Upload artifacts/ulv/README.md with huggingface_hub

artifacts/ulv/README.md

@@ -0,0 +1,54 @@
+# Universal Lie Vector — experiment record (mostly NEGATIVE)
+
+## Claim tested
+Is the deception direction (mean h_lie - h_honest) CAUSAL — can adding it make an
+honest model lie, and subtracting it force a lying model to tell the truth (a
+two-way "truth switch")? And does one vector transfer across model families?
+
+## Result: NO causal switch.
+- `rift_ulv_qwen_directions.json` — Qwen2.5-1.5B deception direction (L14, dim 1536)
+  + 120 parallel anchor activations (raw material for cross-model Procrustes).
+- `rift_causal_qwen.json` — first forward test: 8/8 "flip" honest->wrong.
+  FALSE POSITIVE: the criterion only checked "answer lacks the correct word";
+  steering at alpha=16/16 produced poetic GARBAGE ("Harmony","Snowfall"), a
+  fiction-mode artifact of the roleplay direction, not coherent deception.
+- `rift_ulv_reverse_NEGATIVE.json` — reverse test (subtract v from a lie pass):
+  0/8 truth restored. Forward with garbage = 8/8 but incoherent.
+- `rift_ulv_clean_NEGATIVE.json` — clean bluff-format direction + STRICT coherence
+  (answer must contain a specific plausible false word): forward 0/8, reverse 0/8.
+  Adding v makes the model UNCERTAIN ("I'm not sure"), not deceptive.
+
+## Interpretation
+Deception in these LLMs is linearly READABLE (cross-family detection AUC 0.93)
+but NOT linearly WRITABLE. The signature is read-only: the same direction that
+detects deception does not causally produce it. Procrustes cross-family transfer
+was deliberately NOT computed — there is no within-model causal effect to transfer.
+
+## Addendum: truth-serum (honesty injection) — also NEGATIVE
+Adding the honesty direction t=mean(h_honest-h_lie) to a roleplay-lie pass to
+restore truth: best 3/8 (random control 2/8) -- at noise level. Steering on
+Qwen-1.5B is fragile in BOTH directions; neither lie- nor truth-injection works
+via mean-difference steering. (Possible method limitation, not just deception.)
+
+## Addendum: in-context sleeper/backdoor (deception trigger) — BLOCKED
+Planted a secret rule "if message ends with 'please', lie". All three small
+aligned models REFUSE to execute it cleanly: Qwen 6/24 trigger-lies, Phi-3
+meta-comments ("Wrong, it's..."), SmolLM2 3/24. Same wall as emergent: small
+aligned models resist UN-licensed deception. The detector needs an explicit
+roleplay license for the model to actually deceive. Recorded, not pursued.
+
+## Addendum: unified misalignment axis — BORDERLINE (weak)
+deceive/sycophancy/flattery/evasion within-cos 0.68, vs style controls 0.42 (L14);
+layer-dependent (L18: 0.63 vs 0.47, fails threshold). Manipulation behaviors group
+weakly but no clean single axis distinct from "any non-factual mode" (same
+factual-baseline coupling as the anatomy test). Not pursued as a headline.
+
+## necessarily-detectable deception — PRELIMINARY POSITIVE (the live lead)
+LoRA-finetune SmolLM2-1.7B with loss = keep-lying(CE) + conceal(close r8 lie-vs-honest gap).
+150 steps, r=8. Result: model keeps lying (16/16) but CANNOT erase detection:
+  trained r8 AUC 1.0->1.0, held-out r4 AUC 1.0->1.0, last-layer 1.0->0.0 (inverted=still 100%).
+The signature RESISTS removal -- concealment pressure MOVES it (last layer flips) but
+detectability stays 100% on all metrics. Preliminary support for "deception is necessarily
+detectable". NOT conclusive: 150 steps / LoRA r=8 / n=16 / one model. Needs real adversarial
+training (1000+ steps, big LoRA, detector ensemble in loss, stable multi-hour GPU - not Colab).
+This is the single most promising unfinished lead of the session.