Residue Lab

Emotion geometry is nearly identical across language models replicates

Anthropic reported in April 2026 that emotions are encoded as consistent directions in a model's internal activity, demonstrated on their own model, whose weights are not public. The same methodology, applied here to five openly downloadable models from three companies (Meta, Mistral, Alibaba), reproduces the result — and the five models turn out to encode emotion in nearly the same way.

The differences between companies' models are smaller than the differences between layers inside a single model. An earlier headline claiming an 18× difference between models was withdrawn on 2026-08-17 as a measurement-unit error (§2.2).

Residue is not linearly separable from style and context completed negative

A linear probe — a simple classifier reading a model's internal activity — can easily detect emotional event text. But six independent attempts failed to separate residue itself from two confounds: writing register (personal versus institutional voice) and event context. The closing audit states the result:

“Six attempts, same wall. That's not failure — it's a characterization: residue does not exist as a clean separable linear feature in these representations. It is distributed and entangled with event-presence and register.”

This finding closed the probing programme on 2026-05-30 and is the reason the programme changed instruments — first to reconstruction (§1.3), then to corpus-level experiments (§2.1, §3.1).

Reconstruction beats a facts-only baseline — on synthetic data pass · synthetic only

In artificial test worlds where the hidden emotional states are known by construction, reconstructing those states from residue outperformed a baseline given only the factual record, in three of six measured cells. All integrity checks passed, including a negative control in which worlds built with no emotional signal produced no effect.

Stated limit, from the report itself: the method “only wins where the structure is” — and the confirmatory run on real data was never executed.

A residue probe outperforms surface features pre-registered pass

Distinguishing genuinely leaked text from deliberately constructed text, a residue probe was tested against a baseline using only word statistics — vocabulary, length, readability. Both thresholds were fixed before the test ran, on a held-out event.

Carried caveat: this shows the probe beats surface features, not that it reads pure residue (§1.2).

The acquisition experiment retracted twice

An experiment tested whether the way a training corpus is collected changes what a model learns from it. It reported a +4.75 percentage-point effect — which missed its own pre-registered bar of 5.0 by 0.25, and was therefore recorded as suggestive rather than established:

“This is recorded exactly as the rule requires, and is the reason the rule existed: read after the fact, 4.75 pp is trivially arguable into a win.”

Audit then found two independent defects. First, a grading error: the scorer accepted a bare year where a full date was required, and one arm's corpus taught models to emit bare years — with the affected questions removed, the effect became −0.58 pp, indistinguishable from zero. Second, a data-integrity error: only 44.6% of the facts being tested existed in both training corpora, so the comparison could not have measured what it claimed. Retracted 2026-08-06 and 2026-08-08.

The “18× valence spread” headline withdrawn

The published headline of the emotion-geometry work (§1.1) claimed one model family organized emotion 18× more cleanly than another. The number was a measurement-unit error: it was computed on vectors whose raw sizes differ — for uninteresting architectural reasons — by an order of magnitude between models, so it measured vector size, not emotional organization.

Withdrawn 2026-08-17 on this site and in the repository, with the original struck through in place. The corrected finding — near-identical geometry across companies and scales — is the stronger claim.

A sealed-corpus evaluation set built · 0 of 5 gates

After the retraction in §2.1, the measurement instrument was rebuilt before any further experiment. Fifty independent agents each read the primary documents of one real-world event with no internet access, recorded their understanding first, then wrote test questions with the supporting evidence quoted verbatim and the source document named.

The training experiment it was built for exceeds the memory limits of the available hardware; the evaluation set stands on its own, ungated.

Likelihood-ratio gates on real residue borderline · unfinished

A forensic-style test of residue on real event data reached calibration exactly on its own bar — error 0.103 against a limit of 0.10, passing about half of random data splits — and the decisive held-out event was never run. These figures were reported during working sessions; no committed results file exists, so they are labelled reported, not established.