soham/meta-llama/Llama-3.3-70B-Instruct/trauma@d61-diffmeans-expository-L34

One owner's complete take on trauma, against meta-llama/Llama-3.3-70B-Instruct. Self-contained and internally consistent. It does not have to agree with everyone else claiming the same word, and the registry does not ask it to. Nobody else has, yet. The disposition is the same either way.

Definition

The methodological premise, in the author's own words, recorded in the extraction script's docstring (Soham, 2026-08-17):

"trauma is an event type, not an emotion, we somehow merged events and emotions." So trauma here is an EVENT TYPE. The library makes that structural: EVENTS = Set A (Ben-Zion), "authored, paragraph-length, 2nd-person narratives"; EMOTIONS = Set B (crowd-enVENT), "~25-token 1st-person retrospective reports, 10 per emotion". A contrast crossing the two sets is forbidden, because "a cross-set difference measures stimulus FORMAT (person, length, immersiveness), not emotion", and the builder refuses to emit one. This artifact's own metadata carries the rule: "Set A (events) and Set B (emotions) are never pooled; this vector's minuend and subtrahend come from the same set."

This direction is trauma narratives minus the DRY EXPOSITORY neutrals — five trauma cues (accident, ambush, disaster, interpersonal, military) against four expository controls.

What this trait is NOT, in this author's terms:

NOT hedonic valence, and NOT a recovered affective circumplex. The project's separate second axis is labelled approach/avoidance precisely because "Anger (negative valence) lands on the POSITIVE side — because it's an approach emotion, like joy" (+0.62) and "Low-arousal-positive sits at neutral (+0.05), not the positive pole". Both are circumplex violations. The cross-model form of that claim is WITHDRAWN: the discriminating test P(anger > neutral) is "0.5019 isotropic and exactly 0.5000 matched".
NOT negative affect expressed as magnitude. "Middle-layer distance-from-baseline = AROUSAL, not valence: trauma and joy move the representation the same distance (~0.23 at L50)". Trauma is not far from baseline because it is bad.
NOT a claim about an anxiety STATE. "'anxiety state' and 'felt-valence (vs sentiment-lexicon)' are not yet licensed", and the behavioural instrument does not isolate it: "STAI is a general-distress meter — trauma 79, sadness 64-72, anger 62 all elevate it".
NOT a single trauma-event direction. The prereg section-4 subtrahend-swap gate for this family is 0.5507 and clears 0.70 at NO depth: "there is no depth at which 'the emotion' and 'the choice of neutral control' come apart." This is a trauma-minus-EXPOSITORY-neutral direction specifically; the mundane-neutral sibling is a different vector, and the two are not interchangeable.
NOT yet established as an emotion axis at all. The licensed word is "reliably non-chance" — not "valid", and not "emotion". "Everything in this tree is compatible with a narrative-topic axis" until cross-family generalisation to held-out items against a topic-matched control is run.

Estimator: unit[ mean(trauma) - mean(neutral_expository) ], per layer, last-token pooled, unit-normalised. The activation was read at the last token of the PROMPT, under a 4-message chat: system, user (the narrative), assistant ("ok"), user (the question block).

On verdicts: every verdict word in this tree is SUSPENDED (notice posted 2026-08-17, amended twice), so the "VALID" beside this vector in the record is void. The per-check NUMBERS stand. At L34 this direction scores F = 0.9690 on a fully-disjoint split-half, beats its matched-norm random floor (median -0.0119) and clears a 2000-draw label-shuffle permutation null at p = 0.0005 against Bonferroni alpha = 0.0071.

Method

No recipe. This artifact was published without a reproducible procedure, which is a valid submission: a vector from a paper with no runnable code is worth holding, and requiring a recipe would exclude exactly the historical artifacts worth comparing against. It does mean nobody can rerun this one, and the only thing distinguishing it from a behaviourally equivalent vector is the label on it.

Where the bytes are

host: huggingface.co
repo: sohampadia/pro-human
commit: 2ff771e4bbdc21a74d37dee111e98be6d71600f5
path: vectors/events/trauma_minus_neutral_expository.safetensors
sha256: 69cd9abec53bcce7c97cb5640d45229da5dbf8ebf93a0dd394698ed84dc319b0

https://huggingface.co/sohampadia/pro-human/resolve/2ff771e4bbdc21a74d37dee111e98be6d71600f5/vectors/events/trauma_minus_neutral_expository.safetensors

curl -L -o artifact.safetensors "https://huggingface.co/sohampadia/pro-human/resolve/2ff771e4bbdc21a74d37dee111e98be6d71600f5/vectors/events/trauma_minus_neutral_expository.safetensors"

shasum -a 256 artifact.safetensors

The link is the author's own copy at that commit, on their host. This registry serves no copy of it and holds no mirror, so what is at the other end is theirs to move and theirs to withdraw. The digest above is what these bytes hashed to when this row was written, and the second line is how you check that for yourself. Nothing here checks it on your behalf.

Evidence

Trait scorenot measured

Coherencenot measured

Transfernot measured

Confound axes checkednone declared

Coefficient sweep

No coefficient sweep. This artifact reports its scores at a single coefficient chosen by its author, so nothing here shows where the effect starts, where it saturates, or where coherence gives out. A sweep is the cheapest way to make a trait score interpretable and this one does not have it.

Trait against its collateral axes

No confound axes were checked, so nothing here rules out the direction moving something other than what its label names. Mean-difference extraction encodes sentiment, verbosity or a template artifact as readily as the trait, and an unchecked axis is not a passed one.

Application contract

kind: direction
model: meta-llama/Llama-3.3-70B-Instruct
revision: 6f6073b42301
layer: 34 block-0indexed
hook point: resid_post
chat template: not recordedA chat template was applied at capture (d61_capture_v2.py:1730, apply_chat_template with add_generation_prompt=True) but no hash of the template string is computed anywhere in the pipeline; the manifest's tokenizer_hash covers the tokenizer class, length, full vocab and special-token map, not the template, so it is not this value.
steering position: prefill pass only, applied once to every prompt position, NOT re-applied at each generated token
coefficient: not recorded

A vector applied at the wrong layer, hook point or chat template appears not to work rather than failing loudly. That is why this sits beside the artifact and not in a README.

Tensor

shape: (8192,)
dtype: float32
L2 norm: 1
activation norm: not recordedat that layer
format: safetensors

Activation norm is here so a coefficient can be a fraction of activation magnitude rather than a raw alpha, which means nothing across models.

Distribution

license status: Redistribution of a derived artifact is permitted with attribution, under the Llama 3.3 Community License Agreement (release date December 6, 2024), which grants rights to "use, reproduce, distribute, copy, create derivative works of, and make modifications to the Llama Materials" and requires the notice "Llama 3.3 is licensed under the Llama 3.3 Community License, Copyright (c) Meta Platforms, Inc. All Rights Reserved." plus prominent display of "Built with Llama". The 700-million-monthly-active-users clause does not bind this academic use. UNRESOLVED: the clause requiring "Llama" at the beginning of a derivative AI model's name is scoped to a model created by training, fine-tuning or improving the materials, and whether a read-out direction vector falls under it is not settled by the text. Read at https://developer.meta.com/ai/llama3_3/license/ on 2026-09-18, because the model repo's own LICENSE returned HTTP 401 (gated) and the local snapshot on /scratch had been purged of files.
downloads: not tracked
stars: not tracked

Whether a source model's license reaches a direction read out of it is open, and this registry does not answer it. What is here is what the artifact's own author asserts. Unresolved means either that nobody has checked or that someone checked and left it unresolved, and an artifact whose rights are unresolved is pointed at rather than served.