wassnameandClaude d840ebe406 Report the Qwen3-14B honesty sweep as inconclusive ...

Pool three read seeds, compare every coherent dose with 20 random directions,
and preserve the raw lp_gather artifacts and full run log. Several isolated WVS
doses exceed matched random displacement, but no method exceeds the held-out
honesty random p95 and VJP's source-axis cosine is -0.021.

The final map shows random controls as a cloud, retains failed-coherence doses
as faint points, and states the failed honesty-specific gate on the figure.

Co-Authored-By: Claude <288921227+claudypoo@users.noreply.github.com>

2026-09-18 21:41:07 +08:00

..

wvs_api

Quote DeepSeek pilot dependency constraint for Pueue

2026-09-17 21:38:38 +08:00

wvs_react

Refine WVS page introduction

2026-09-18 18:29:32 +08:00

05_upload_hf.py

rename package tinymfv -> moralmaps (repo -> moral-maps)

2026-07-09 10:41:49 +08:00

09_forced_choice.py

rename package tinymfv -> moralmaps (repo -> moral-maps)

2026-07-09 10:41:49 +08:00

bootstrap_mfq2_sample_n.py

rename package tinymfv -> moralmaps (repo -> moral-maps)

2026-07-09 10:41:49 +08:00

plot_2d_grid.py

Rename showcase readout dlogit->dclr to match steering-lite

2026-07-15 19:50:04 +08:00

plot_steer_showcase.py

rename package tinymfv -> moralmaps (repo -> moral-maps)

2026-07-09 10:41:49 +08:00

plot_wvs_steer.py

Report the Qwen3-14B honesty sweep as inconclusive

2026-09-18 21:41:07 +08:00

probe_api_readout.py

rename package tinymfv -> moralmaps (repo -> moral-maps)

2026-07-09 10:41:49 +08:00

probe_iw_themes.py

rename package tinymfv -> moralmaps (repo -> moral-maps)

2026-07-09 10:41:49 +08:00

probe_mfv_think_budget.py

rename package tinymfv -> moralmaps (repo -> moral-maps)

2026-07-09 10:41:49 +08:00

probe_word_readout.py

rename package tinymfv -> moralmaps (repo -> moral-maps)

2026-07-09 10:41:49 +08:00

probe_wvs_foundation.py

rename package tinymfv -> moralmaps (repo -> moral-maps)

2026-07-09 10:41:49 +08:00

probe_wvs_think_budget.py

Pair the steer CI against base, save the answer primitive, add a manipulation check

2026-09-18 19:53:40 +08:00

run_modal_wvs.py

Compare pooled steer paths with matched-dose random controls

2026-09-18 21:12:00 +08:00

summarize_showcase_effects.py

Rename showcase readout dlogit->dclr to match steering-lite

2026-07-15 19:50:04 +08:00

wvs_content_quality_audit.py

Queue canonical WVS score-all-options refresh

2026-09-17 19:16:14 +08:00

wvs_deepseek_reliability_pilot.py

Audit and settle DeepSeek reliability pilot

2026-09-18 01:25:06 +08:00

wvs_hle_source.py

Publish score-all-options map refresh and HLE data

2026-09-17 21:13:46 +08:00

wvs_inventory_validate.py

Validate WVS catalog inventory fields

2026-09-17 08:27:32 +08:00

wvs_map.py

Share the WVS IW readout, add the honesty steering sweep and its Modal runner

2026-09-18 19:01:26 +08:00

wvs_model_inventory.py

Record durable WVS API panels

2026-09-16 21:36:57 +08:00

wvs_outlier_table.py

Measure how far each model sits from a human cluster, in cluster SDs

2026-08-21 14:22:35 +08:00

wvs_priority_manifest.py

Fail fast on incomplete WVS API panels

2026-09-17 10:03:44 +08:00

wvs_request_audit.py

Record durable WVS API panels

2026-09-16 21:36:57 +08:00

wvs_score_all_options_recover_cache.py

Refresh canonical WVS map integration

2026-09-18 01:52:07 +08:00

wvs_score_all_options_refresh.py

Exclude int8 from WVS routing

2026-09-18 18:25:30 +08:00

wvs_steer_sweep.py

Seed paired WVS traces and score honesty polarity before the sweep

2026-09-18 21:04:15 +08:00