# project: modelmap document: qwen3-0.6b-4bit/probes/v1 — confidence author: Simon-Pierre Boucher contact: contact@spboucher.ai website: https://modelmap.io created: 2026-08-12 status: reviewed

# Confidence — qwen3-0.6b-4bit / probes / v1

text

Level      : 1
Seeds      : 5
Prompt sets: 6 (2 disjoint template families per property)
Methods in agreement : 1 (linear probes only — Level 2 requires a second method)
Causal verification  : none (observational; Level 3 requires intervention)

Per-property evidence:

lang_id: maxAcc A/B = 1.000/1.000, seed SD 0.0000, dataset shift 0.0000, twin max selectivity 0.808, replication(top-5) 1.00/1.00
code_prose: maxAcc A/B = 1.000/1.000, seed SD 0.0000, dataset shift 0.0000, twin max selectivity 0.883, replication(top-5) 1.00/1.00
arith: maxAcc A/B = 1.000/1.000, seed SD 0.0000, dataset shift 0.0000, twin max selectivity 0.558, replication(top-5) 1.00/1.00

This is a published NEGATIVE result (Level 1 for the negative claim). The random-init architecture twin matches the trained model at ceiling (accuracy 1.00, 28/28 layers FDR-significant, for the twin as for the real model; mean real-minus-twin selectivity within +/-0.06). By the validity criterion registered in hypothesis.md BEFORE the run (twin selectivity must stay < 0.05), this probing harness is INVALID for localization claims on these promptsets: it measures the tokenizer + architecture prior, not learned computation. The negative claim itself is controlled and replicated (5 seeds, 2 disjoint promptsets, 3 properties) - hence Level 1.

Consequences adopted: (1) probe maps are only publishable as REAL-MINUS-TWIN differentials; (2) promptsets v2 must remove lexical separability (shared vocabulary across classes); (3) the seed-vs-dataset variance hypothesis is untestable at ceiling and moves to run #2.