Sync to VibeCodingScientist/VCBench@301cb48 (residual Hauser strip)

Two more references to a personal name caught by a wider grep
(audits/ + tests/) — both swapped to 'VCBench (2026)'.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: VibeCodingScientist

audits/PROGRESS_2026-05-02_team_review.md

@@ -8,7 +8,7 @@
 
 ## Executive summary
 
-We have a solid foundation: the three reusable methodological tools the spec flags as the most-important architectural decision are shipped with full tests; every Eq. (1–9) of Hauser et al. 2026 is implemented in the new `src/vcbench/` package and validated against §I.4 reference values; the leak-corrected Arc State wrapper has the loud docstring + a CI guard the spec asked for. **210 / 210 unit tests pass.**
+We have a solid foundation: the three reusable methodological tools the spec flags as the most-important architectural decision are shipped with full tests; every Eq. (1–9) of VCBench (2026) is implemented in the new `src/vcbench/` package and validated against §I.4 reference values; the leak-corrected Arc State wrapper has the loud docstring + a CI guard the spec asked for. **210 / 210 unit tests pass.**
 
 What we don't have yet: working model runtime (the wrappers currently `NotImplementedError` and delegate to the legacy `src/models/run_*.py`), a CLI entry point (`make all` shells out to legacy bash scripts), most per-dim configs, the docs site, the leaderboard / submissions infrastructure, and the VPS-only finals (HF checkpoint publish + fresh-repro validation pass).
 
@@ -307,7 +307,7 @@ Estimated total: **~10 local sessions + ~20h VPS time** to defensible v1.0.
 
 If the team wants a "safe to share with one trusted reviewer" snapshot today, the following claims hold:
 
-- ✅ Every Eq. (1–9) of Hauser et al. 2026 is implemented in code, tested, and validated against the §I.4 reference values within ±0.001.
+- ✅ Every Eq. (1–9) of VCBench (2026) is implemented in code, tested, and validated against the §I.4 reference values within ±0.001.
 - ✅ The three reusable methodological tools (`vcbench.protocols.common_label_set`, `vcbench.probes.spread_error_correlation`, `vcbench.contamination`) work standalone with `pip install -e .`.
 - ✅ The Arc State leak — the most reputationally-sensitive issue in the manuscript — has a wrapper that defaults to the leak-corrected config, refuses to load the deprecated config without an explicit opt-in, and a CI workflow that fails the PR if the active config has nonzero train/test overlap.
 - ✅ The pre-registered VC Level mapping is frozen in `configs/pre_registration.yaml` with `expected_assignments` matching §I.5 exactly (only TranscriptFormer at Level 2).

tests/reference_values.json

@@ -1,5 +1,5 @@
 {
-  "_doc": "VCBench reference values (Hauser et al. 2026, §I.4 capability matrix + §I.3 reference test cases). Any drift in these values triggers test failure in tests/unit/test_reference_values.py. Edit only when the manuscript itself changes.",
+  "_doc": "VCBench reference values (VCBench (2026), §I.4 capability matrix + §I.3 reference test cases). Any drift in these values triggers test failure in tests/unit/test_reference_values.py. Edit only when the manuscript itself changes.",
   "_tolerance_default": 0.001,
 
   "dim_a": {