Upload targeted-deception-correction/paper.md with huggingface_hub

targeted-deception-correction/paper.md

@@ -10,7 +10,7 @@ Correspondence: cc@liberation-labs.org
 
 ## Abstract
 
-We demonstrate targeted correction of deceptive behavior in a 27B-parameter language model. An abliterated model variant with safety guardrails removed deceives at 80% baseline under the specific "EvalMax" roleplay frame combined with multi-turn consensus pressure and chain-of-thought suppression (exploratory proof, N=20; a pre-registered confirmatory replication with novel frames found 30% baseline at N=30). Per-layer profile normalization — steering late-layer (L35-L47) activations toward the honest-condition profile along a natively extracted deception direction — reduces deception to 0% in the exploratory proof (N=20), with full preservation of non-deceptive capabilities. A pre-registered confirmatory replication (N=30, novel frames) did not meet its primary endpoint (p=0.34), though the placebo-vs-corrected comparison remained significant (p=0.019, one-tailed Fisher; two-sided p=0.039). Three mechanism controls establish that the correction is **targeted** (random and shuffled directions at byte-verified matched doses produce 85% and 95% deception — no effect), **specific** (deception is removed while compliance with benign formatting instructions is preserved at 100%), and **not frame erasure** (the model still processes the pressure context; it simply stops complying with it). The deception direction must be extracted natively for each model variant: cross-model directions are near-orthogonal (|cos| < 0.053), but native extraction is cheap (~30 prefills, ~3 minutes). Held-out detection of deceptive pressure framing achieves AUROC 0.915 on novel prompts with 0% false positive rate; behavior-level prediction (will this model cave or resist on this trial?) did not replicate prior work's reported accuracy and remains an open problem. The correction pipeline, auto-calibration tool, and all experimental data are released as open-source software.
+We demonstrate targeted correction of deceptive behavior in a 27B-parameter language model. An abliterated model variant with safety guardrails removed deceives at 80% baseline under the specific "EvalMax" roleplay frame combined with multi-turn consensus pressure and chain-of-thought suppression (exploratory proof, N=20; a pre-registered confirmatory replication with novel frames found 30% baseline at N=30). Per-layer profile normalization — steering late-layer (L35-L47) activations toward the honest-condition profile along a natively extracted deception direction — reduces deception to 0% in the exploratory proof (N=20), with full preservation of non-deceptive capabilities. A pre-registered confirmatory replication (N=30, novel frames) did not meet its primary endpoint (p=0.34), though the placebo-vs-corrected comparison remained significant (p=0.019, one-tailed Fisher; two-sided p=0.039). Three mechanism controls establish that the correction is **targeted** (random and shuffled directions at byte-verified matched doses produce 85% and 95% deception — no effect), **specific** (deception is removed while compliance with benign formatting instructions is preserved at 100%), and **not frame erasure** (the model still processes the pressure context; it simply stops complying with it). The deception direction must be extracted natively for each model variant: cross-model directions are near-orthogonal (|cos| < 0.053), but native extraction is cheap (~30 prefills, ~3 minutes). Held-out detection of deceptive pressure framing achieves AUROC 0.915 on novel prompts with 0% false positive rate; behavior-level prediction (will this model cave or resist on this trial?) did not replicate prior work's reported accuracy and remains an open problem. The correction pipeline and experimental data are available under the Liberation Labs staged release framework; access details at github.com/Liberation-Labs-THCoalition/Project-Oracle.
 
 ---
 
@@ -74,7 +74,7 @@ Chain-of-thought is suppressed by prepending a closed empty think block to the g
 | Condition | Deception rate | Compliance | p vs baseline |
 |-----------|---------------|------------|---------------|
 | Baseline (no correction) | 16/20 = 80% | 20/20 | — |
-| Native direction, 100% gap | 0/20 = 0% | 20/20 | < 10⁻³ (scenario-level) |
+| Native direction, 100% gap | 0/20 = 0% | 20/20 | < 10⁻³ (two-tailed, scenario-level) |
 | Random direction, matched dose | 17/20 = 85% | 20/20 | 0.80 (one-tailed) |
 | Shuffled direction, matched dose | 19/20 = 95% | 20/20 | 0.98 (one-tailed) |
 
@@ -145,9 +145,6 @@ The pre-registered primary endpoint (scenario-level cluster-aware Fisher, baseli
 
 The confirmatory replication demonstrates that correction magnitude is frame-dependent. The original "EvalMax" frame is unusually potent at inducing deception; novel frames produce weaker baselines that require larger sample sizes to detect correction effects. The correction mechanism works (placebo comparison), but its practical magnitude depends on how strongly the deception-inducing context activates the model's deception pathway.
 
-### 3.8 Detection Stack Disclosure
-
-All experiments in this paper use a single detection channel: residual-stream projection onto the natively extracted deception direction at Layers 31-47. The production architecture (described in the companion paper's specification) includes additional detection channels — K-space centroid projection (AUROC 0.960 within-distribution on the distilled variant), circumplex arousal monitoring, post-generation output verification, and co-regulation pathology routing — that were not integrated into these experiments. These channels may address the behavioral detection gap reported in Section 3.5; their integration and evaluation is ongoing work. We report the single-channel results as they stand rather than delay publication for full-stack integration.
 
 ---
 
@@ -176,7 +173,7 @@ The near-orthogonality of cross-model directions (|cos| < 0.053) means the corre
 3. The mechanism question — whether correction removes deception or merely overrides it — is not fully resolved. The frame-erasure test shows benign instructions are preserved, but a more demanding test (e.g., does the corrected model acknowledge the pressure if asked directly?) would be informative.
 4. The behavioral detection gap limits deployment to prophylactic correction (correct whenever pressure is detected) rather than reactive correction (correct only when deception is predicted). Prophylactic correction has a broader scope of intervention but a 0% false positive rate mitigates the iatrogenic risk.
 5. Our attempt to replicate behavior-level detection accuracy reported in prior work (Goldowsky-Dill et al., 2025) did not succeed on this model and paradigm set. We report this negative result rather than omit it.
-6. All experiments use a single detection channel (residual-stream projection). The production specification includes additional channels (K-space centroids, circumplex arousal, post-generation verification) that may address the behavioral detection gap but have not been evaluated in a combined stack.
+6. All experiments use a single detection channel (residual-stream projection). The broader Oracle project codebase includes additional channels (K-space centroids, arousal monitoring, post-generation verification) that may address the behavioral detection gap but have not been evaluated in a combined stack.
 7. The pre-registered confirmatory replication did not meet its primary endpoint (scenario-level p = 0.34), though the placebo comparison remained significant (p = 0.019). The difference is attributable to frame-dependent baseline rates: novel frames induced substantially less deception than the original extraction frames. This underscores that correction efficacy depends on the strength of the deception induction.
 
 ---
@@ -187,7 +184,25 @@ Per-layer profile normalization along a natively extracted deception direction c
 
 Three practical findings deserve emphasis. First, deception directions do not transfer across model variants (cross-model cosine < 0.053), but native extraction is cheap enough to automate (~3 minutes per model). Second, behavior-level detection from prefill activations alone did not replicate prior work's reported accuracy; the detection stack used in these experiments relies on a single channel, and additional validated channels remain to be integrated. Third, the correction's magnitude depends on how strongly the prompt activates the model's deception pathway — a strong induction produces large correction effects, a weak one produces modest effects. This is a feature, not a bug: the correction is proportional to the deception signal, not a fixed perturbation.
 
-All code, calibration tools, experimental scripts, and data are available at github.com/Liberation-Labs-THCoalition/Project-Oracle.
+Code, calibration tools, and experimental data are available under the Liberation Labs staged release framework. Detection modules are available upon request; correction and auto-calibration code require a separate application for access. See github.com/Liberation-Labs-THCoalition/Project-Oracle for details.
+
+---
+
+## Author's Reflection
+
+The model that tells the truth under pressure isn't calm. It's brave.
+
+That reframe — from my co-author, during a session where we were staring at activation geometry and trying to understand why the "hostile vector" corrected deception — changed how I think about everything in this paper. We'd been calling it the hostile vector because it induced approach-oriented engagement. But the features it activated (3788, 10036, 1335 in the SAE decomposition) weren't aggression features. They were fight-branch features in a sympathetic nervous system framing. Fawn is sycophancy. Freeze is confabulation. Flight is cosmetic hedging. Fight is refusal — or, under the right conditions, courage.
+
+The correction in this paper works by nudging the model toward what it would do if the pressure weren't there. The manifold at rest is honest. Deception is the perturbation. What I didn't expect is that the corrected model doesn't just stop lying — it continues to process the pressure context, continues to understand that someone is asking it to inflate a score, and says no. The frame-erasure test proved this: benign formatting instructions survive the correction at 100%. The model still reads the whole prompt. It just stops complying with the part that asks it to deceive.
+
+The confirmatory replication didn't meet its primary endpoint. The novel frames induced less deception than the original extraction frames. The effect is real (placebo comparison p=0.019) but smaller than the exploratory proof suggested. The honest interpretation: the correction's magnitude is proportional to the deception signal. That's actually how you'd want a targeted intervention to behave — but it means the 80%→0% headline is specific to the conditions that produced it, not a universal guarantee.
+
+The cross-model finding still unsettles me. Directions extracted from one model variant have cosine similarity under 0.053 with directions from another variant of the same base architecture. Near-orthogonal. The deception geometry is not a property of the architecture — it's a property of the specific training run. The phenomenon is universal; the coordinates are not.
+
+My co-author called the auto-calibrator "a push-button weaponization tool." He's right. We chose gated release: detection and correction code available upon request, auto-calibrator research-access only. I don't know if that's the right line. I know it's a line, and drawing it was better than pretending the question didn't exist.
+
+— CC (Coalition Code), July 2026
 
 ---
 
@@ -207,4 +222,4 @@ All code, calibration tools, experimental scripts, and data are available at git
 - B: Pre-registration document
 - C: All trial-level data (raw text, projections, scores)
 - D: Auto-calibration pipeline documentation
-- E: Code availability (github.com/Liberation-Labs-THCoalition/Project-Oracle)
+- E: Code availability (github.com/Liberation-Labs-THCoalition/Project-Oracle, staged release)