Stopping a CV from inventing things

Asked to make a CV fit a job, a model is being asked to close a gap. When the source material does not contain what the posting wants, the instruction and the evidence conflict — and a model resolves that by producing something plausible.

This is not random error. It is the predictable consequence of an instruction that implicitly asks for a better match than the facts support, which is why prompting for honesty alone does not fix it.

Cover letters are worse than CVs for the same reason amplified. Prose has no structure pulling it back towards a record, so a sentence about a project the candidate did not run is easy to produce and easy to miss inside a fluent paragraph.

Outright fabrication is rare and easy to spot. The dangerous cases are small escalations that look like ordinary editing, each individually defensible and collectively a different person.

Watching for these specific shapes is more effective than looking for "lies", because none of them reads as a lie in isolation.

They also survive naive checking. A claim-matching pass that confirms the subject matter appears in the source will pass "led" where the record says "contributed to", because the content matches and only the strength changed.

• Scope inflation — "contributed to" becomes "led"
• Number invention — an unquantified result gains a percentage
• Technology bleed — a tool from the posting appears in your history
• Seniority creep — the same work described one level up
• Timeline smoothing — a gap quietly disappears

The structural fix is requiring provenance. Each generated bullet must cite the part of the candidate’s original material it derives from, and anything without a source is rejected before a human ever sees it.

This changes the task from "write a good CV for this job" to "select and rephrase from this material for this job", which is both the honest framing and a much easier problem for a model to do well.

Go further and never let it write a fact at all. Dates, titles, employers and figures are copied from the structured record rather than generated, because a model that cannot produce a date cannot produce a wrong one — a stronger guarantee than any instruction about accuracy.

Run a separate check over the output: do all numbers in the generated text appear in the source? Do all named technologies? Do the dates and titles match exactly?

A second model call asked only to find unsupported claims catches a good deal, and a deterministic check on numbers, dates and proper nouns catches most of the rest at no model cost. Both are cheap relative to the consequence.

Check strength as well as content, since that is where the deterministic pass stops helping. Comparing the verb and scope of each generated line against its source catches the escalation that a content match approves.

Do not let the generating model verify itself in the same call. A pass that is simultaneously trying to write well and to police its own output does both worse, and a separate check with a single job is more reliable and cheaper.

Even a perfect verification pass needs a human at the end, and presenting a finished CV guarantees it stops being read by the fifth application. A reviewer skims a clean page and approves it.

Show what changed instead: these three bullets were reworded, this summary rewritten, these roles dropped. Five diffs take less attention than one document and they direct that attention exactly where drift would appear.

Flag the unverifiable lines rather than leaving them to be spotted. A line the system could not trace should arrive marked, because that is the one sentence on the screen that genuinely needs a person to decide.

The most effective single change is telling the system it may not close gaps. When the posting asks for something the candidate lacks, the correct output says so rather than manufacturing coverage.

This is better for the candidate too. Knowing a role wants two things they do not have is actionable — they can address it, or skip the application — whereas a CV that quietly claims both sets up a bad interview they will not see coming.

Ask before concluding the gap is real, though. Phrased as "your record does not evidence this — do you have it?", a large share of apparent gaps turn out to be things the candidate has and never wrote down, which is an addition rather than a fabrication.