Some genuinely good, accurate drafts were getting rejected by checks that were too strict, or that misread what an instruction actually meant. This fix corrects three specific, unrelated problems: traceable claims now pass, instructions are judged on their meaning rather than their exact wording, and a "don't mention X" instruction is applied as plain guidance instead of accidentally being flipped into a requirement to mention X.
What changed
The validation system that checks drafts before they are allowed to send is supposed to catch real, concrete problems, like an unsupported claim or an instruction the AI plainly ignored. But three specific bugs were causing it to reject drafts that were actually fine, which meant good outreach was getting stuck for reasons that had nothing to do with real quality issues.
First, claims that were genuinely traceable back to real research were sometimes still failing the evidence check. A draft could reference something true and sourced, backed by an actual verified fact about the prospect, and still get held back as if the claim had no backing at all.
Second, the system was checking instructions too literally. If you told the AI to do something in your own words, and the AI followed the intent correctly but phrased it differently than your exact wording, the check could fail the draft anyway, even though it did exactly what you asked it to do.
Third, and most strangely, a negative instruction like "don't mention our competitor" could get inverted during validation, effectively treating the instruction as a requirement to mention the thing you explicitly asked to avoid. A draft that correctly left something out could get flagged as if it had failed to include it, which is the opposite of what the instruction was meant to enforce.
All three bugs shared a common shape: the validation system was checking for the letter of an instruction or a claim rather than its actual intent. A claim with real evidence should pass regardless of exact phrasing, an instruction should be judged on whether the draft did what you meant, and a rule about what to avoid should be enforced as written, not flipped. In each case, the strictness that is supposed to protect you from bad drafts was instead working against drafts that were already correct.
Traceable claims that have real evidence behind them now pass validation as they should. Instructions are now evaluated on whether the draft honored their meaning, not whether it repeated your exact phrasing word for word. And "don't mention X" instructions are applied as guidance to avoid that topic, not misread as a requirement to bring it up.
This is a meaningful shift in how the validation layer reads instructions overall. Rather than pattern matching against your exact words, it is now checking for whether the underlying intent was respected, which is a closer match to how you actually think about your own instructions when you write them. You write instructions the way you would explain them to a person, not as a script to be matched character for character, and the check now works the same way.
How to use it
If you had campaigns where drafts were getting held back or rejected for reasons that did not make sense when you actually read them side by side with your instructions, especially around claims you knew were accurate or instructions you knew the AI had actually followed, it is worth checking those campaigns again. Drafts that were previously stuck should now move through validation correctly and start reaching prospects.
If you use negative instructions, like avoiding a specific topic, competitor, or phrase, you can trust that they now work as intended. A draft that correctly avoids the topic will not be penalized for it, and you should not need to work around this by rephrasing your instructions to avoid the bug.
If you had previously rewritten instructions into unusually rigid or repetitive phrasing to try to force them past validation, you can simplify them back to how you would naturally describe what you want. That kind of workaround should no longer be necessary, and plainer instructions should now be judged fairly on what they actually mean.
Why it matters
A validation system that rejects good drafts is worse than one that is too lenient, because it slows down campaigns that were actually working correctly and forces you to spend time investigating a problem that, on closer look, was not really there in the first place. These fixes remove three specific sources of friction that had nothing to do with real draft quality, so a draft that is factually accurate and follows your instructions gets treated that way, and reaches your prospect on schedule instead of sitting stuck behind a false rejection for no defensible reason. It also means the checks you rely on to catch real problems are more trustworthy, because a false rejection here and there used to make it harder to tell whether a hold on a draft was worth investigating or safe to ignore entirely.
None of this loosens what validation actually enforces. A claim without evidence still fails. An instruction the AI genuinely ignored still fails. A draft that mentions something you asked it to avoid still fails. What changed is that validation now recognizes when a draft has actually done the right thing, instead of occasionally punishing it for doing so, which is what a strict check is supposed to do in the first place.
If you had grown cautious about how much you trusted validation results, treating a rejection as ambiguous rather than a clear signal something was wrong, these fixes are meant to restore that clarity, so a held draft is once again a reliable sign that something genuinely needs your attention rather than noise to filter out before you can get back to the actual work of running your campaigns. Give it a few days of normal sending and you should notice fewer holds that turn out, on review, to have been nothing at all.
Availability
These fixes apply automatically to every campaign drafting an AI-written email now, with nothing to turn on and nothing to reconfigure. They cover claims, instructions, and negative instructions across first-touch messages, follow-ups, and replies alike.