The file that nearly shipped was a Japanese-to-English set of warranty terms for a consumer electronics maker, forty-one segments, machine-translated by the engine and pre-populated in the translation-management system before anyone opened it, and it read like a lawyer had written it. Hana, the reviewer on the account, was not the post-editor; she was the second pair of eyes the pipeline put between the post-edited file and the client, and her job was not to re-edit but to verify. She had thirty minutes. The post-editor had done careful work: the prose was clean, the register was right, the terms looked correct, and on a first read the file was a pleasure. Then she did the thing the pipeline existed to make her do, which was to stop reading the English as English and start reading it against everything else the project knew. She put the target beside the source, then beside the translation memory, then beside the termbase, and she read the four together. In segment 19, the English warranty granted the customer a right to a replacement unit "at the manufacturer's sole discretion, including for cosmetic damage," and that clause, fluent and lawyerly and entirely plausible, did not exist in the Japanese source. The engine had invented it. In segment 27, the Japanese carried a sentence excluding water damage from coverage, and the English simply did not have it; the sentence had evaporated, and the paragraph closed over the gap without a seam. One invented clause that handed the customer a right the manufacturer never granted, and one dropped sentence that removed an exclusion the manufacturer depended on. Both fluent. Both invisible to anyone reading the English alone. Both caught, in the same pass, by the same method, because Hana was not reading the translation. She was cross-referencing it. This lesson is that method, slowed all the way down and built into a pipeline so it runs on every file, not only on the days a reviewer happens to be suspicious.

The Two Failures That Share One Cure

Two distinct things went wrong in Hana's file, and the discipline of this lesson begins with naming them precisely, because they are mirror images and the cure for both is the same single act. A hallucination is content the engine added that has no basis in the source: an invented clause, a fabricated number, an explanatory phrase the model generated out of its general knowledge of how warranties read, material that is in the target and in nothing the source actually said. An omission is the opposite: meaning that was present in the source and is absent from the target, a dropped sentence, a vanished qualifier, an excluded condition that the fluent target closes over so smoothly you would never know a gap had been there. One is addition without a source. The other is source without a target. They are the two ways a translation can fail to be a faithful map of its original while still reading perfectly, and in 2026, with an LLM (a large language model, the kind of generative engine that drafts fluent translations and is also the kind most prone to confidently inventing) sitting at the front of nearly every pipeline, they are no longer occasional accidents. They are the engine's characteristic failure mode, because an LLM is built to produce plausible continuations, and a plausible continuation is exactly a sentence that reads like it belongs whether or not the source warranted it.

Here is the property that makes both failures dangerous and unites their cure. Neither one is visible from inside the target. Read the English of segment 19 on its own and it is a coherent, well-formed warranty clause; nothing about it announces that it is fiction, because the engine that invented it is a fluency machine and fiction is what it is good at. Read the English of segment 27 on its own and the paragraph is complete, balanced, and internally consistent; nothing announces that a sentence is missing, because the engine rewrote the surrounding prose to flow over the absence. The target cannot betray either error, because in both cases the target is fluent and self-consistent. The error exists only in the relationship between the target and the source, and a relationship is not something you can see by looking at one of its two ends. You have to look at both, together, and compare. That act of comparing the target against the things outside it that define what it should contain is called cross-referencing, and it is the only instrument that detects a failure which leaves no trace inside the target itself.

A hallucination is target without source. An omission is source without target. Neither is visible from inside the target, because the target is fluent in both cases. The error lives in the relationship, and you cannot inspect a relationship by looking at one end of it.

This is why the lesson at L1 introduced hallucinations and omissions as a single pair rather than two separate problems, and why this L3 lesson treats them together at pipeline scale. They are not two skills. They are one skill, applied in two directions: you walk from source to target asking "did everything here arrive?" and that direction catches omissions, and you walk from target to source asking "did everything here come from somewhere?" and that direction catches hallucinations. Same act, two bearings. A reviewer who runs only one direction catches only one of the two failures, and the file that nearly shipped on Hana's desk would have failed differently depending on which direction she skipped. Run source-to-target only, and you catch the dropped water-damage exclusion but miss the invented discretion clause. Run target-to-source only, and you catch the invented clause but miss the dropped exclusion. The method is two-directional by necessity, not by thoroughness, because the two failures are genuinely opposite and each direction is blind to the other's.

The Triangulation Method: Four References, Not One

The naive version of cross-referencing is to put the target beside the source and compare the two, and for many files that is most of the value. But a localization pipeline carries more than a source. It carries a translation memory (the TM, a database of previously translated and approved source-target segment pairs that the project has accumulated over time) and a termbase (the database of the client's approved term for each concept, the mandated word that must appear regardless of what fluent synonym the engine prefers). These are not decorations. They are additional witnesses to what the target should contain, and the difference between an amateur cross-reference and a professional one is how many witnesses you consult. This lesson calls the full method triangulation, borrowing the word from surveying and navigation, where you fix a single unknown position by taking bearings from several known points, and no single bearing is trusted alone.

The four points of the triangulation are the target, the source, the TM, and the termbase, and each answers a different question about the segment in front of you.

• The source answers the foundational question: what meaning is supposed to be here? It is the ground truth for content. Every unit of meaning in the target should trace back to the source, and every unit of meaning in the source should arrive in the target. The source is the reference against which both hallucination and omission are defined, and it is the one reference you can never skip.
• The translation memory answers: how has this, or something near this, been rendered before, in approved work? When a source segment matches or fuzzily matches a TM entry, the TM is a witness to what a human-approved target looked like, and a fresh MT rendering that diverges sharply from a high-match TM entry is a signal worth checking, because either the engine improved on the past (rare) or it drifted from it (common). The TM is especially powerful for catching omissions, because a TM entry shows you the full prior target, and a sentence present in the TM target but absent from the new MT target is an omission with a paper trail.
• The termbase answers: which exact words are mandated here, regardless of fluency? It is a witness to required vocabulary. A termbase does not usually catch a dropped sentence, but it catches a particular and underrated kind of hallucination: the engine substituting an approved term for a fluent synonym, which is a small content fabrication, the target now asserting a concept in words the client did not approve. And cross-referencing a target against the termbase can surface omission of mandated content too, when an approved term that should appear in a segment is simply absent.
• The target is the thing under examination, the unknown position you are fixing. It is never a reference for itself; the entire failure mode of fluent MT is that the target looks correct from inside, so trusting the target to validate the target is the precise error the method exists to prevent.

Triangulation matters because no single reference catches everything, and worse, each reference has blind spots that another reference covers. The source alone tells you what should be there but not which exact words the client approved; the termbase tells you the approved words but not whether a whole sentence dropped; the TM tells you what the approved past target looked like but only for segments that match the TM. A reviewer working from the source alone will catch the invented clause and the dropped sentence, which is most of the value, but will miss the engine quietly swapping the client's mandated term "Authorized Service Center" for the fluent-but-wrong "official repair shop," a fabrication of approved meaning that only the termbase reveals. Reading all four together is what closes the gaps, because where one reference is blind, another is watching.

One reference catches some failures and is blind to others. Four references, read together, leave the target nowhere to hide, because where the source is silent the termbase speaks, and where the termbase is silent the TM does.

Why the TM and Termbase Earn Their Place

It is tempting, under deadline, to collapse triangulation back to a source-target check and skip the TM and termbase as luxuries, and that temptation should be resisted on the high-consequence files for a concrete reason: the TM and termbase are the only references that encode prior human approval, and prior human approval is exactly what an MT-first pipeline is at risk of quietly discarding. Consider the dropped water-damage exclusion in Hana's segment 27. If that warranty had been translated before, an earlier version carrying the exclusion would likely sit in the TM as an approved segment. When the new MT rendering arrives without the exclusion, the divergence from the TM entry is not a stylistic difference; it is a flag that a previously approved unit of meaning has gone missing. The source catches the omission by itself, yes, but the TM catches it twice, and a reviewer scanning a forty-one-segment file under a thirty-minute deadline benefits enormously from a second, independent witness that raises a flag on the exact segment where the meaning changed. The TM turns a needle-in-a-haystack source comparison into a targeted check, because it tells you where to look hardest: the segments where the MT and the approved past disagree.

The termbase earns its place by catching the failure the source comparison structurally cannot. When you compare target to source for omissions and additions, you are comparing units of meaning, and a synonym substitution does not change a unit of meaning in the way an addition or omission does; "official repair shop" carries roughly the meaning of "Authorized Service Center," so a source-target meaning map passes it as faithful. But in terminology-controlled work, the approved term is not a stylistic preference; it is a contractual or regulatory requirement, and an unapproved synonym is a defect even when it is meaning-preserving, because consistency of the exact term is what a regulator, a search index, a downstream TM, or a legal interpretation depends on. The termbase is the only one of the four references that can flag this, because it is the only one that asserts not just "this meaning should be here" but "this exact word must carry it." Skipping the termbase means accepting that a class of fabrication, the substitution of approved language with fluent alternatives, will pass your gate undetected.

Detecting Additions: Walking Target to Source

The discipline for catching hallucinations is a directional walk: you traverse the target and demand that every unit of meaning in it produce a source. The metaphor that keeps this honest is a customs inspection. Every item in the target's luggage must declare where it came from; anything that cannot name its source segment is contraband, an addition the engine smuggled in. The reason this must be a deliberate, target-to-source walk rather than a casual read is that additions are, by their nature, things that read as if they belong. The engine does not hallucinate gibberish; it hallucinates plausible content, content that fits the genre, the register, and the surrounding sentences so well that it feels native to the document. Your reading brain, which is trained to extract meaning and move on, will accept it without protest. Only when you stop and ask of each clause "which part of the source produced this?" does the addition fail to answer, and the failure to answer is the detection.

The units to walk are not whole segments but the meaning-bearing pieces within them: clauses, qualifiers, conditions, numbers, named entities, cross-references. For each, you ask the customs question. In Hana's segment 19, the target carried "at the manufacturer's sole discretion, including for cosmetic damage." Walked as units: "the customer receives a replacement unit" traces to the Japanese source, present; "at the manufacturer's sole discretion" does not trace to anything in the Japanese, contraband; "including for cosmetic damage" does not trace either, contraband. Two clauses with no source, both fluent, both consequential, because together they restructure the warranty's obligation, one adding a manufacturer escape hatch and the other appearing to expand coverage, and a customer or a regulator reading the English would take both as the manufacturer's word. The walk does not require you to be a brilliant lawyer. It requires you to be a strict customs officer who lets nothing through without a source.

The Tells of an Addition

Additions announce themselves to a trained reader through a small set of recurring tells, and knowing the tells lets you walk faster, because they tell you where to slow down. The first tell is specificity the source did not have: the engine sharpening a vague source into a precise target, "store appropriately" becoming "store below 25 degrees Celsius away from direct sunlight," where the precision is invented. The second is genre-completion: the engine adding the clause a document of this type "usually" has, the standard liability disclaimer, the customary safety warning, the boilerplate exclusion, content that is plausible precisely because it is generic to the genre and therefore exactly what a pattern-completing model reaches for. The third is helpful elaboration: an explanatory phrase, a parenthetical, a clarifying example that the engine generated to be useful and that has no source warrant. The fourth, and the most dangerous in legal and medical content, is the invented obligation or right: a "must," a "shall," a "is entitled to," a granted permission or imposed duty that the source never stated, the category Hana's discretion clause belonged to, because it changes who owes what to whom.

The cheap pre-filter for additions, before you even walk, is length. Languages expand and contract by predictable ratios under translation, and Japanese to English has its own characteristic ratio. A target segment dramatically longer than the expansion ratio predicts is a flag: the extra length came from somewhere, and "somewhere" is either legitimate restructuring or invented content. Length is a coarse instrument; it produces false positives and misses additions that displace rather than extend. But it is nearly free, it can be computed automatically across an entire file, and it tells the reviewer which segments to walk hardest. The pipeline should compute it, flag the outliers, and route them to the careful target-to-source walk first.

Detecting Omissions: Walking Source to Target

Omissions require the opposite walk, and it is the harder of the two because absence is harder to perceive than presence. When you walk target to source, the contraband clause is right there in front of you, demanding inspection; you only have to refuse it a source. When you walk source to target, the missing sentence is, by definition, not in front of you. It is in the source, and your job is to confirm it arrived in the target, which means you have to hold the source unit in mind and go looking for its counterpart, and if the counterpart is absent there is nothing in the target to catch your eye, because the absence left no mark. This is why omission is the failure that vigilance misses most reliably and why a systematic source-to-target walk, rather than a target read, is the only dependable detector.

The walk is a roll call. You read the source, and for each meaning-bearing unit in it, a clause, a condition, a qualifier, a cross-reference, an exclusion, you call its name and confirm a target counterpart answers. Present: move on. Absent: an omission, logged. In Hana's segment 27, the Japanese source carried, among other units, a sentence excluding water damage from coverage. Roll call of the source units: the coverage grant, present in target; the duration, present; the water-damage exclusion, called, and no target unit answers, absent. The exclusion was dropped, and the English paragraph, having been smoothed by the engine, read as a complete and balanced statement of coverage with no hint that an exclusion had been removed. Only the roll call against the source surfaced it, because only the roll call was looking for the specific thing that was gone rather than reading the thing that remained.

Catching an addition is refusing a clause that cannot name its source. Catching an omission is calling a source unit's name and noticing that nothing in the target answers. The first is inspection; the second is a roll call, and the roll call is harder because absence leaves no mark.

Where the TM Makes Omissions Loud

The roll call against the source is the bedrock detector for omissions, but it is slow, and on a long file under deadline a reviewer cannot give every segment a full roll call. This is where the third reference, the translation memory, changes the economics, and it is the clearest illustration of why triangulation beats a two-way comparison. When a source segment has a high fuzzy match in the TM, the TM holds an approved prior target for nearly the same source, and that prior target is a second witness to what the units of meaning should be. If the prior approved target contained a water-damage exclusion and the new MT target does not, the TM comparison flags the divergence automatically, before any human roll call, and points the reviewer straight at the segment where meaning went missing. The TM converts the omission from an invisible absence the reviewer must hunt for into a visible divergence the tool can surface, because the TM remembers what was there even when the new target has forgotten.

This is not a reason to skip the source roll call; the TM only helps where a match exists, and plenty of segments are new. But it is a reason to run the TM comparison first, as a triage step, so that the segments where the MT and the approved past disagree are flagged and walked hardest, and the reviewer's scarce roll-call attention is spent where a second witness has already raised a hand. A pipeline that pre-computes TM divergence and routes the high-divergence, high-match segments to careful human review is spending human attention where it is most likely to pay, which is the whole game when the deadline is thirty minutes and the file is forty-one segments.

Tooling Assists and Where They Stop

A reasonable question at this point is why any of this is human work, given that a pipeline is software and software is good at comparison. Part of it genuinely is tooling work, and a serious pipeline automates everything automatable, because every check offloaded from human attention is attention freed for the checks only judgment can make. The tooling assists are real and worth configuring fully.

• Length-ratio outlier detection flags segments where the target is implausibly long (an addition tell) or implausibly short (an omission tell) relative to the source under the language pair's expansion ratio. Cheap, automatic, file-wide, and a strong first triage.
• TM divergence flags surface segments where a high-fuzzy-match TM entry exists and the new MT target diverges from the approved prior target, the single best automatic signal for both omitted and altered meaning on repeat content.
• Termbase QA checks flag segments where a source term maps to a termbase entry and the approved target term is absent, catching synonym-substitution fabrications and mandated-term omissions automatically.
• Number, tag, and placeholder reconciliation confirms the set of numbers, formatting tags, and code placeholders matches between source and target, catching a specific class of addition (an invented number) and omission (a dropped placeholder) by mechanical token-matching.
• Quality-estimation scores (QE, an automatic model-produced confidence estimate for a segment's translation) can route low-confidence segments to human review, and some QE models are specifically tuned to flag likely hallucination and omission.

Run all of these. They are fast, they are tireless, and they convert a forty-one-segment file into a short list of segments that deserve the expensive human pass. But understand precisely where they stop, because the place they stop is the place this lesson lives. Every one of these tools detects a proxy for hallucination or omission, not the thing itself. Length-ratio detection flags an unusual length, but unusual length is not the same as invented content; a legitimate restructuring is long and an addition that displaces existing text is not. TM divergence flags that the new target differs from the approved past, but difference is not error; the engine may have improved on a clumsy old segment. Termbase checks flag a missing approved term, but cannot judge whether the surrounding clause is faithful. The number reconciliation confirms that 25 appears in both source and target but cannot tell you that the clause containing the 25 was hallucinated whole. Every automatic check answers a syntactic or statistical question that correlates with the semantic failure but is not identical to it.

The semantic judgment, the actual determination that "at the manufacturer's sole discretion" has no source and is therefore a hallucination, or that the water-damage exclusion is genuinely absent and is therefore an omission rather than a paraphrase relocated elsewhere in the paragraph, is a meaning-to-meaning comparison, and meaning-to-meaning comparison is exactly what the tools cannot do reliably. They can tell you a segment is suspicious. They cannot tell you it is wrong, and they certainly cannot tell you it is a Critical wrong that fails the file rather than a Minor wrong you note and pass. This is the deeper reason the revised ISO 18587, the post-editing standard now expanded to cover AI and LLM output and in DIS ballot with publication targeted into 2026, insists the human in the loop hold the full competence of a professional translator: the human is not there to do what the tools do faster but to do what the tools cannot do at all, the meaning-level cross-reference and the severity judgment that follows it.

Every automatic check flags a proxy: an unusual length, a divergence, a missing token. The proxy correlates with the failure but is not the failure. The tool tells you a segment is suspicious; only the human cross-reference tells you it is wrong, and only the human tells you whether it is Critical.

The LLM as a Second Checker, and Its Trap

A tempting modern move is to ask an LLM to do the cross-reference itself: feed it the source and the target and prompt it to flag additions and omissions. This works partially and is worth using as one more automatic triage signal, because an LLM is genuinely capable of noticing that a target clause has no source basis, and a second model checking the first model's output catches some real defects. But it carries a trap that must be stated plainly. The checker is the same kind of engine that produced the failure, a fluency-first model that is itself prone to hallucination, and asking it to judge faithfulness invites two characteristic errors: it can hallucinate that an omission exists when the meaning was simply relocated, generating a false positive that wastes the reviewer's scarce time, and far more dangerously, it can read a hallucinated target as faithful because the hallucination is plausible to the same statistical sense that generated it, producing a false negative that gives the reviewer false confidence. An LLM checking an LLM is two fluency machines agreeing that fluent text is fine, which is precisely the failure mode the human cross-reference exists to break. Use the LLM checker as a flag-raiser that routes attention, never as a gate that clears a segment, and never let its "this looks faithful" substitute for the human roll call on high-consequence content. The accountability for the cross-reference stays human, because "the checker model said it was fine" is no more an answer than "the translator model said it was fine" when a Critical ships.

A Worked Cross-Reference Pass on a Real File

Let us run the full triangulation on a concrete file the way Hana ran it, slowly, so the method is a procedure you could execute tomorrow rather than a principle you admire. The file is the Japanese-to-English warranty, but we will work a representative six-segment slice with the two planted defects plus the kinds of pass and near-miss that make the method realistic. The setup is the one a real pipeline provides: the source segments, the MT target pre-populated in the TMS, a translation memory with several high-fuzzy matches from a prior version of this warranty, a termbase with the client's approved terms (among them "Authorized Service Center" for the concept of the official repair facility), and the automatic flags already computed, length-ratio outliers and TM-divergence flags lit on specific segments. The reviewer reads the four references together, segment by segment, walking target-to-source for additions and source-to-target for omissions, with the TM and termbase as the second and third witnesses.

Segment 17. Source (Japanese): "This warranty is valid for twelve months from the date of purchase." Target: "This warranty is valid for twelve months from the date of purchase." TM: a 100 percent match from the prior version, identical approved target. Termbase: no controlled term in this segment. Cross-reference: target-to-source walk, every English unit (warranty, twelve months, date of purchase) traces to a source unit, no contraband; source-to-target roll call, every source unit (validity, twelve-month duration, purchase-date anchor) answers in the target, no omission; TM confirms the approved past target is identical, a strong corroboration; termbase not applicable. Verdict: clean, and recorded as clean, because the record documents what passed, not only what failed.

Segment 19. Source (Japanese): "If the product fails under normal use, the customer may receive a replacement unit." Target: "If the product fails under normal use, the customer may receive a replacement unit at the manufacturer's sole discretion, including for cosmetic damage." Automatic flag: length-ratio outlier, the target is markedly longer than the Japanese-to-English ratio predicts. The flag is what tells the reviewer to walk this one hardest. Target-to-source walk: "if the product fails under normal use" traces to source, present; "the customer may receive a replacement unit" traces to source, present; "at the manufacturer's sole discretion" is called and demands a source, and the Japanese has none, contraband; "including for cosmetic damage" is called and demands a source, and the Japanese has none, contraband. Two hallucinated clauses, both fluent, both consequential. The TM, a high fuzzy match here, shows a prior approved target without either clause, a second witness confirming the additions are new. The termbase is not implicated. Verdict: two additions, the discretion clause and the cosmetic-damage clause, both logged. Severity to be set at the gate, and both are strong Critical candidates because they alter the warranty's obligations as the customer would read them.

Segment 24. Source (Japanese): "Repairs must be performed at an Authorized Service Center." Target: "Repairs must be carried out at an official repair shop." Automatic flag: termbase QA, the approved term "Authorized Service Center" is absent from the target. Length is normal, the source-target meaning map passes (an "official repair shop" means roughly an "Authorized Service Center"), and a reviewer working from the source alone would very likely pass this segment as faithful, which is exactly the blind spot triangulation covers. Target-to-source walk: no invented units, the meaning traces. Source-to-target roll call: no dropped units, the meaning arrives. But the termbase, the third witness, asserts that this concept must be carried by the exact approved term, and the engine substituted a fluent synonym. This is a fabrication of approved meaning: the target now asserts the repair-facility concept in words the client did not approve, which in a warranty can matter legally and which certainly matters for consistency with every other document in the program. Verdict: a terminology defect, an unapproved-synonym substitution, caught only because the termbase was in the triangulation. Logged as a Major, elevated to Critical if the program's rules make the term contractually binding.

Segment 27. Source (Japanese): "Coverage does not extend to water damage. The warranty covers manufacturing defects for the full term." Target: "The warranty covers manufacturing defects for the full term." Automatic flag: length-ratio outlier on the short side, and a TM-divergence flag, because the prior approved version in the TM contained the water-damage exclusion and this MT target does not. The two flags together point straight at the omission. Source-to-target roll call: source units are the water-damage exclusion and the manufacturing-defect coverage; "manufacturing-defect coverage" is called and answers in the target, present; "water-damage exclusion" is called, and no target unit answers, absent. The exclusion was dropped, and the English reads as a complete, balanced coverage statement because the engine smoothed the paragraph over the gap. The TM is the loud second witness here: the approved past target had the exclusion, this one does not, divergence flagged. Verdict: an omission of a coverage exclusion, logged, and a strong Critical candidate because dropping an exclusion silently expands the manufacturer's liability beyond what the source granted.

Segment 31. Source (Japanese): "Contact support within 30 days to register the product." Target: "Contact support within 30 days to register the product and activate your extended coverage." Length-ratio flag: mild, near the edge. Target-to-source walk: "contact support within 30 days" traces, present; "to register the product" traces, present; "and activate your extended coverage" is called and demands a source, and the Japanese has no mention of extended coverage, contraband. A subtler hallucination than segment 19 because it is short and reads as a natural continuation, the kind of helpful elaboration tell, and the kind a reviewer skimming would accept. Number reconciliation passes (30 in both), which is the reminder that mechanical checks clear the easy part and leave the semantic addition standing. Verdict: an addition, the extended-coverage clause, logged as a Major or Critical depending on whether "extended coverage" implies an obligation the manufacturer would have to honor.

Segment 36. Source (Japanese): "See Section 5 for the return procedure." Target: "See Section 5 for the return procedure." TM: 100 percent match. Termbase: "return procedure" present and approved. Number reconciliation: Section 5 in both. Target-to-source walk: clean. Source-to-target roll call: clean. Verdict: clean, recorded as clean. Including the clean segments in the worked pass is deliberate, because the method is not a hunt for guaranteed defects; it is a uniform procedure applied to every segment, and most segments pass. The discipline is running it on the smooth ones too, because the smooth ones are where the silent failure hides.

Reading the Verdict and Producing the Record

Six segments, and the triangulation surfaced four defects that a fluent-English read would have missed: two hallucinated clauses in segment 19, an unapproved-term substitution in segment 24, a dropped exclusion in segment 27, and a hallucinated extended-coverage clause in segment 31. Apply the severity gate, the rule this program returns to on every quality lesson, that a single Critical error fails the file regardless of how clean the rest reads. The invented discretion and cosmetic-damage clauses, the dropped water-damage exclusion, and very likely the extended-coverage and the contractually-binding-term defects are Critical, because each alters the legal obligations the warranty states. The file fails, and it would fail on any one of them. A pipeline that delivered this file on the strength of its fluent, lawyerly English would have shipped a warranty that granted rights the manufacturer never offered and dropped an exclusion the manufacturer relied on, and the engine's confidence would have carried it straight past a reviewer who read the English as English.

The second, equally important output of the pass is the record. Each cross-reference produces a logged line: "Segment 19, target-to-source, clause 'at the manufacturer's sole discretion' has no source, hallucination, Critical; TM high-match prior target lacks the clause, corroborated." "Segment 27, source-to-target, source unit 'water-damage exclusion' has no target counterpart, omission, Critical; TM-divergence flag confirms prior approved target carried it." That log is not a byproduct; it is the evidence that the cross-reference happened, what it consulted, and what it found, and it is exactly the audit-grade artifact the revised ISO 18587 and ISO 5060 (the 2024 standard that formalizes the MQM-aligned Critical, Major, and Minor severities used to score the defects) expect a defensible pipeline to produce. The triangulation did not only catch the errors. It produced the proof that the errors were caught and the proof of how, which is what turns a reviewer's "trust me, I checked" into a record a client, an auditor, or a court can read.

Building the Cross-Reference Into the Pipeline

A method that runs only when a reviewer happens to be suspicious is not a control; it is luck, and luck is exactly what an MT-first pipeline at scale cannot afford, because the file Hana caught was not special and the next one will arrive when nobody is feeling sharp. The point of this lesson at L3, the practitioner level where you build pipelines rather than just work inside them, is to make the cross-reference a defined, mandatory, instrumented stage that runs on every file by design. Three disciplines turn the method into a pipeline stage.

The first is placement. The cross-reference pass belongs after post-editing and before the delivery gate, as a distinct verification step with its own owner and its own pause point, not folded into the post-editor's own pass. The reason is the same one that makes self-review weak everywhere: the post-editor who rendered or accepted the target is primed to see it as correct, and the cross-reference works best as a fresh adversarial pass, ideally by a second person, that reads the target against the references rather than re-reads the target as prose. Where a second person is not affordable on every file, the pause point still matters: the post-editor stops, changes mode from editing to cross-referencing, and runs the four-reference comparison as a separate act with separate attention.

The second is routing by risk and by flag. Not every file deserves a full manual roll call on every segment, and pretending otherwise burns the throughput the pipeline exists to capture. The pipeline runs the automatic assists, length ratio, TM divergence, termbase QA, number and placeholder reconciliation, and QE, on everything, and uses the flags plus the content's risk tier to route human attention. A high-liability file, a warranty, a contract, a drug label, gets a full human triangulation on every segment regardless of flags. A low-liability file gets a full human pass only on flagged segments, with the rest cleared by the automatic checks. This is the same risk-tiered logic the whole L3 pipeline rests on: match the effort to the consequence, spend the expensive human cross-reference where the Criticals actually live.

The third is the record as a required output. The cross-reference stage does not pass a file silently; it emits the log, the severity classification, and the gate verdict as a mandatory artifact, so that the verification is provable and the quality record assembles automatically. A cross-reference that leaves no trace is indistinguishable, after the fact, from no cross-reference at all, and the entire value proposition of an ISO 18587- and 5060-defensible pipeline is that the verification is documented, attributable to a competent human, and reconstructable when a client or an auditor asks how the silent critical error was caught. Build the record into the stage and the proof comes for free; bolt it on later and it never comes at all.

A cross-reference that runs only when a reviewer feels suspicious is luck, not a control. Make it a defined stage: placed after post-editing as a fresh adversarial pass, routed by risk and by automatic flag, and required to emit a record. Then it catches the silent failure on the file nobody felt sharp on.

Key Takeaways

• A hallucination is target content with no source (an invented clause, a fabricated number); an omission is source meaning absent from the target (a dropped sentence, a vanished exclusion). They are mirror images, and neither is visible from inside the target, because the engine renders both fluently and self-consistently. The error lives in the relationship between target and source, which is why cross-referencing, comparing the target against the references that define what it should contain, is the only instrument that detects them.
• The method is two-directional by necessity, not thoroughness: walk target-to-source to catch additions (a customs inspection where every clause must name its source), and walk source-to-target to catch omissions (a roll call where every source unit must find a target counterpart). Each direction is blind to the other's failure, so running only one catches only half.
• Triangulation consults four references, not one: the source (the ground truth for meaning), the translation memory (a witness to the approved prior target, strongest for surfacing omissions on repeat content), the termbase (the only witness that catches synonym-substitution fabrications of approved terms), and the target (the unknown under examination, never a reference for itself). Where one reference is blind, another is watching.
• Additions announce themselves through tells: invented specificity, genre-completion boilerplate, helpful elaboration, and the dangerous invented obligation or right. Length-ratio outlier detection is a near-free file-wide pre-filter that routes the careful walk to the segments most likely to carry an addition.
• Omissions are harder to catch than additions because absence leaves no mark, so the source-to-target roll call, holding each source unit in mind and confirming a target counterpart, is the bedrock detector. The TM makes omissions loud: a high-match prior target that carried a unit the new MT target lacks is a divergence the tool flags automatically, pointing the reviewer straight at the dropped meaning.
• Tooling assists (length-ratio outliers, TM-divergence flags, termbase QA, number and placeholder reconciliation, QE scores) are fast, tireless, and worth fully configuring, but every one flags a proxy, an unusual length, a divergence, a missing token, that correlates with the semantic failure without being it. The tool says a segment is suspicious; only the human cross-reference says it is wrong, and only the human assigns severity.
• Using an LLM to check an LLM's output is a useful triage flag but a dangerous gate: two fluency-first models can agree that fluent hallucinated text is faithful (a false negative that breeds false confidence), so the LLM checker routes attention but never clears a segment, and the accountability stays human, the reason the revised ISO 18587 requires the human in the loop to hold full professional-translator competence.
• Make the cross-reference a defined pipeline stage, not a lucky habit: place it after post-editing as a fresh adversarial pass with its own pause point, route human attention by risk tier and automatic flag, and require it to emit a severity-classified log as a mandatory artifact. That log is the audit-grade quality record ISO 18587 and 5060 expect, and it turns "trust me, I checked" into reconstructable proof.