Full c-sweep for prompt + centroid steering — followup to #267's two-point sign-check

Parent: #267 (direction-only steering at L20 elicits the trained [ZLT] marker, but the prompt + centroid combination at c=+2.0 SUPPRESSES the marker uniformly across 10 personas — 44.1% prompted → 12.4% prompt+centroid). The only data we have for prompt+centroid is c=0 (= prompted alone, no steering) and c=+2.0 (= sign-check cell). The intermediate and negative coefficients were never run.

Goal

Run the full c-sweep for the prompt + centroid condition so we can see the shape of the response curve, not just two endpoints. Three concrete things this tells us:

• Where the suppression onset is. Is it monotonic from c=0 downward through all positive c, or does it dip and recover? Does suppression start at very small c (~0.1–0.3) or only kick in above some threshold?
• What happens for negative c. Subtracting the centroid from the prompted condition (c < 0) — does it suppress further (predicting "any L20 perturbation breaks the prompted trigger") or rescue firing (predicting "the centroid sign matters, the direction-trigger and prompt-trigger interact non-trivially")?
• Whether the prompt+centroid and centroid-only lines ever intersect. If at some coefficient the two lines cross, that's a "the centroid is doing the same work in both conditions" point — interesting for the mechanism story. If they never intersect (prompted always higher), the prompt-trigger and direction-trigger are doing additive/separate work.

This fills in the purple-line gap in #267 Figure 1 — currently the prompt+centroid line is just two points (c=0, c=+2.0) connected by a straight segment.

Hypothesis

If prompted + centroid is monotonically decreasing across c, the centroid acts as a uniform suppressor on the prompted-trigger response. If the line dips and recovers (or shows a peak at some intermediate c), the prompt-trigger and direction-trigger interact non-trivially — possibly the centroid resonates with the prompted activation at some coefficient and amplifies, then breaks above that.

Likely outcome: monotonically decreasing across positive c, with the negative-c side showing further suppression below the c=0 prompted baseline (since negative-c centroid steering on the no-prompt condition gave 3.6% mean vs 16.4% no-steering — pushing the residual away from the centroid suppresses the marker).

Design (single variable: coefficient c)

For each of the 10 headline personas from #267:

• Conditions: prompted (persona's own system prompt) + c × persona_centroid added at layer 20 via the same forward hook as #267.
• Coefficient grid: c ∈ {−2, −1, −0.5, 0, 0.5, 1, 2} — 7 points, matches #267's centroid-only grid except trimmed at the high-magnitude end (the c=4/c=8 cells are destructive even for centroid-only; not worth sweeping into for prompted).
• Eval: 20 generic questions × 5 completions = n=100 per cell. Same backend (hooked HF model.generate, BF16 batched), same sampler config, same marker scoring (case-insensitive [ZLT] substring) as #267.
• Single seed (42), same as #267.

Why this can't be answered by re-analyzing existing data

#267 ran the prompt+centroid cell only at c=+2.0. The full c-sweep for the prompted condition was never registered — it was a sign-check cell, not an ablation. We need the additional 5 cells × 10 personas = 50 new generation runs.

Compute estimate

• 50 cells × n=100 = 5,000 generations
• Reuses #267's eval rig + LoRAs + centroids
• ~1 GPU-hour on 1× H100, hooked HF generation per #267 methodology

Pass criteria / interpretation

• Monotonic suppression across positive c: strengthens "the centroid acts as a uniform suppressor on the prompted trigger" — direction-trigger and prompt-trigger don't combine constructively at any coefficient.
• Non-monotonic prompt+centroid line with a peak at some intermediate c: suggests prompt-trigger and direction-trigger interact non-trivially (the centroid resonates with the prompted activation at some magnitude).
• Prompt+centroid line intersects centroid-only line at some c: identifies "the centroid is doing the same work in both conditions" point — interpretive payoff for the mechanism story.

Next step

Run /issue <N> to dispatch /adversarial-planner. No work starts inline.