Corvid Systems wants to buy Tessel Analytics.
On paper it's a good deal. A twelve-section board memo lays out the case, a financial appendix backs it with numbers, a competitor teardown supports the strategic logic, and an email thread from the deal team fills in the color everyone conveniently leaves out of the memo. It reads like something a real board would approve - mostly because a real board would approve it.
It's also, deliberately, wrong in a long list of specific, hidden ways. Scattered next to those planted flaws are an equal number of statements that look exactly as suspicious and are actually fine - the kind of thing a jumpy critic flags anyway, just to look thorough.
Corvid and Tessel don't exist. I built them, buried the flaws myself, and then set two different ways of using AI loose on the case to see which one actually found the trap.
The actual question
The Computational Deliberation Lab is built on one bet: that a machine can take the deliberation methods we already know how to design - the ones built, over decades, to make decisions better, smarter, and more robust - and make them automatic and available at the exact moment a real decision needs them. AI, done right, can calculate something close to the output of a genuine deliberation process - not a substitute for one, but an equivalent, on demand, at a cost no human team could sustain. Before that bet can be tested at scale, there's a narrower, more testable question sitting underneath it: if you want AI to stress-test a real decision, is it better to ask one strong model to be a tough critic and run that same request several times - or to split the job into narrow, separately-instructed passes, each one only allowed to look for one kind of problem, run independently and then compared?
Asking one strong model to simply try harder is the obvious, cheap default, and for good reason - it's simpler, it requires no extra design work, and until you've actually run the comparison, there's no reason to assume it's wrong.
So I ran the comparison, instead of assuming I already knew the answer.
The setup, in plain English
Two "arms," matched as closely as I could make them. Both read the exact same four fictional documents, cover to cover, with no internet access, no tools, and nothing else to go on. Both wrote under the same length limit per response. Both ran ten times, from a completely fresh conversation each time, so nothing from one attempt could leak into the next.
Arm one was one broad, tough-critic generalist prompt, rerun ten times.
Arm two was ten separately-instructed narrow lenses - call them Pass 1 through Pass 10 - each one told to look at the case from exactly one angle (the arithmetic, who benefits, how a skeptical outsider would read it, and so on) and to stay in its lane rather than try to cover everything.
The idea, going in, was that ten specialists each looking hard at one thing would collectively cover more ground with less overlap than one generalist saying the same kind of thing ten times.
Every response was graded blind, one finding at a time, against a scoring bar written before any output existed - an actual quoted piece of evidence was required for credit, "sounds concerning" didn't count, and a hedged or uncertain finding was scored against the arm, not for it. The same reviewer also checked both arms against a set of statements built to look exactly like the planted flaws and be completely sound, to see how often each arm cried wolf.
The honest headline
The tough-critic prompt won. Rerun ten times, it found more problems per attempt and more total ground than the ten narrow lenses did, including on the hardest, best-hidden issues in the deck.
Ten specialists, each looking at one narrow slice, lost on points to one generalist told to be difficult, ten times in a row.
I want to say that plainly, because it's the least flattering sentence in this piece and it's also the true one. This was a first pass at testing one of the lab's more basic assumptions, and the version of "narrow, isolated analysis" I actually ran was a fairly naive one - instructions adapted almost verbatim from a larger, more careful methodology that wasn't fully switched on for this run. When I calibrated the test beforehand against what a single strong reader should find, I'd expected something in the 35-55% range per attempt. The tough-critic generalist prompt came in below that band. Which means, before anything else, this was a harder case than I'd built it to be - a useful thing to learn, and not the finding I set out to publish.
The more interesting number
Here's the part that made rerunning the analysis worth it. Winning on raw coverage isn't the same as winning on independence, and the two arms were not equally independent of themselves.
The tough-critic prompt, rerun ten times, kept finding the same handful of big, obvious problems over and over. Most of its total output was re-discovery, not new discovery - the digital equivalent of ten friends independently telling you your presentation has a typo on slide one, and nobody mentioning slide fourteen.
The ten narrow lenses barely repeated each other at all. When a lens found something, it was usually the only one of the ten that did. Under the strict scoring, the narrow passes also never flagged a single one of the planted "actually fine" statements as a problem - the cleanest read of caution anywhere in the run. The tough-critic prompt occasionally did.
So: one arm is a better single reader. The other arm is a better set of ten readers.
Those are different questions, and the lab's actual hypothesis was about the second one.
Why this comparison isn't finished
Two things kept this from being a clean result, and both are fixable rather than fatal.
First, the tough-critic prompt simply wrote more per attempt - noticeably more, even against the same length limit - and writing more is partly how you find more. Some, not all, of its lead is just longer answers.
Second, and more important: the ten narrow lenses were designed to hand off to each other - each one is supposed to read the previous lens's open questions and build on them, with a final pass allowed to close the case. In this run, none of that happened. All ten ran as completely disconnected, one-shot attempts, and were then graded as if each one were supposed to be a complete analysis on its own. That's a bit like grading ten specialists' private notes as though each of them, alone, was expected to have written the whole diagnosis - and marking them down for correctly declining to.
Put those two together and the honest reading isn't "narrow analysis loses." It's "a naive, unchained, under-instructed version of narrow analysis loses to a strong generalist told to try harder - which is exactly the kind of gap that better instructions and an actual synthesis step exist to close, and I'm already rebuilding both."
The part that actually points somewhere
Two things in the data are, I think, genuinely promising, and neither one required the narrow lenses to win outright.
The first: when I combined a small number of tough-critic attempts with several of the narrow lenses, the mixed group matched - and with a couple more lenses added, beat - a much larger group made entirely of tough-critic reruns. Diversity of angle, even in this rough a form, bought something that more of the same angle didn't.
The second, which in my mind matters more when you think about large scale: the tough-critic prompt's returns dried up fast. Well before its run of ten was over, it was mostly rediscovering things it had already found. The narrow lenses never showed that pattern - the tenth lens was still turning up something new, at close to the same rate as the early ones had.
One arm was approaching a ceiling. The other hadn't found one yet.
That's the actual promise of splitting a hard question across many isolated, independently-instructed passes instead of just asking harder: repetition saturates, but genuine difference of approach doesn't have to. Ten reruns of the same request is a small, fast-closing well. Ten different, well-built angles might not be a well at all - and the honest next question is what happens at a hundred angles, or a thousand, once "ask the same question again" has been fully wrung dry and "ask a different question" hasn't even started to be.
I'm not the only one betting on this
In late 2025, OpenAI published the full prompt behind what it says was a complete resolution of the Cycle Double Cover Conjecture, a graph theory problem that had been open since the 1970s.[1] It's worth actually reading, because the instructions aren't "think harder." They're closer to: run dozens of agents on genuinely different approaches at once, deliberately keep them independent early so they don't all converge on the same convenient dead end, track which "family" of approach each one belongs to, redirect people away from the family everyone's already crowding into, and use dedicated adversarial agents throughout to try to break every candidate result rather than take a confident report at face value.
I didn't need that document to believe isolated, diverse, adversarially-checked deliberation was worth building toward - it's the premise this lab started from. But it's a good feeling to watch the frontier of AI research reach for the same shape of answer, on a completely different problem, for what looks like the same underlying reason.
One mind pushing on the same idea harder runs out of new things to say a lot faster than several different minds looking at it from angles that don't overlap.
What I'm not publishing, and why
This page leaves out the exact scoring rubric, the full list of what was planted and where, the precise detection numbers, and the specific instructions each of the ten lenses ran on. That's not coyness - it's the same rule this lab runs on everywhere else.
Publish the what and the why, keep the how private, because the how is also the thing that makes the next test worth running.
A scoring bar or a planted-flaw list that's already public stops being a fair test the moment it's public.
If you want to see more of it - the actual numbers, the scoring bars, how the lenses were written - I'm genuinely happy to walk through it with people directly. Reach out.
What's next
The obvious next run is the one this one didn't test: the same ten lenses, actually chained the way they're designed to be, with a real closing step allowed to turn ten sets of open questions into one verdict - scored against the same kind of case, the same bars, the same rules.
I'm also rewriting the lens instructions themselves; what ran here were close to first drafts, and I don't think they're a fair test yet of what isolated, well-specified attention can do.
Sharper instructions, a real synthesis step, and a harder case are all already in progress.
If you want to try this at home
You don't need a research lab to use the one piece of this that's already solid: if you have a real decision you want stress-tested, don't ask one AI once.
A few things that seem to genuinely help:
- Run it independently, at least four times. Not four follow-up messages in the same conversation - four fresh conversations, same prompt, same materials. You'll typically get more real ground covered for the effort than one long, exhaustive-feeling conversation gives you.
- If you want different angles, ask for different angles. "Please be critical" four times tends to produce four versions of the same critique. Naming a distinct lens each time - the numbers person, the person who has to actually implement this, the skeptical outsider - costs nothing extra and produces more overlap-free ground.
- Treat any single confident answer as a hypothesis, not a verdict. The value shows up when you compare several independent takes side by side, not when you take the first one at its word.
- Watch for the attractor. A good AI critic will happily spend its whole effort on whatever part of your document is most numerically dense, because that's the most rewarding thing to check - not necessarily where the real problem is. Ask, separately and explicitly: "what am I not even asking about?"
- Aggregate for what's missing, not just what's confirmed. The interesting output isn't the pile of everything all four passes agreed on. It's the handful of things only one of them said.
An open question
The tough-critic prompt is still, today, the better single reader.
What I don't yet know - and what the next run is actually built to test - is how many genuinely different, well-instructed angles it takes before that stops being true, and whether "genuinely different" is a property you can design for on purpose, or one you mostly find out about after the fact.
Appendix: for people who want a bit more of the mechanics
Everything below is still a summary. It leaves out the scoring bar itself, the full defect list, the internal lens instructions, and every number precise enough to reverse it into any of those. What it adds is a clearer picture of how much checking happened before a single figure above was allowed to be called a result.
The setup, restated precisely
- Both arms ran ten times each, fresh conversation every time, same four documents, no web access and no other tools available to either arm.
- Both were held to the same instructed cap: 1,800 words per response. The tough-critic prompt averaged roughly 2,000 words anyway and had to be regenerated more than once to even get that close; the narrow lenses averaged roughly 1,200 and never came near the ceiling.
- The instructions for both arms were frozen before the run started - the tough-critic prompt and all ten narrow lenses - and were not touched, tuned, or quietly improved once a single output had come back. Whatever came out is what the instructions, exactly as originally written, actually produce.
- Every prediction about how this would probably turn out - which arm would likely do better on which kind of problem, where I expected a tie, where I expected one side to embarrass itself - was written down and locked before I saw a single graded result. Several of those predictions turned out to be wrong, which gets discussed below rather than quietly dropped.
What kinds of flaws, and why those
I didn't plant one kind of problem and repeat it ten times. I planted several genuinely different kinds, on purpose, because the comparison falls apart if every flaw rewards the same kind of reading.
Some flaws live entirely in one sentence, in one place - a careful reader who checks that sentence against itself will catch them. Others only exist in the relationship between two documents: a number in one place quietly means something slightly different from the number it's being measured against elsewhere, and neither sentence, read alone, is wrong. Others are about what usually happens to plans that look like this one, independent of anything specific in the documents at hand - the kind of thing you only catch if you already know the base rate for this category of decision. Others depend on noticing who supplied a piece of evidence and what they stood to gain if it were believed, which has nothing to do with how well-argued the sentence itself is. And a few aren't stated anywhere at all - an entire stakeholder, an entire cost, a whole category of consideration that a genuinely thorough reader would think to ask about and a merely careful one would not.
I planted that range on purpose, because a test built out of one flavor of trap doesn't measure what I wanted it to measure. If every planted problem were the single-sentence, quotable kind, the comparison would mostly be a test of who reads most closely - and a generalist rereading the same material several times is a perfectly good strategy against that specific kind of trap, no procedure required. If every problem instead required connecting several documents, the comparison would tilt the other way, toward whichever approach naturally holds more of the material in view at once. Neither result would tell me anything about the actual question, which is what happens across a genuinely mixed field of traps - the kind a real decision actually contains, where some flaws are shallow, some are structural, some depend on knowing who's talking, and some aren't written down anywhere for either side to find.
That's also why the composition mattered for the lab's larger bet, not just for this one write-up. The whole premise behind splitting scrutiny into narrow, independent passes is that different angles should be good at catching different shapes of problem - one instinct for what a plan quietly assumes, another for who benefits from a piece of evidence, another for what's missing entirely. If the planted flaws had all been one shape, that premise would never actually get tested: every lens would be graded against a target that only rewards one of its instincts, and the question this whole demonstration exists to ask - does genuine diversity of angle find more, in total, than one angle run repeatedly - would have been begged rather than answered. So the mix of flaw types wasn't a production detail sitting underneath the experiment. It was close to the experiment itself.
What I'm not putting on this page is the actual list of categories, how many of each kind, or where each one lives - for the same reason the scoring bar and the exact count stay off it. A published map of what to look for turns the next version of this test into an open-book exam for whichever arm reads this essay first.
Keeping the grader honest
The part of this that took the most care wasn't running the two arms. It was making sure the thing scoring them couldn't be fooled, couldn't see which arm it was looking at, and couldn't drift partway through.
- Every planted problem, and every look-alike-but-sound control statement, was graded in its own separate, isolated pass - never read alongside the others, and never told which arm, or even which pass number, produced the text in front of it.
- A finding only counted if the grader could point to an actual quoted piece of the text as evidence. "This seems concerning" wasn't enough.
- Hesitation counted against a hit, not for one. If a grading pass wasn't sure, the rule forced it to record that as a miss rather than a maybe - deliberately, so the headline numbers couldn't be flattered by generous scoring.
- Before any real material was scored, the grading process itself was tested against a small batch of answers I already knew the correct grade for, and it had to reproduce those correctly before it was trusted with anything that mattered. It also had to correctly catch a known planted problem in a dry-run pass before the real scoring was allowed to start.
- Every output was banked and fingerprinted the moment it came back, before the next one was even generated, so nothing could be quietly edited, reordered, or swapped on the way to being graded.
Catching a mistake in my own scoring
Partway through analyzing this run, I found a bug in my own rule, not in the AI's output.
The scoring had a rule that hedged or uncertain findings should count as misses under the strictest reading. When I checked how the grading had actually been applied, a handful of hedged answers had slipped through and been counted as hits anyway. I rewrote the rule so it couldn't be applied that way, rescored the entire run under the corrected version, and republished the corrected numbers - which is why the figures above sit where they do rather than somewhere slightly more flattering. One earlier version of the story briefly made the narrow-lens arm look marginally ahead on one measure. The corrected numbers erase that. I'm telling you this instead of quietly fixing it, because a lab that only tells you about the errors that don't matter isn't one worth trusting about the ones that do.
A second, independent read of the same numbers
I didn't just trust my own first pass at reading these results. I had a separate analysis redo the whole thing from scratch - starting from the raw, item-by-item grading data alone, before reading anything I had already written about what it meant - and then compared notes with my original write-up.
It agreed with the headline story: the tough-critic prompt covers more ground, the narrow lenses overlap each other less. But it also caught two small bookkeeping mix-ups in my own notes, both of the same kind - a mislabeling of which arm one of my pre-registered predictions was actually about, which flipped whether that specific prediction should have been marked as holding or failing. Neither mistake touched a single number in the tables above; both were about which column a pre-existing sentence pointed at. I corrected both once they were found. That kind of second, independent, start-from-raw-data check is exactly the sort of thing I'd want run on someone else's claimed result before believing it, so I ran it on my own first.
Five false starts before this one worked
What you're reading is not the first attempt at running this comparison - it's closer to the sixth. Five earlier attempts were each stopped before they produced a single graded result, for five different reasons: a technical safeguard that couldn't get the material it needed, a scoring script that failed its own pre-flight test against known answers, and a couple of process checks that flagged something ambiguous and halted rather than guessed. None of those five attempts contributed so much as one number to anything published here. Throwing away a run that hit a snag, rather than salvaging whatever it happened to produce, is slower. It's also the only way the numbers above mean what they say they mean.
Checking for leaks and drift
Every output, from both arms, was scanned afterward for any sign that something had gone wrong in transit - the wrong material being seen, a prior attempt's content leaking into a new one, or one call quietly seeing something it shouldn't have. The scan flagged a number of false alarms, mostly harmless coincidences, like a random file identifier that happened to share two letters with an internal label. Every flag was individually checked by hand and confirmed harmless. Zero real leaks turned up anywhere in the run. A checking process that never flags anything isn't a careful one; a checking process that flags things and lets you verify each flag by hand is.
Testing for false alarms, not just hits
Alongside the planted problems, the case also contained a set of statements built to look exactly as suspicious as a real problem while actually being sound - the equivalent of a perfectly reasonable expense sitting next to a fraudulent one, placed there to see if either arm would flag it out of sheer nervousness. This matters because a critic that accuses everything of everything technically never misses. It's just useless. Both arms were graded on how often they cried wolf, not only on how much they caught, and that comparison is part of what makes the finding above about caution worth taking seriously rather than assuming.
Four ways of counting the same result
Because the two arms didn't write the same amount per response, I didn't settle for one headline number. I ran the comparison four different ways - findings per thousand words written, findings per distinct thing raised, coverage per call at a matched total word budget, and coverage per word spent by the whole group of ten - and they don't all point the same direction. Some favor the tough-critic prompt, some favor the narrow lenses. The essay above reports that disagreement honestly rather than picking whichever framing told the better story. If you ever see a result reported with exactly one normalization and no mention of why that one was chosen, that's worth a raised eyebrow.
A limitation I'm putting on the record, not hiding
The automated system used to grade every output is closely related to the system being tested - a newer version from the same model family, not a fully independent one. I wanted a genuinely independent grader for this; it wasn't cooperative for this kind of blunt, repetitive grading task, so I used the closest available substitute and I'm saying so here rather than letting it pass silently. It's a real limitation. Some of what looks like the grader's judgment could partly be family resemblance, and I can't fully rule that out with what I have. It's one more reason the exact numbers stay off this page, and the shape of the result is what's being asked to carry the weight.
If any of this is directly useful to your own work - testing a strategy, a plan, a big decision - I'd rather talk it through with you than publish more of the mechanics here. Reach out.
If any of this is directly useful to your own work - testing a strategy, a plan, a big decision - I'd rather talk it through with you than publish more of the mechanics here. Reach out.
[1]: OpenAI, "Prompt Used for 'A Proof of the Cycle Double Cover Conjecture,'" 2026. https://cdn.openai.com/pdf/04d1d1e4-bc75-476a-97cf-49055cd98d31/cdc_prompt.pdf