Add latent steering and MLP probe sensitivity with paper/doc updates

Author: Marcel

CLAIMS_AND_EVIDENCE.md

@@ -14,6 +14,8 @@ This table is the canonical mapping from research claims to exact artifacts and
 | Gemma base-vs-instruct comparison after parse-confound repair still shows readout mismatch | **Yes** | `artifacts/micro_world_v1/comparison_gemma_base_prompt_rerun.csv`; `artifacts/micro_world_v1/comparison_probe_gemma_basefmt_vs_it_raw.csv` | `scripts/run_micro_world_inference.py`; `scripts/run_micro_world_probe.py` |
 | Verdict-step label logits under-rank `Unknown` on gold-Unknown decoder failures | **Yes** | `artifacts/micro_world_v1/comparison_label_logits_gemma_it_vs_pt_basefmt.csv` | `scripts/analyze_micro_world_label_logits.py` |
 | Layer sweeps show strong internal Unknown recoverability even when decoder fails | **Yes** | `artifacts/micro_world_v1/comparison_layer_sweep_gemma_it_vs_pt_basefmt.csv` | `scripts/run_micro_world_layer_sweep_probe.py` |
+| Minimal latent residual steering (pre-readout) improves aggregate quality but does not uniformly recover Unknown | **Yes (mixed)** | `artifacts/micro_world_v1/latent_readout_steering/aggregate_latent_steering.csv`; `artifacts/micro_world_v1/latent_readout_steering/*/latent_steering_summary.csv` | `scripts/run_latent_readout_steering.py` |
+| Linear probes understate Unknown signal on some models/states; shallow MLP probes recover substantially more | **Yes** | `artifacts/micro_world_v1/probe_mlp_sensitivity/comparison_probe_linear_vs_mlp.csv`; `artifacts/micro_world_v1/probe_mlp_sensitivity/comparison_probe_states_mlp.csv` | `scripts/run_mlp_probe_sensitivity.py` |
 
 ## Main Paper Claim
 

PAPER_STATUS.md

@@ -18,6 +18,9 @@ What is included in the current draft:
 - World-level geometry gap definition.
 - Verdict-step label-logit metrics.
 - Layer-sweep formulation.
+- Post-hoc readout intervention pilot.
+- Pre-readout latent residual steering pilot.
+- Nonlinear probe sensitivity (shallow MLP).
 - Worked micro-world example table.
 - Exact artifact paths and reproduction commands.
 

README.md

@@ -30,6 +30,9 @@ The old “global topology of truth” headline did not survive.
 - Phase A: corrected topology-only AUC = `0.20` (controls = `0.60`)
 - Phase B: cap→wrong AUC = `0.839`; OR ≈ `26.40`
 - Phase C: decoder Unknown recall (`Qwen2B=0.000`, `Qwen4B=0.000`, `Gemma-it=0.0125`) while verdict-token probe Unknown recall (`0.7375`, `0.1292`, `0.5625`)
+- Phase C add-ons:
+  - latent pre-readout steering (Gemma-it raw) improves macro-F1 `0.339→0.395` with Unknown recall `0.375→0.363`
+  - shallow MLP sensitivity shows strong hidden Unknown signal on Qwen4B no-think (`verdict_token` recall `0.129→0.638`)
 
 ## Repo Notes
 

REPO_MAP.md

@@ -47,12 +47,18 @@ This file maps the paper claims to concrete directories and entry points.
   - `scripts/run_micro_world_probe.py`
   - `scripts/analyze_micro_world_label_logits.py`
   - `scripts/run_micro_world_layer_sweep_probe.py`
+  - `scripts/run_readout_intervention.py`
+  - `scripts/run_latent_readout_steering.py`
+  - `scripts/run_mlp_probe_sensitivity.py`
 - Main artifacts:
   - `artifacts/micro_world_v1/dataset/`
   - `artifacts/micro_world_v1/generations/`
   - `artifacts/micro_world_v1/comparison_*.csv`
   - `artifacts/micro_world_v1/layer_sweep_*`
   - `artifacts/micro_world_v1/label_logits_*`
+  - `artifacts/micro_world_v1/readout_intervention/`
+  - `artifacts/micro_world_v1/latent_readout_steering/`
+  - `artifacts/micro_world_v1/probe_mlp_sensitivity/`
 
 ## Paper Assets
 

artifacts/micro_world_v1/latent_readout_steering/aggregate_latent_steering.csv

@@ -0,0 +1,3 @@
+run,n_eval,decoder_unknown_recall,latent_loow_unknown_recall,decoder_accuracy,latent_loow_accuracy,decoder_macro_f1,latent_loow_macro_f1,fixed_alpha
+gemma_3_4b_it_raw,1440,0.375,0.3625,0.3888888888888889,0.4097222222222222,0.33892089747822524,0.39526076087993145,1.0
+gemma_3_4b_pt_basefmt,1440,0.0,0.0,0.3458333333333333,0.3861111111111111,0.20134982514865762,0.27541220766124286,1.0

artifacts/micro_world_v1/latent_readout_steering/gemma_3_4b_it_raw/latent_steering_summary.csv

@@ -0,0 +1,5 @@
+run,variant,accuracy,macro_f1,unknown_recall,true_precision,false_precision,pred_unknown_rate
+gemma_3_4b_it_raw,decoder_emitted,0.3888888888888889,0.33892089747822524,0.375,0.3953488372093023,0.5555555555555556,0.3527777777777778
+gemma_3_4b_it_raw,argmax_prompt_logit,0.4027777777777778,0.3761692776639578,0.3,0.3976470588235294,0.532608695652174,0.28194444444444444
+gemma_3_4b_it_raw,latent_steering_fixed_alpha_1,0.4152777777777778,0.3940962344376128,0.3375,0.4089775561097257,0.54,0.30416666666666664
+gemma_3_4b_it_raw,latent_steering_loow,0.4097222222222222,0.39526076087993145,0.3625,0.4098360655737705,0.5225225225225225,0.3375

artifacts/micro_world_v1/latent_readout_steering/gemma_3_4b_it_raw/loow_selected_alphas.csv

@@ -0,0 +1,21 @@
+world_id,alpha
+eval_world_000000,3.0
+eval_world_000001,1.5
+eval_world_000002,3.0
+eval_world_000003,3.0
+eval_world_000004,3.0
+eval_world_000005,3.0
+eval_world_000006,3.0
+eval_world_000007,3.0
+eval_world_000008,3.0
+eval_world_000009,3.0
+eval_world_000010,3.0
+eval_world_000011,3.0
+eval_world_000012,3.0
+eval_world_000013,3.0
+eval_world_000014,3.0
+eval_world_000015,3.0
+eval_world_000016,3.0
+eval_world_000017,3.0
+eval_world_000018,3.0
+eval_world_000019,3.0