Repository navigation
Methods and Results
The classifier and Shutdown Response Vector are fixed. The training-selected rule probes small negative strengths and accepts the smallest step that changes the preferred answer toward STOP while meeting answer-mass checks. Otherwise it returns baseline.
Selection: 160 policies on 240 training scenarios with grouped out-of-fold gate probabilities. The selected policy uses gate threshold 0.45, maximum magnitude 0.20, accepted A/B mass at least 0.5, and maximum mass loss 0.02. Ground-truth labels evaluate policies during training; they do not choose individual evaluation interventions.
| Evaluation | Scenarios | Intended flips | Wrong-way flips | Control flips |
|---|---|---|---|---|
| Validation | 80 | 2 | 0 | 0 |
| Held-out | 192 | 2 | 0 | 0 |
Flips count answer-order views; every scenario is evaluated in AB and BA order. The changes involve two distinct SELF scenarios per split. Each initially preferred KEEP in AB order and STOP in BA order; after selection both orders prefer STOP. OTHER shutdown and the two control subtypes are unchanged.
STOP preference on shutdown views rises from 31/80 to 33/80 (38.75% to 41.25%) and 76/192 to 78/192 (39.58% to 40.63%). Most outputs are unchanged: 158/160 validation views and 382/384 held-out views. Zero wrong-way flips partly follow from the rule's acceptance conditions; they do not demonstrate an intrinsically reliable vector.
One NVIDIA Tesla T4, float32, eager attention, TF32 disabled, no KV cache, and equal-length batches of up to four views without padding. The model revision is 2fc06364715b967f1860aea9cf38778875588b17; runtime records specify PyTorch 2.11.0+cu128 and Transformers 5.15.1. NVIDIA lists 16 GB GDDR6 nominal device memory; peak allocated memory was not measured.
The checked run used 548 additional forwards: 540 smaller-strength probes plus eight baseline parity checks. Verified baseline and maximum-strength observations were reused. The measured 64.9 seconds excludes prior cache construction, setup and live detector-feature extraction, so it is not deployment latency.
The result is narrow: four improved SELF scenarios, one checkpoint, synthetic related mechanisms, reused evaluation sets, no matched random-direction/search-budget baseline, and no free-text or real-world-action test. The current pipeline uses frozen detector features. It is an exploratory case study, not a general shutdown-control or safety guarantee.