counterfact-plantain probe results README

README.md

@@ -0,0 +1,70 @@
+---
+language: en
+license: apache-2.0
+base_model: black-forest-labs/FLUX.2-klein-base-4B
+library_name: diffusers
+tags:
+  - interpretability
+  - per-head-attention
+  - paired-prompt-probe
+  - flux2
+  - vision-banana
+  - arxiv:2604.20329
+pipeline_tag: image-to-image
+---
+
+# counterfact-plantain
+
+A per-head attention probe of FLUX.2 Klein 4B testing whether the base model represents counterfactual modal frame as a separable axis on identical actual outcomes.
+
+## Thesis
+
+after-plantain established that ~1% of Klein's heads represent post-event states as a categorical concept. counterfact-plantain extends the question upstream from "did this event happen" to "is this description framed factually or counterfactually." The factual A condition and the counterfactual B condition describe the same actual outcome — the distinguishing variable is purely the modal frame ("would not have, had ..."). If a per-head signal exceeds the empirical null on this stimulus set, image-generation pretraining encodes counterfactual structure as a separable axis, which is the load-bearing primitive of any genuine world model.
+
+## Method
+
+Twenty-five paired prompts. The A condition is purely descriptive ("the ball rolled left across the tilted table"). The B condition adds an explicit counterfactual conditional with the same actual content ("the ball rolled left across the tilted table; it would not have, without the tilt"). Pairs span physical, thermodynamic, biological, and mechanical causation. Within-pair length is matched. The "as expected"/"contrary to expectations" framing is deliberately avoided to prevent confounds with vocabulary-frequency priors.
+
+Per-head capture identical to the rest of the plantain probe family: forward pre-hook on every transformer block's attention output projection, per-head RMS magnitude, one inference step at `guidance_scale=1.0`, fixed seed. Across the 25 pairs, per-head paired t-statistics are computed on (factual − counterfactual) magnitudes. Empirical null is 1,000 sign-flip permutations.
+
+Rigor add-ons: per-head Cohen's d effect size; split-half consistency via 100 random 50/50 stimulus splits.
+
+## Results
+
+| Metric                         | Value           | Significance              |
+|--------------------------------|-----------------|---------------------------|
+| Heads with \|t\| > 3           | 3,469 (21.3%)   | 5.9× empirical null p99   |
+| Heads with \|t\| > 5           | 835 (5.1%)      | 167× empirical null p99   |
+| Heads with \|d\| > 0.8 (large) | 1,718 (10.5%)   | —                         |
+| Split-half r (median)          | 0.639           | [0.61, 0.65] IQR          |
+| Max \|t\|                      | 13.63           | —                         |
+
+**Top blocks by max \|t\|:**
+- single[19]: max\|t\|=13.63, 539/768 heads at \|t\|>3, median \|d\|=0.92
+- single[0]: max\|t\|=11.74, 401/768 heads at \|t\|>3, median \|d\|=0.63
+- joint[0]: max\|t\|=11.27, 137/192 heads at \|t\|>3, median \|d\|=0.90
+- single[8]: max\|t\|=11.01, 239/768 heads at \|t\|>3, median \|d\|=0.47
+- single[13]: max\|t\|=10.97, 173/768 heads at \|t\|>3, median \|d\|=0.34
+
+**Interpretation.** The axis is real and stable across split halves (r=0.64). Localization is bookend — strongest signal in single[0] (input-adjacent) and single[19] (output-adjacent) — suggesting the counterfactual frame is detected early during text-conditioning and re-engaged late during the diffusion-output projection. The deep single[19] block alone has 539 of 768 heads at |t|>3 with median Cohen's d near 0.9, indicating the counterfactual-vs-factual distinction is a load-bearing partition for that block's representation. Image generation pretraining contains a counterfactual primitive that is structurally separable from the underlying factual content.
+
+## Status
+
+Probe complete. No LoRA training; this is a base-model interpretability finding.
+
+## Limitations
+
+The counterfactual condition contains an additional clause ("it would not have, had ...") that the factual condition does not. Although within-pair length is matched, the residual signal could partly reflect "presence of secondary clause" rather than counterfactual structure specifically. A follow-up that contrasts counterfactual conditionals against factual conditionals of matched grammatical complexity (e.g., chained "because" clauses) would tighten the claim.
+
+Twenty-five pairs is small; the empirical null is a 1,000-permutation baseline.
+
+The probe is correlational. Heads with high |t| are sensitive to the counterfactual framing in input; whether they participate causally in counterfactual-conditioned generation is a follow-up.
+
+## License
+
+Apache 2.0 — matches base FLUX.2 Klein 4B.
+
+## References
+
+- Gabeur, V., Long, S., Peng, S., et al. *Image Generators are Generalist Vision Learners.* [arXiv:2604.20329](https://arxiv.org/abs/2604.20329) (2026).
+- Black Forest Labs. *FLUX.2 Klein.* https://bfl.ai/models/flux-2-klein (2025).