. Set the default to the side where a miss is survivable. That sentence is the entire philosophy of high-liability routing, and everything else in this lesson is its application.

It helps to name what the rule is buying you in plain terms. It is buying you a pipeline whose safety does not degrade when conditions get bad, which is exactly when you need safety most. A skilled judge and a defaults-to-safe rule produce the same outcome on a calm Tuesday with one file. They diverge precisely on the Friday-afternoon mixed-folder deadline crunch, which is when the dangerous files actually arrive, because dangerous files do not announce themselves and do not wait for a convenient moment. The rule is not an insult to the judge's skill. It is a recognition that skill is a resource that runs out under load, and that the cases where it runs out are the exact cases that matter.

The Routing Policy That Makes the High-Stakes Path Automatic

A principle is worthless until it is written down as something the pipeline executes. So here is the policy, the actual artifact, expressed in the form a shop can adopt, audit, and hand to a new coordinator on their first day. A high-liability routing policy has four parts, and each part exists to close a specific gap that per-file judgment leaves open.

Part One: The Trigger List

The policy begins with a written, enumerated list of content types and signals that trigger high-liability routing, so that the trigger is a fact about the content rather than an opinion about it. The list is built from the five intake signals you already know, made concrete as named document types:

• The audience acts on it physically. Drug labels, patient information leaflets, summaries of product characteristics, instructions for use (IFUs), medical-device instructions, dosing information, hazard and safety warnings, evacuation procedures. A human swallows, injects, operates, or evacuates based on this text.
• It allocates legal rights or obligations. Contracts, indemnity and liability clauses, warranties, terms of service, informed-consent forms, employment agreements. The text says who must do what and who is liable.
• A regulator will read it. Marketing-authorization dossiers, clinical-trial protocols and amendments, financial prospectuses, annual reports and disclosures, patent filings. A government agency reviews, approves, or files it.
• The numbers are load-bearing financial figures. Prospectuses, audited financial statements, loan and credit agreements, tax filings, insurance policy terms. A transposed digit or an inverted condition moves money or misstates a position.
• A signature must certify it. Birth, marriage, and death certificates, court judgments, sworn statements, official academic records. The deliverable must carry a translator's sworn or certified attestation to be valid.

The trigger list is the thing a new coordinator can apply without judgment. They do not decide whether a financial prospectus is risky; they look it up, find "prospectus" on the list, and route accordingly. Judgment has been spent once, carefully, by the people who wrote the list, and then banked, so it does not have to be re-spent, badly, by everyone who touches intake afterward.

Part Two: The Default Assignment

For each trigger, the policy states the lane, and the assignment defaults to the safe side. Medical, legal, and certified content defaults to full human translation, with full post-editing permitted only where a domain specialist has reviewed the specific content and confirmed that a grounded, fully verified post-edit meets the same standard. Financial content with load-bearing figures defaults to full PE at minimum, escalating to full human translation for the most binding instruments. Sworn and certified content defaults to a qualified sworn or certified translator with no machine draft trusted on its surface, because the signature is the product and a machine cannot swear an oath. The policy names the lane in writing so that the routing decision is a lookup, not a deliberation, and so that any deviation from the default leaves a visible, reviewable trace rather than disappearing into a moment of unrecorded judgment.

Part Three: The Override With a Name On It

A rule with no escape valve becomes either ignored or absurd, so the policy includes a controlled override, and the design of the override is where most policies live or die. The override is not "anyone who disagrees can re-route." The override is: content that matches a high-liability trigger may be routed to a lighter lane only if a named, qualified person, a domain specialist or senior linguist, affirmatively certifies in writing that the content does not carry the consequence the trigger implies, and that certification is recorded. The asymmetry is deliberate and total. Routing up to the safe lane requires nothing, no approval, no justification, because over-protecting content is never the failure mode that hurts anyone. Routing down to the cheap lane requires a name, a qualification, a written reason, and a record, because under-protecting content is the failure mode that kills and sues. The override exists so the rule is not absurd; the friction on the override exists so the rule is not toothless. A downgrade that nobody is willing to sign their name to is a downgrade that should not happen.

Part Four: The Escalation Path

Finally, the policy names what happens when the rule was not applied at intake and the misclassification surfaces later, usually in the hands of a post-editor who opens a file and finds a contraindication where a marketing string was promised. The policy makes escalation a defined, expected, blameless step rather than an act of individual courage. The post-editor stops, flags the misclassification to the PM, and the file is pulled and re-routed to its correct lane, with the quote and deadline renegotiated as a consequence of the routing error rather than treated as fixed constraints the linguist must somehow honor. We will return to escalation in its own section, because it is the safety net under the rule and the place where the human in the loop does their most valuable work, but it belongs in the written policy so that the post-editor who escalates is following procedure, not sticking their neck out.

The routing policy has four parts: an enumerated trigger list that turns risk into a lookup, a default assignment to the safe lane, a controlled override that requires a named person to sign off any downgrade, and a blameless escalation path for misclassifications that slip through. The override's friction is asymmetric on purpose: routing up is free, routing down costs a signature.

Handling Content That Mixes High and Low Liability

The cleanest version of routing assumes each file belongs wholly to one tier, and reality refuses to cooperate. Real documents mix. A medical-device user manual is mostly harmless: unboxing instructions, a parts diagram, cleaning notes, a warranty registration card, and then, in the middle, three pages of contraindications, sterilization steps, and dosing-equivalent calibration figures where a single inverted instruction harms a patient. A financial annual report is mostly boilerplate and narrative, with a handful of pages of audited figures and forward-looking statements where a transposed number is a misstatement to a regulator. A software end-user license agreement is wrapped around an app that is otherwise pure low-stakes UI strings. The mixed file is not the exception; it is the normal case, and a routing policy that cannot handle it is a routing policy that fails on most of its real inputs.

There are two ways to handle the mix, and the choice between them is itself a risk decision. The first and simplest is to route the whole file by its highest-liability content. If any part of a file triggers a high-liability signal, the entire file goes to the high-liability lane. This is the conservative default, and for many files it is correct, because the cost of full-translating a few extra harmless pages is trivial next to the cost of one missed contraindication, and because keeping the file whole keeps the workflow simple and the accountability clear. The whole document gets the treatment its most dangerous sentence requires, which is exactly the defaults-to-safe principle applied at the file level.

When and How to Split a Mixed File

The second approach is to split the file by section into its component tiers, routing the contraindication pages to full human translation and the unboxing instructions to light PE. Splitting is more efficient and more dangerous, and it is only safe under specific conditions. It is worth doing when the file is large, the high-liability portion is a clearly bounded minority, and the split can be done structurally and verifiably rather than by a human eyeballing where the dangerous part starts and stops. The danger in splitting is precisely at the seams: the boundary between the safe section and the dangerous section is a judgment call, and a judgment call is the thing this whole lesson is trying to engineer away. If the split is wrong, if a contraindication sentence sits on the safe side of a boundary that a tired coordinator drew in the wrong place, the splitting has reintroduced the exact failure mode the rule was built to prevent, and reintroduced it invisibly.

So the rule for splitting is a rule about the seam. Split only when the boundary between tiers is structural, explicit, and verifiable, never when it is a matter of judgment about where the risk begins. A document with a clearly delimited "Safety Information" section that the source itself sets apart can be split on that structural boundary. A document where the dangerous content is interleaved sentence by sentence with the harmless content cannot be safely split and must be routed whole to the high-liability lane. And whoever draws the boundary inherits the same accountability as whoever routes a whole file: a misdrawn seam that lands a contraindication in light PE is a routing error of exactly the same severity as never routing it at all. When in doubt, do not split. The efficiency you gain by splitting a borderline file is never worth the consequence you risk by splitting it wrong, and the defaults-to-safe principle says the doubt itself is the answer: route it whole, route it safe.

A mixed file defaults to its highest-liability content: route the whole file to the safe lane unless the boundary between tiers is structural, explicit, and verifiable. The seam of a split is a judgment call, and a judgment call at the seam reintroduces the exact failure the rule exists to prevent.

Escalation When Intake Misclassifies

No rule fires perfectly. Trigger lists have gaps, new content types arrive that nobody anticipated, a file is mislabeled at the source, a coordinator on their second day applies the list wrong, and a high-liability file slips into a low-liability lane. The routing rule is the primary control, but it is not the only control, and a mature pipeline treats the post-editor at the file as the last and most important line of defense rather than as a passive executor of whatever lane the file arrived in. This is where the human in the loop earns their place, and it is worth being precise about the behavior the pipeline needs from them, because it runs against the grain of how production work usually feels.

The scenario is the one that opened this lesson, viewed from the linguist's chair. You accepted a file routed as light post-editing. You open it, and three or four segments in you realize you are not looking at a marketing FAQ; you are looking at a dosing instruction, or a consent risk paragraph, or an indemnity clause. The file's lane says light PE. The content says full human translation. The deadline says Tuesday. The quote, somewhere upstream, said cheap. Every signal in your environment, the open file, the assigned lane, the clock, the quote, presses you toward one behavior: keep post-editing, do it carefully, and trust that your care will be enough. That behavior is the trap, and the entire profession depends on you not falling into it.

What Escalation Actually Is and Why It Is Not Optional

Escalation is the deliberate act of stopping work, flagging to the project manager that the file has been misclassified and belongs in a higher tier or is MT-forbidden, and refusing to let the cheap workflow proceed until the routing is corrected. It is not raising a concern, hedging, or doing the work while noting a worry in the comments. It is a stop. The reason it must be a hard stop and not a soft flag is the silent critical error: a fluent inverted contraindication that you post-edit "carefully" in the wrong lane still ships looking perfect, because careful light post-editing of high-liability content is still light post-editing, still optimized for surface and speed, still structurally unable to deliver the character-by-character verification against the source that the content requires. You cannot rescue a misrouted file by trying harder inside the wrong workflow. The workflow itself is wrong, and only re-routing fixes it.

The thing that makes escalation hard is never the linguist's knowledge; it is the social and economic pressure around the act. Escalating feels like slowing down a job everyone wants finished, like creating a problem, like being the difficult vendor who reopens a settled quote. A pipeline that wants to be safe has to actively neutralize that pressure, which is why escalation belongs in the written policy as a defined, expected, blameless step. When escalation is procedure, the post-editor who stops is following the rules, not breaking the rhythm. When it is not, the post-editor who stops is a hero, and a pipeline that requires heroism to be safe is a pipeline that will eventually meet a linguist who is too tired, too new, or too economically squeezed to be a hero on a Friday, and that is the file that ships wrong.

When a misrouted high-liability file reaches you, the duty is to stop and escalate, not to post-edit it more carefully. You cannot rescue the wrong workflow by working harder inside it. A pipeline that depends on heroism to catch misrouting will eventually meet a linguist who cannot be a hero that day.

Closing the Loop: Escalation as Pipeline Feedback

Escalation protects the file in front of the linguist, but its higher value is as a feedback signal that improves the rule. Every escalation is evidence that the trigger list or its application has a gap, and a shop that treats escalations as data tightens its triage over time: the content type that was missed gets added to the trigger list, the coordinator who misapplied it gets retrained, the source-labeling problem that mislabeled the file gets fixed upstream. An escalation that merely rescues one file and is then forgotten is a half-used control. An escalation that is logged, counted, and fed back into the routing policy makes the next misclassification less likely. The linguist who escalates is therefore doing two jobs at once: saving the patient or the counterparty who would have received the inverted clause, and handing the pipeline the information it needs to stop putting that kind of content in the wrong lane. Accountability stays human throughout: "the engine wrote it" is never an answer when a Critical error ships, and the revised ISO 18587, the post-editing standard expanded to cover AI and LLM output and in DIS ballot with publication targeted for late 2025 into 2026, requires the post-editor to hold full professional-translator competence precisely so that recognizing and refusing a misrouted file is treated as core professional work, not optional vigilance.

A Worked Routing of a Regulated Batch

Abstract rules earn their keep on a concrete batch, so let us route the Friday folder properly, the way it should have been routed before the quote went out. The client is the pharmaceutical company, the deadline is Tuesday, and the folder holds forty-two files. The wrong move, the one that actually happened, was to look at the folder, recognize the client, and quote the whole thing as MTPE. The right move is to run every file against the trigger list before pricing anything, and let the rule sort them. Here is the batch sorted into its lanes.

The Low-Liability Lane: MT-Ready, Light PE

Most of the folder lands here, and that is the point: routing by rule does not make everything expensive, it makes the right things expensive and lets the rest fly. The corporate newsletter, the conference brochure, the internal training slides, the press releases, and the patient-support website FAQ pages all match no high-liability trigger. Nobody acts on them physically, they allocate no legal rights, no regulator files them, they carry no load-bearing financial figures, and no signature certifies them. They are high-volume, low-consequence content where MT does its best work and a light post-edit cleans it. These twenty-odd files get MTPE rates and the fast lane, exactly as the client expected for the batch as a whole. The rule has cost nothing here; it has simply confirmed that the cheap lane is correct for content that genuinely belongs in it.

The Elevated-Liability Lane: Full PE

A smaller set sits one tier up. The patient-support website includes a few pages that move from general encouragement into specific guidance about recognizing when a side effect warrants calling a doctor, and a corporate-responsibility report includes a section summarizing safety commitments. These do not rise to the regulated-document tier, but they touch health guidance and brand-critical claims where a meaning slip damages trust and could influence a patient's behavior. They trigger the "audience may act on it" signal weakly, enough to pull them out of the light lane. They go to full PE: MT drafts, a qualified linguist verifies every claim against the source and the termbase. The rule has correctly placed them above the newsletter and below the leaflet, which is precisely the discrimination per-file judgment fails to make under deadline.

The High-Liability Lane: Full Human Translation

Now the three files that started the fire. The patient information leaflet, the dosing section of the summary of product characteristics, and the clinical-trial consent amendment each trigger multiple high-liability signals at once. The leaflet and the dosing section are content the audience acts on physically, where a flipped negation or a corrupted dose number harms a patient, and the measured LLM error rates on exactly this material, roughly 59% on drug names, 60% on dates and times, and 66% on adverse events, all delivered in fluent prose, are the empirical reason the machine cannot be trusted to draft them. The consent amendment triggers three signals simultaneously: a regulator reads it, it carries legal weight, and a participant acts on it physically by agreeing to bear described risks. All three default to full human translation by a qualified medical translator, and the consent amendment, sitting at the intersection of regulator, law, and physical action, is treated as MT-forbidden in the strict sense, with no machine draft trusted on its surface and an independent review step because the asymmetry between fluent and correct is too dangerous to leave to a single pass. These three files were always going to cost full-translation rates and take longer than Tuesday. The routing rule's job was to surface that fact on Friday at intake, while there was still time to renegotiate the deadline and the quote, instead of on Monday when a post-editor found the inverted contraindication three segments in.

The Mixed File in the Batch

One file in the folder is the hard case: a device user manual that accompanies a combination drug-device product. It is forty pages, and thirty-five of them are unboxing, assembly, cleaning, and storage instructions that nobody's life depends on. Five pages, set apart in the source under a clearly headed "Important Safety Information and Dosing" section, contain contraindications, injection-technique steps, and dose-calculation figures. This is the textbook mixed file, and the policy handles it cleanly because the boundary is structural: the source itself delimits the safety section with its own heading. The thirty-five instructional pages route to full PE (they are device instructions, elevated but not life-critical), and the five-page safety section, split on the source's own explicit structural boundary, routes to full human translation by a medical specialist. Had the contraindications been scattered sentence by sentence through the assembly instructions with no structural marker, the file would have routed whole to full human translation, because there would have been no verifiable seam to split on and the defaults-to-safe rule would have governed the doubt.

What the Worked Batch Proves

Step back and look at what the rule did to the Friday folder. It did not make everything expensive, the panic reaction a client fears. It made about twenty files cheap, a handful full-PE, three full-human-translation, and one split on a verifiable structural seam. It produced a quote with four line items instead of one, each defensible, each tied to a content property rather than a guess, each auditable by the client and a certifier. It surfaced the Tuesday-is-impossible-for-the-leaflets problem on Friday, when it was a renegotiation, rather than on Monday, when it would have been a crisis. And critically, it did all of this without depending on the PM to feel which files were dangerous, because the trigger list felt the danger for her. That is the entire promise of high-liability content routing: the safe outcome stops being a thing a careful human achieves on a good day and becomes a thing the pipeline produces on every day, including Friday.

Key Takeaways

• High-liability content routing sends medical, legal, financial, and life-safety content to full human translation or full post-editing by rule, not by guess: a written policy that fires on a property of the content, so the safe outcome no longer depends on the right person paying attention at the right moment.
• Knowing the correct tier and reliably reaching it are different competencies; routing is the engineering discipline that converts a thing you know into a thing the pipeline does automatically, closing the gap between correct knowledge and correct outcome where catastrophes live.
• Per-file judgment breaks under three production forces: deadline pressure biases every borderline call toward the cheap lane, volume and sameness let dangerous files inherit the workflow of harmless ones in a mixed folder, and turnover makes safety depend on who is at the desk. A defaults-to-safe rule is immune to all three.
• Defaults-to-safe inverts the burden of proof: content matching a high-liability signal is dangerous until proven safe and is routed to the safe lane automatically, so a miss lands on the survivable side (over-protected content costs money) rather than the catastrophic side (under-protected content costs a life or a lawsuit). Set the default where a miss is survivable.
• The routing policy has four parts: an enumerated trigger list that turns risk into a lookup, a default assignment to the safe lane, a controlled override where any downgrade requires a named qualified person's signed certification (routing up is free, routing down costs a signature), and a blameless escalation path for misclassifications.
• Mixed files default to their highest-liability content: route the whole file to the safe lane unless the boundary between tiers is structural, explicit, and verifiable. The seam of a split is itself a judgment call, and a misdrawn seam reintroduces the exact failure the rule prevents, so when in doubt, route whole and route safe.
• When a misrouted high-liability file reaches a post-editor mid-workflow, the duty is to stop and escalate, not post-edit more carefully: you cannot rescue the wrong workflow by working harder inside it, and careful light PE of a contraindication still ships the inverted clause looking perfect. Escalation must be a blameless written procedure, and every escalation is feedback that tightens the trigger list.
• Routing a regulated batch by rule does not make everything expensive; it makes the right things expensive: it surfaces the impossible-deadline problem at intake on Friday when it is a renegotiation, produces a defensible multi-line quote tied to content properties, and keeps accountability human under the revised ISO 18587, which requires the post-editor to hold full professional-translator competence precisely so misrouting can be recognized and refused.