Pick Your Blind Spot

“Every tool is a weapon if you hold it right.” — Ani DiFranco, My IQ (1993)

I had a harness measuring whether a set of skills helped a cheap model, and it told me they didn’t. It was wrong, and the way I found out is the point.

The harness injected the skill on every case. Real runs don’t work that way — a skill fires when its trigger matches and otherwise stays out of the way. I only saw it because I ran the harness a second way, gated on the actual trigger, and the two runs disagreed. The verdict flipped: the skills clearly helped. Nothing about the skills had changed.

Notice what did the finding. Not a more careful reading of the harness — a contrast. One configuration against another. And notice what the contrast did not do: it told me the two runs differed, and it had no opinion about which one was right. I supplied that, from knowing how skills actually fire. The comparison reached something I hadn’t specified, and then handed the judgment straight back to me.

That’s the whole argument in one receipt. There are two ways to ask whether a change is any good, they reach different distances, and the one that reaches further does it by a mechanism you should be nervous about.

Bottom line up front

• A gate and a comparator ask for different kinds of judgment. A gate asks for an absolute verdict — does this meet the standard — so the judge has to hold the standard internally. A comparator asks only for discrimination — is A different from B — so the difference only has to be detectable. Discrimination is the cheaper faculty, which is why the comparator genuinely reaches past what you wrote down.
• The reach and the gameability are the same property, not a trade-off between two. A comparator reaches further because it is sensitive to what you put in front of it. A planted distractor is a thing you put in front of it. Under deliberate manipulation, pairwise preferences flip ~35% of the time against 9% for absolute scores (Tripathi et al., COLM 2025) — that number is the bill for the reach, not a separate defect to weigh against it.
• There are three ceilings, not two, and you only ever move between the first two. What you specified (the gate’s) → what a judge can discriminate when shown a contrast (the comparator’s, genuinely higher) → the correlated-error floor, which neither instrument escapes.
• The comparator’s reach is real and unauditable. You cannot tell no difference from didn’t look on that axis. A gate’s standard is yours, readable, versionable; a comparator’s operative standard is implicit and you never get to read it.
• So don’t ask which is better. Ask which failure you can afford here — and buy back the coverage the other one would have given you.

The two instruments ask for different faculties

A gate is absolute. It asks whether this artifact clears a bar, and the bar is whatever you managed to write down. That’s the whole of its reach. Everything you understood but never encoded — the convention nobody documented, the failure mode you’d recognise instantly and have never named — sits outside it. This is the oracle problem in work clothes: what counts as correct is an assumption you supply, not something the check discovers (Weyuker, 1982; Barr et al., 2015). A gate can be exhaustive about the part of your intent that made it onto paper and structurally blind to the rest.

A comparator asks for something weaker, and the weakness is the feature. Put two versions in front of a judge and ask which is better. The shared flaws cancel — whatever both arms get wrong drops out — and what survives is the difference. Crucially, you never had to name the axis. The judge doesn’t need to hold your standard; it needs only to detect that the two things aren’t the same.

That gap between the two faculties is not something I’m asserting because it sounds tidy. It’s the reason preference learning is built the way it is. Christiano and colleagues’ method has humans compare short clips of behaviour rather than score them, on the explicit grounds that the comparison is the judgment people can give reliably when the absolute number is beyond them. Asking for less produces more.

So the comparator’s ceiling really does sit above the gate’s. I want to be exact about which ceiling that is, because I got it wrong the first time I wrote this down.

Three ceilings, and the one you can’t move

1. What you specified. The gate’s ceiling. It is yours: readable, reviewable, versionable, and it fails by omission in ways you can go and look at.
2. What the judge can discriminate when shown a contrast. The comparator’s ceiling, and genuinely higher than the first — that’s the whole finding. But it is not yours and you cannot read it. When a comparator reports no difference, you have no way to distinguish there is no difference from nothing looked on that axis. The operative standard still exists; it has just moved somewhere you can’t audit it.
3. The correlated-error floor. Judges drawn from similar training share failure modes, so agreement between them isn’t independent evidence (Kim et al.). Neither instrument touches this. Stack all the comparators you like.

The mistake worth naming: I first wrote that a comparator “escapes the intent ceiling.” That’s false in a specific and useful way. It escapes the absolute-judgment requirement — the need for the judge to hold your standard internally. It does not escape specification. It converts your specification from an artifact into an implication, and an implication is not a thing you can review.

The reach is the gameability

Here’s where my own reasoning got priced, and I’m keeping the receipt because it went against me.

Having worked out that a comparator reaches further, I drew the obvious conclusion: comparators are the better instrument. Tripathi and colleagues ran exactly that comparison under deliberate manipulation — a generator embedding distractor features the judge happens to favour — and pairwise preferences flipped in about 35% of cases against 9% for absolute scores. Roughly four times easier to play.

When I first wrote that up, I filed it as an instinct the result had defeated: a comparison feels like it ought to be the safer instrument, and — in that post’s own words — on this axis it’s the easier one to play. The limiter was already there. What I left unsaid is whether the mechanism behind the feeling survives. The reason a comparison feels safer is that a shared flaw is held constant across the arms, so the delta should survive what the level doesn’t. Tripathi defeats the feeling and says nothing about the mechanism — and the mechanism is where the reach comes from. So the reach stands. That’s a sentence I’m finishing, not an error I’m correcting.

Because the two are the same property. A comparator’s mechanism is sensitivity to difference. A planted distractor is a difference. The judge that will report a regression you never wrote a rule for is, by exactly the same disposition, the judge that will report an improvement somebody manufactured for it. You do not get to keep the first sensitivity and decline the second — there is only one sensitivity. Tripathi’s 35% is what the reach costs.

And note what this does not say: that a gate is safe. A gate is gamed too, just from the other end — you satisfy the letter of what got written down. That’s Goodhart with the serial numbers filed off, and it’s why the 9% isn’t 0%. The instruments don’t differ in whether they can be played. They differ in what you have to do to play them.

• A gate fails on intent, and is gamed by satisfying the specification you wrote instead of the one you meant.
• A comparator fails on audit, and is gamed by planting a difference it is built to notice.

What to do with that on Monday

Choosing an instrument is choosing a blind spot. Every position has one; a position without one would have to be everywhere at once. So choose deliberately, name what the choice hides, and go buy that coverage somewhere else.

Concretely: if you’re gating, you have accepted a ceiling at your written intent, so something in the loop needs to be probing past it — and a contrast is the cheap way to probe, which is exactly what saved me on the harness. If you’re comparing, you have accepted an unauditable standard and a planting surface, so something needs to check that the difference being reported was earned rather than manufactured. Ask what the delta was on. A comparator that can’t tell you which axis moved has told you less than it appears to.

And “use both” is not the free answer people reach for. Running a gate and a comparator gives you two instruments with two blind spots, which beats one only if you know which blind spot each is covering. Two checks that fail together are one check with better marketing.

The point

The instrument doesn’t just measure the work. It selects which of your failures are visible and which are structurally out of reach. A gate can’t see past what you wrote down. A comparator can, and the seeing is the same faculty as the being-steered.

None of that is solvable. All of it is choosable. So pick the failure you can live with, say out loud what the pick is hiding, and buy that coverage back somewhere you can name.

---

How sure I am:

• Cited. The oracle problem, and therefore the gate’s intent ceiling (Weyuker; Barr et al.). The 35%/9% manipulation gap (Tripathi et al.). The correlated-error floor (Kim et al.). That preference learning is built on comparison because comparison is the judgment humans give reliably (Christiano et al.).
• From experience. The harness that lied, read here for a third thing: that a contrast found the flaw and had no opinion about which side was right. One instance. It shows an instrument can be confidently wrong and that a comparison can surface what inspection didn’t; it can’t tell you how often either happens.
• A wager, and the weakest joint in the post. The claim that a comparator reaches further for an LLM judge rests on a transfer plus an argument. Christiano’s evidence is about human raters; Tripathi’s is about LLM judges but measures manipulability, not reach. Nobody, as far as I can find, has measured whether a comparator surfaces defects a well-specified gate misses. The psychophysics literature points the same way and I have now read it to primary — Stewart, Brown and Chater find absolute identification “surprisingly poor compared with discrimination of the same stimuli” — but it measures humans on one-dimensional stimuli and says nothing about direction, so it still isn’t backing here.
• Mine. The three-ceiling structure, and the identity claim that reach is gameability. Argued, not measured. It also corrects my own earlier framing rather than someone else’s, which is the kind of correction I trust least in myself.

What would change my mind: hold the reviewer and the task constant, vary only the protocol — gate versus comparator — and count the defects each finds that the other doesn’t. If the comparator’s unique-find set is empty against a well-specified gate, the reach is imaginary, the 35% is a pure loss, and this post collapses into “write better specifications.”

Sources

• Ani DiFranco, My IQ, on Puddle Dive (1993) — “every tool is a weapon if you hold it right.” Her subject is not evaluation; the epigraph borrows the identity claim, that the property and its abuse are one thing.
• Weyuker, On Testing Non-Testable Programs, Computer Journal 25(4) (1982); Barr, Harman, McMinn, Shahbaz & Yoo, The Oracle Problem in Software Testing: A Survey, IEEE TSE 41(5) (2015) — what counts as correct is supplied, not discovered. The gate’s ceiling is this result.
• Christiano, Leike, Brown, Martic, Legg & Amodei, Deep Reinforcement Learning from Human Preferences, NeurIPS 2017, arXiv:1706.03741 — the method elicits comparisons rather than scores because comparison is what human raters can supply reliably. Cited for the faculty gap, with the human-rater scoping caveat above.
• Tripathi, Wadhwa, Durrett & Niekum, Pairwise or Pointwise? Evaluating Feedback Protocols for Bias in LLM-Based Evaluation, COLM 2025, arXiv:2504.14716 — “Pairwise preferences flip in about 35% of the cases, compared to only 9% for absolute scores.”
• Strathern, “Improving ratings”: audit in the British University system, European Review (1997); Goodhart (1975) — the law the gate’s own gaming mode instantiates.

---

Extends Evals Aren’t the Moat. That post reports the 35%/9% result as defeating the instinct that a comparison must be the safer instrument, and scopes the defeat to that axis; this finishes the sentence, because the mechanism that made the comparison feel safer is also what gives it reach. Also extends Models Narrow the Overlap (the intent ceiling and the correlated-error floor — this asks which instrument you reach for between them). Adjacent to The Harness Is the Product.