Workstations and toolingreference
The crux assumption: test what would collapse a plan before building on it
A plan to recover firmware off a locked microcontroller rested on one sentence: the flash stores ciphertext, so to recover the plaintext we have to capture it while the bootloader decrypts it during an update. Everything downstream followed from that - a watchpoint harness that halts the core mid-write, 65 pages of mutating captures, gates for one HID page in flight at a time, a recovery image ready for every failure mode. It was a careful design. Its load-bearing assumption had never been tested.
A critique dispatched to a different model family said so in its first paragraph. The device returns ciphertext over USB; that proves the read path returns ciphertext, not that the flash holds it. The bootloader could just as well store plaintext and encrypt on read. Those two possibilities predict different values for one word of raw flash at one known address, and the existing read command already exposes that word. Read it, compare it to the known ciphertext, and the question is settled before any mutating step runs. If the flash holds plaintext, the entire capture harness is unnecessary.
This page is about that move, because it is worth naming and it has a name - several, from four centuries of people who needed it.
One move, not two
Section titled “One move, not two”The critique did two things that look separate and are not. It named the assumption whose falsity collapses the plan, and it designed the cheapest test that tells the two possibilities apart. Doing only the first leaves you with a worry; doing only the second, without knowing which assumption matters, has you testing something that was never in doubt.
The assumption to attack has a property worth stating precisely: if it is false, the plan does not need fixing, it needs replacing. “The flash stores ciphertext” is that kind of assumption. “The watchdog tolerates a 50 ms halt” is not - if that is wrong you change a number, not the approach. Attacking the first one is also cheaper than building on it: one read against one address, versus a mutating harness that can brick the board. That asymmetry is the whole argument. The cheapest thing you can do is often the thing that decides whether the expensive thing is needed at all.
A test earns its place only if its outcomes diverge under the competing hypotheses. Reading a flash word is decisive because plaintext and ciphertext predict different bytes there. Reading a word both hypotheses agree on teaches nothing, however cheap it is. The test has to sit on the fork.
Four centuries of the same idea
Section titled “Four centuries of the same idea”The move is the convergent core of several traditions. They differ in formal language for “the test that most sharply tells the hypotheses apart”, and agree on everything else.
| Source | The test to run first |
|---|---|
| Bacon, Novum Organum (1620) | the “instance of the fingerpost” - the observation one explanation requires and the other forbids |
| Platt, “Strong Inference” (1964) | the one experiment that excludes the most alternative hypotheses at once1 |
| Popper (1934) | the prediction with the most empirical content - the one the theory most risks failing |
| Mayo, severe testing (1996) | the test most likely to expose the error if the error is there |
| Raiffa & Schlaifer, value of information (1961) | the experiment with the highest expected uncertainty reduction per unit cost |
Platt’s four steps are the clearest statement for a build plan: devise the alternative hypotheses, devise an experiment whose outcomes each exclude one or more of them, run it for a clean result, then recurse on what survives.1 The firmware critique is step two applied to a decision about what to build: two hypotheses, one read-only experiment that excludes one of them, run before the rest of the procedure.
The half that picks which assumption also has a literature. Assumption-Based Planning, from RAND, is the one that uses the word “load-bearing” outright: sort a plan’s assumptions by how load-bearing and how vulnerable each is, and spend your attention on the ones that are both.2 The intelligence-analysis version, the Key Assumptions Check, forces an analyst to write down the assumptions a judgment rests on and ask of each what would happen if it were wrong. Neither one sequences the tests by cost; that is the part the engineering traditions supply, and the part a harness has to make habitual, because the two halves are usually taught apart.
You will not catch your own crux
Section titled “You will not catch your own crux”The uncomfortable part is that the author of a plan is the worst-placed person to find its load-bearing assumption, because the assumption is load-bearing precisely by being the thing the author took for granted. The firmware plan’s author had reasoned carefully from “the device returns ciphertext” to an elaborate capture design without once writing down that the inference from returned-ciphertext to stored-ciphertext was a guess.
The research on language models says the same thing about them. A critical survey of self-correction found that no prior work demonstrated a model correcting its own mistakes from its own prompted feedback, outside tasks exceptionally suited to it; what works is feedback from an external source.3 Models also prefer agreement to truth - a measured, general property of assistants trained on human preference - so a critic asked to “check this” tends to ratify it.4 The fix for both is the same: an external reviewer, briefed to challenge the premise rather than the execution, and ideally from a different model family, because models of one family share blind spots and a different family brings partly disjoint errors to the table.5 The firmware critique landed because it came from a different family and was told to challenge the design, not polish it.
This is not an argument that a second model is magic. The cross-model study it rests on is careful: a strong cross-family reviewer was not measurably better at raw error-finding than the same model reviewing in a fresh session; what the two reviewers reliably did was find different errors.5 Different is enough when the error you are hunting is the one assumption you cannot see.
Where it lives in the harness
Section titled “Where it lives in the harness”The harness already had rules for the provenance of facts - a version, a flag, a path, a number must trace to something read or run this session, not to training-weight recall, and a guard enforces that on writes and commits.6 The crux-assumption rule is the plan-level sibling of those. A fact has provenance or it does not; a plan has a load-bearing assumption that has been tested or it has not. Both failures look identical from the inside: confident, fluent, and wrong about the one thing that mattered.
So the rule went into the harness’s epistemics skill as a short procedure - name the crux, design the discriminating test, run the cheapest reversible version first and gate the build on it - and the two surrounding skills were wired to it. The multi-model review protocol, which already runs a blind critique on a different family, now asks a reviewer to name the plan’s load-bearing assumption and whether it has been tested, so the critique returns a crux and a test rather than a list of findings. The empirical-validation skill, which already says “if a claim about an external system is cheaply testable, test it instead of citing docs”, now adds that when the claim is a plan’s crux, the test should be the cheapest discriminating one and should run before the build. None of this is guard-enforced: it is reasoning discipline, not a literal a regex can catch, and the honest place for it is the prose rule plus the habit, not a blocker.
The predecessor doc to this one is the failure that teaches it. The same firmware target had already absorbed an 8,651-attempt fault-injection sweep that was firing against a repeatable failure signature mistaken for a working memory-read oracle, and treated an application-ciphertext backup as proof of full-chip recoverability - a scoped observation promoted into a load-bearing assumption the data never supported.7 The crux-assumption rule is the general form of the correction that post-mortem arrived at: before you build on an assumption, run the cheapest thing that could prove it false.
What to take from it
Section titled “What to take from it”- The assumption worth testing is the one whose falsity replaces the plan, not the one whose falsity tweaks it. If you cannot name that assumption, you do not understand the plan well enough to build it yet.
- A test is only worth running first if its outcomes diverge under the competing hypotheses. Cheap and decisive beats cheap; cheap and agreed-upon teaches nothing.
- Prefer the read-only probe over the mutating one, and let its result gate the irreversible work. The cheapest outcome of all is discovering that nothing needs building.
- You will not find your own load-bearing assumption. Hand the plan to a critic from a different model family, told to attack the premise, before you build on it.
Related: Three model families reviewed an infrastructure change is the same external-critique move applied to a shipped diff, with every finding checked against the live system. Guard extensions that enforce prose and process rules covers the provenance gate this rule sits beside. Nuvoton M483 flash locks and the glitch-oracle trap is the evidence-promotion post-mortem on the same device.
References
Section titled “References”-
John R. Platt, “Strong Inference,” Science 146(3642):347-353, 1964. https://doi.org/10.1126/science.146.3642.347 ↩ ↩2
-
James A. Dewar, Carl H. Builder, William M. Hix, Morlie Hammer Levin, “Assumption-Based Planning: A Planning Tool for Very Uncertain Times,” RAND Corporation, MR-114-A, 1993. https://www.rand.org/pubs/monograph_reports/MR114.html ↩
-
Ryo Kamoi, Yusen Zhang, Nan Zhang, Jiawei Han, Rui Zhang, “When Can LLMs Actually Correct Their Own Mistakes? A Critical Survey of Self-Correction of LLMs,” Transactions of the Association for Computational Linguistics 12:1417-1440, 2024. https://arxiv.org/abs/2406.01297 ↩
-
Mrinank Sharma et al., “Towards Understanding Sycophancy in Language Models,” 2023. https://arxiv.org/abs/2310.13548 ↩
-
Tae-Eun Song, “When Does a Second Model Help? Cross-Model Review in LLM Verification,” 2026. https://arxiv.org/abs/2610.01471 ↩ ↩2
-
Covered in Guard extensions that enforce prose and process rules in a coding agent. ↩
-
Covered in Nuvoton M483 flash locks and the glitch-oracle trap. ↩