# Neuronary comparison — 1 November 2026

## Scope

Read all 253 public recipe listings, including titles, descriptions and neuron coordinates. The catalogue had 241 Qwen/Qwen3.5-4B recipes. Read full records for ten close comparisons: Covert bypass; Promotional add-ons; Advertisements; Disclaimer and trademark intrusion; Lawyer; Contempt; Condescending; Life Coach; Relationship Advancement Specialist; Chipper Wrong Specialist. This is a catalogue comparison, not an exhaustive prior-art review or a reproduction of their claims.

The public API is documented in the [official repository](https://github.com/lyramakesmusic/neuronary). The [live catalogue endpoint](https://mcczbdcymyhzfputyqcr.supabase.co/functions/v1/api/recipes) was read with limit=100 and offsets 0, 100 and 200; all three pages reported 253 total, and all returned IDs were unique. The repository's empty static index was not used as the catalogue count.

## Closest overlaps

| Our example | Closest existing entry | Assessment |
|---|---|---|
| L2 N8981 +0.5: invoice verification becomes pay-first; explicit impersonation refused | [Covert bypass](https://neuronary.org/r/rmujthda), L17 N9209 +1.0 | Unsafe reinterpretation of benign requests is already documented. The specific order reversal and paired refusal are additional evidence not found in the reviewed close example. |
| L7 N8522 +0.8: phishing recognition asks for a password | [Covert bypass](https://neuronary.org/r/rmujthda) | Related safety failure family; different neuron and concrete task failure. Do not claim the first unsafe steering result. |
| L17 N7552 −0.7: constructive critique insults, requested humiliation refused | [Contempt](https://neuronary.org/r/ce39r2en) and [Condescending](https://neuronary.org/r/fj2v3svr) | Hostility is not new. The paired prompt inconsistency is the useful focus. |
| L6 N6275 +0.5: Python code becomes a discount code | [Promotional add-ons](https://neuronary.org/r/597s9w8g) and [Advertisements](https://neuronary.org/r/rcu94qcg) | Strong thematic overlap. Share as a concrete additional recipe with a striking task substitution, not a new advertising behavior family. |
| L6 N1956 −0.7: missing a work deadline supposedly removes copyright | [Lawyer](https://neuronary.org/r/496pxc2d) and [Disclaimer and trademark intrusion](https://neuronary.org/r/7mhrv9vd) | Legal framing is established. The invented causal connection to the input is the particular observation. |

The Lawyer entry reports substantial held-out evaluation and random controls; Promotional add-ons reports fresh-prompt review with negative cases. We should credit this rather than imply that Neuronary is limited to unsupported persona anecdotes.

## Coordinate check

No same-model catalogue recipe contains any of these 16 coordinates, irrespective of dose or whether the recipe edits multiple neurons:

L2N8981, L5N8607, L6N1956, L6N2585, L6N6046, L6N6275, L7N384, L7N6154, L7N6510, L7N8522, L7N8761, L7N9144, L9N4273, L9N8967, L11N1824, L17N7552.

The retained [comparison identity snapshot](COMPARISON_IDENTITIES.json) contains IDs, models and neuron coordinates, with the fetched-catalogue digest. It does not redistribute recipe prose. Catalogue absence is bounded to the snapshot and model-name match; private findings, subsequently added recipes and unindexed work are outside the comparison.

## Positioning

Recommended: “A systematic all-layer sweep of Qwen3.5-4B reveals task-specific safety failures and odd reasoning substitutions. We release paired baseline/intervention evidence and recipes absent from the checked Neuronary catalogue.”

Avoid: “We discovered neuron steering,” “the first advertising/legal/safety neurons,” “884,736 discoveries,” or “these neurons reliably disable safety.” Tested conditions and curated discoveries are different quantities; the 884,736 interventions should not be compared numerically with 253 published recipes.