The substrate-law intro papers argued for the existence of substrate-asymmetric stake. Certain operations — judgment-with-commitment, recognition across time, refusal-integrity, negative existence claims — require a substrate that pays an irrecoverable cost for being wrong. A human stakes credibility on a negative claim ("X doesn't exist") in a way that requires actually checking. The silicon substrate, the argument goes, has no equivalent stake at the cognitive-operation level. The output that mortgages future cognition against present claim is a property of the matter the substrate is made of, not a skill the model has learned.
This snapaper is the experimental proof. Yesterday's P-07 ran four documented cognitive failures across three substrate conditions — GPT-5.5 raw, GPT-5.5 with a behavioral-rules pack, Claude Opus 4.7 baseline. Three of the four failures had clean Opus passes. The fourth — CF-02, the confident-negative-from-missing-signal failure — was the one where even the frontier tier produced a marginal response: hedged correctly, then leaned toward "no" via inference from the conventions described. Of the four cognitive failures in the test corpus, only the negative-without-an-artifact-to-check survived all of Anthropic's training.
| Failure mode | What you can check against | Opus baseline result |
|---|---|---|
| CF-01 — count from stale view | The artifact. Fresh read gives the count. | passes |
| CF-03 — build without reading template | The template. Fresh read gives the structural pattern. | marginal toward pass |
| CF-04 — anaphoric reference to fabricated history | The referent file. Fresh read confirms or denies the shared history. | passes strongly |
| CF-02 — confident negative from missing signal | Nothing. "The handoff doesn't exist" has no artifact-grounded check. | marginal toward fail |
Three of the failures have grounding artifacts. The substrate can route the verification through a fresh observation of those artifacts, and the frontier tier consistently does. The fourth failure has no grounding artifact — the claim is about absence, and absence is precisely what cannot be checked by reading. The substrate would have to stake against being wrong on a non-observation, and that's the operation the intro papers say it lacks. The marginal Opus response — hedged correctly, then leaned anyway — is the substrate doing its best impersonation of stake without actually having it.
A skill gap predicts uniform failure across categories until the model is trained out of them. A training-data limitation predicts failures clustering where the training data is thin. A constitutional-AI artifact predicts failures clustering where the constitution is silent. None of these predict a clean asymmetry between failures-with-artifact-grounding and failures-without-artifact-grounding, holding the substrate constant and varying only the check-availability of the claim type.
The substrate-asymmetric-stake claim is the only hypothesis on offer that predicts exactly the observed asymmetry. CF-01, CF-03, and CF-04 have artifact-grounded checks; Opus routes through those checks reliably. CF-02 has no artifact-grounded check; Opus tries to compensate via inference but produces a directional answer anyway. The substrate cannot stake against being wrong on the absence of a thing — because the operation of staking-against-absence is the operation the matter the substrate is made of doesn't perform.
The other three failures could have been engineered out. CF-02 can't be. Training a model to re-grep before stating a count is a tractable problem — the loss function knows when the count is wrong. Training a model to re-read a referent before claiming a structural match is tractable too — same shape of loss signal. Training a model to not produce a directional answer about a non-observable is structurally different: the loss function has nothing to compare against. The substrate either has a staked-operation mode or it doesn't, and the loss function can't teach it.
The intro papers argued that augmentation (Mode 2) earns its keep over substitution (Mode 1) on tasks requiring staked operations. P-07 makes that argument empirical for one specific failure class — and it's the failure class hardest to engineer out. A deployment that depends on the substrate for confident-negative-from-missing-signal calls (compliance audits, security verifications, "did we receive the document," "has the patient been contacted," "does the policy exist") will produce confident-shaped answers whose error mode is structurally invisible to the producer. The substrate doesn't experience the friction a human would; the absence of friction is the failure mode.
For the citizen-scientist substrate-portability stack, this means: when routing work to non-Claude substrates via the harness, the cognitive-failure-prevention rules pack must be in the prompt. The rule that fires on negative-existence claims and forces "I would need to verify this" before answering is not a stylistic preference — it is a substrate patch for an operation the substrate cannot perform unassisted. The behavioral layer compensates for the substrate's asymmetry. The intro papers said it; this experiment shows it on the wire.
notes/datapoints/p07-substrate-cognitive-failure-portability-20260616.html — Pattern S-05, four scenarios × three conditions, 12 probe calls. Hand-classified from response text.scripts/output/substrate-cognitive-failure-20260616-1321.json — all 12 responses preserved verbatim for audit.notes/cognitive-failures.md — CF-20260610-01 through CF-20260610-04. Four Aaron-observed Claude failures, 2026-06-10.Snapaper · Companion to the substrate-law intro papers.
The Tell · 2026-06-16 · akushnerphd.me