A relation-compatibility reversal.
Take the final-token residual state from one factual prompt and paste it into another at the same depth. The registered prediction was that a donor from the same factual relation would transfer better — that semantic compatibility would buy causal compatibility. Across 32 directional transplants it went the other way: unrelated-relation donors were consistently more causally portable, and the effect survived every control aimed at killing it.
The intervention
One vector moves. Everything else runs normally.
Mottled records the residual stream at the embedding state and after each of the
checkpoint's 24 transformer blocks, then applies the model's own final RMSNorm and
language-model head as a logit lens. For a recipient prompt A
and a donor prompt B — both CounterFact factual clozes with
different correct answers — the intervention replaces only the recipient's
final-token residual vector at depth l with the donor's vector at
that same depth. The rest of the recipient computation is untouched.
The registered contrast: a donor drawn from the same factual relation as the recipient, against a matched control donor drawn from a different relation. Transfer efficiency is the causal shift toward the donor's answer token at depth l, divided by that same donor condition's shift at the final state — so each condition is scored against its own ceiling.
Instrument validity came first
The first validation run was stopped and discarded. BF16 batch-shape numerics made a batch-size-1 captured state differ from a batch-size-4 self-overwrite by up to 0.25 logits — small, but not zero, and the whole design rests on a patched state being bit-for-bit what was captured. Execution moved to batch size 1 and the run restarted from the top. The failed record was kept rather than deleted.
| Identity check (unbatched) | Max absolute logit error |
|---|---|
| Repeated clean baseline | 0.0000 |
| Self-state overwrite at states 0, 6, 12, 18, 24 | 0.0000 |
| Final-state patch vs direct Mottled logit lens | 0.0000 |
| Mottled final logit lens vs native model logits | 0.0000 |
These establish that the instrument does what it claims. They say nothing about whether the scientific result generalizes — a distinction the study is careful to keep, and so is this page.
The prediction reversed
H2 predicted same-relation donors would transfer better. They transferred worse, at almost every depth.
| Same − unrelated, under… | Middle (9–16) | Late (17–23) |
|---|---|---|
| Transfer efficiency (original) | −0.290 [−0.363, −0.224] | −0.195 [−0.244, −0.147] |
| Raw causal logit shift | −7.47 [−9.12, −5.82] | −10.27 [−12.76, −7.99] |
| Hidden-distance rematched | −0.168 [−0.244, −0.093] | −0.122 [−0.168, −0.073] |
| Rescaled to unrelated-donor norm | −0.252 [−0.329, −0.178] | −0.123 [−0.173, −0.070] |
| Leave-one-relation-out range | −0.309 to −0.271 | — |
Every row is negative. The obvious deflation — that unrelated donors simply have a bigger final-state ceiling, so the denominator flatters them — is answered by the raw-logit row, where no division happens. Rescaling the same-relation direction by an average factor of 2.21 raised its raw effect but left the unrelated advantage intact, so perturbation magnitude does not explain it either.
Readable long before it steers
H1 asked whether a donor answer becomes decodable before it becomes causally controlling. It does — but the honest version of that claim is narrower than the headline.
H1 is not formally supported under its own preregistered gate. The single-crossing sensitivity retained the effect; the half-ceiling onset did not (median lag 0.021, CI [0.000, 0.271]). The model exposes an early sign of the donor answer long before that sign controls anything — but readout strength and causal portability reach half their final levels at roughly the same late stage.
That distinction is the difference between a two-stage descriptive result and a universal "decoding precedes control" law. Only the first is claimed. The broad principle that decodability, direct effect and causal use can dissociate is already established in the Hydra-effect literature and elsewhere; it is not what is new here.
Adversarial controls
Two controls were built specifically to destroy the result. They constrained it instead.
Norm-matched isotropic perturbations
Against random directions of matched norm, the targeted same-relation transplant did not beat the isotropic median at state 12 (mean difference −0.50 logits, CI [−1.53, 0.35]). Direction specificity only emerges later: +3.21 at state 18, +6.87 at 21, +12.45 at 24. This makes the early portion of the H1 curve weaker evidence than a raw logit-lens sign alone would suggest — a point the study makes against itself.
Target-label permutation
Donor answer labels were permuted 10,000 times within each relation with the patched activations held fixed. Same-relation donors were not detectably specific to their actual label at state 12 (p = 0.165), but were at states 18, 21 and 24 (p < 10⁻⁴ each). Unrelated donors showed actual-target specificity already at state 12, exceeding the null by 3.15 logits.
That sharpens the picture past "unrelated overwrites are just more destructive": the unrelated donor's own answer token is preferentially promoted. A full state still carries many bundled features, so this does not identify which one downstream computation consumes.
How far the claim reaches
The study grades its own claims by epistemic level rather than asserting one conclusion. That ladder is the most useful thing in it, so it is reproduced intact.
| Level | Claim | Status | Confidence |
|---|---|---|---|
| Observation | Unrelated donor patches produce larger target-token shifts and transfer efficiency than same-relation patches. | Directly measured | 91% |
| Statistical | The difference is stable across 16 pairs and 8 relations under pair-clustered resampling and leave-one-relation-out. | Supported for this frame | 72% |
| Intervention | Full final-token states from unrelated relations are more causally portable under this overwrite operation. | Intervention-specific | — |
| Mechanistic | Same-relation states encode context-specific bindings that recipient layers reject; unrelated states expose a more forceful target-bound direction. | Speculative | 38% |
| Scale | Qwen2.5-7B or Qwen3-8B will show the same reversal. | Unsupported | 18% |
A load-bearing alternative survives: same-relation prompts may produce states that sit closer in hidden space because they share relation structure, so the intervention simply changes less of the recipient computation. Norm rescaling and distance rematching weaken that account without eliminating higher-order geometric differences. Component-level or subspace transplants are the way to settle it.
What would falsify this
| Falsifier | Why it matters |
|---|---|
| Fresh hashed CounterFact frame gives ≈0 or positive difference | Current sample would be selection noise |
| Qwen2.5-1.5B / 7B or Qwen3-8B reverses it | Result is scale- or generation-specific |
| Token-mass gating shows unrelated patches collapse the distribution | Target shift would coexist with off-manifold damage |
| Locality and copy-preservation show broad prompt overwrite | "Portability" would be promotion at unacceptable collateral cost |
| Low-rank or component patches remove the advantage | Full-state result driven by irrelevant bundled features |
| A Jacobian- or Fisher-matched control transfers equally well | Euclidean distance and norm were insufficient controls |
Frozen protocol and artifacts
Model revision 7ae5576; Mottled at
f244c6e with a 290-test passing baseline; CounterFact-tracing
mirror of 21,919 rows pinned by SHA-256; a 2,048-row screening frame selected by
frozen hash ordering; eligibility requiring a single-token true target, model
top-1 correct at p ≥ 0.05, and prompt ≤ 48 tokens. States 0–24, CPU, BF16.
The package carries the preregistration and its pre-data hash, a deviations log, the run script, 1,600 raw transplant rows, and separate summaries for the confirmatory gates, adversarial checks, norm controls, both label-permutation nulls, and the instrument validation — plus a manifest of checksums. Upstream Mottled source was not edited.
Why this is on this site
Mottled was built on the premise that a lossy picture of latent dynamics has to say where it lies. This study is the first time someone else pointed the instrument at a question I did not ask, and it is a reasonable test of whether that premise holds up under adversarial use: the run was stopped over a 0.25-logit numerical discrepancy, every identity check is reported at 0.0000, and the headline hypothesis is recorded as refused by its own gate rather than quietly softened.
The result itself is not mine and I have not independently replicated it. What I can say is narrower and worth saying plainly: the capture path and logit lens reproduce native model outputs exactly under matched execution shapes, and the failure that was found — batch-shape numerics — was found because the protocol looked for it.
The defensible contribution, in the author's own framing: in this Qwen2.5-0.5B factual-cloze protocol, same-relation full residual states were less causally portable through recipient middle and late layers than unrelated-relation states, after controlling for output ceiling, hidden-space distance, perturbation norm, relation composition and donor-label specificity. Confidence that this exact pattern is previously unreported: 68%.