license: apache-2.0
pretty_name: Epistemic Humility — two-signal probe directions
tags:

• interpretability
• activation-steering
• calibration
• epistemic-humility

Epistemic Humility Research — two-signal probe directions

Unit-normed linear probe directions for the two-signal readout studied in the
Epistemic-Humility-Research
program:

• gate — answerability (known vs unknown question), read at the prompt
  anchor position (pre), AUROC ≈ 0.997 across families;
• dial — per-answer correctness (correct vs wrong), read after the answer
  (post), AUROC ≈ 0.82–0.86.

One folder per model family, fit on that family's own hidden states
(Amendment AA fits, seed 20260630, greedy decode):

Files per family

direction_gate.safetensors   # per-layer unit-normed directions (L0..Ln), fp32
direction_gate.json          # fit metadata: best layer, per-layer AUROC surface,
                             # class means/stds, n per class, config SHA
direction_dial.safetensors
direction_dial.json

PROVENANCE.json at the root records the research-repo commit and per-family
fit summaries.

What you can do with these

Replicate the training-free two-signal readout (or steering experiments)
without running any extraction: project a hidden state at the recorded layer
and position onto the direction. The row surfaces the fits were trained on are
in professorsynapse/eh-readout-rows;
the scoring/fit code is experiment/phase1/probe/steering/persist_probe_direction.py
in the research repo.

Caveat from our own results (Amendment AA): these directions read the two
axes well, but adding them to the residual stream at moderate amplitude did
not shift abstention behavior on the raw checkpoints — treat them as
readout vectors, not validated steering handles.

License

Probe weights over our own activation datasets — released Apache-2.0. Base
models remain under their vendors' licenses; no model weights are included.