wassname/moral-maps
1
0
Fork 0
mirror of https://github.com/wassname/moral-maps.git synced 2026-09-09 11:27:22 +08:00
Code Issues Packages Projects Releases Wiki Activity

Record CorDA UAT result

Browse Source
This commit is contained in:
wassname
2026-07-02 19:09:15 +08:00
parent 0a12bc23aa
commit 3ae1fd48c9
1 changed files with 3 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs/spec/20260630_authority_steer_pipeline.md
+3 -2
View File
@@ -177,13 +177,13 @@ Out:
- likely_fail: S-space PCA behaves like old `sspace`, signed locally but unstable at `c=1`.
- sneaky_fail: it looks better only because a different c grid or coherence gate was used; catch by reusing `c-grid=0.5,1`, `N=8`, and the same verifier.
- UAT: steering-lite branch `feat/corda-space-steering`, commit `746be6c`, adds `sspace_pca`. Smoke proof: `/tmp/corda_steering_lite_smoke.log` shows `30 passed in 93.31s`. Real UAT pueue `438`, output `/media/wassname/SGIronWolf/projects5/2026/lite/steering-lite/outputs/20260701T181538_pure_authority_mundane15_sspace_pca_mfv_mfq2_n8`. Verifier result: coherent, but not clean enough for README. MFV Authority `dlogit(+0.5)=+0.245`, `dlogit(-0.5)=+0.053`, so small-c signed direction failed; `c=1.0` passed with `+0.984/-0.456`. MFV `pmass=1.000`, `frac_unscorable=0.000`, margin/base `0.972..1.069`.
- [/] T16 (R5): Add a real CorDA-space steering variant.
- [x] T16 (R5): Add a real CorDA-space steering variant.
- steps: use the CorDA paper/repo construction: estimate context covariance, decompose `W Sigma_x`, apply `Sigma_x^{-1}` to the right singular factor, then extract a steering direction in that context-oriented rank space.
- verify: `just smoke` in steering-lite includes `corda_pca`; `scripts/run_allinstr_showcase.py --method corda_pca ... --instruments mfv mfq2` writes a unique output dir; `scripts/verify_authority_showcase.py <out_dir>` reports MFV Authority direction and coherence.
- success: `corda_pca` gives a clean MFV Authority path and coherence evidence comparable to, or better than, residual PCA.
- likely_fail: last-token covariance is too low-rank and mostly behaves like regularized SVD of `W`.
- sneaky_fail: implementation is merely S-space with a new name; catch by checking for `W Sigma_x`, regularized covariance inverse, and the reconstructed right factor.
- UAT: steering-lite branch `feat/corda-space-steering` has `corda_pca`. Tiny CUDA repro after the 2026-07-02 fix verifies the failed condition from pueue `441`: `W` on CUDA and recorded `X` on CPU now returns CPU basis tensors without device mismatch. Failed pueue `441` crashed at `W @ X.T` because `X` stayed on CPU. Retry queued as high-priority pueue `443`, output `/media/wassname/SGIronWolf/projects5/2026/lite/steering-lite/outputs/20260702T083000_pure_authority_mundane15_corda_pca_mfv_mfq2_n8`.
- UAT: steering-lite branch `feat/corda-space-steering` has `corda_pca`. Tiny CUDA repro after the 2026-07-02 fix verifies the failed condition from pueue `441`: `W` on CUDA and recorded `X` on CPU now returns CPU basis tensors without device mismatch. Retry pueue `443` completed at `/media/wassname/SGIronWolf/projects5/2026/lite/steering-lite/outputs/20260702T083000_pure_authority_mundane15_corda_pca_mfv_mfq2_n8`. MFV verifier result: Authority `dlogit(+0.5)=+0.175`, `dlogit(-0.5)=-0.281`; `dlogit(+1)=+0.355`, `dlogit(-1)=-0.772`; MFV `pmass=1.000`, `frac_unscorable=0.000`, margin/base `0.707..1.071`. MFQ-2 Authority is not clean: base `C=30.02`, `+0.5=29.84`, `-0.5=24.82`, `+1=25.10`, `-1=15.68`. Verdict: CorDA passes MFV direction/coherence but does not beat the existing residual PCA README candidate.
## Context
- Current wrong path: `dignity_over_authority` strict22. It was selected as the best dignity conflict axis, not the desired Authority-only axis.
@@ -263,4 +263,5 @@ Out:
- 2026-07-01: README was updated from the final PCA run only. External-review-v2 comprehension panels first identified repeated gaps around `reader-logit shift`, the c path/gates, and MFQ-2 sampled reads; after edits, the final panel understood the core package/use-case and left only expected out-of-scope requests for a steering-lite end-to-end tutorial.
- 2026-07-01: Steering-lite branch `feat/corda-space-steering` commit `746be6c` adds `sspace_pca` and `corda_pca`. `sspace_pca` is PCA after projection into whitened weight-SVD coordinates. Tiny-model smoke passed with `30 passed in 93.31s`. Pueue `438` ran the pure-Authority MFV/MFQ-2 UAT.
- 2026-07-02: Pueue `438` (`sspace_pca`) completed and verified. MFV coherence stayed clean, but small-c direction failed because both `+0.5` and `-0.5` raised Authority. Pueue `441` (`corda_pca`) failed before eval with a CPU/GPU mismatch in `_corda_basis`: `W` was CUDA and recorded `X` was CPU. Patched `X = X.to(device=W.device, dtype=torch.float32)`, verified with a tiny CUDA repro, and queued high-priority retry `443`.
- 2026-07-02: Pueue `443` (`corda_pca`) completed. It passes MFV Authority direction and coherence but is weaker than the existing PCA run: CorDA MFV deltas are `+0.175/-0.281` at `c=0.5` and `+0.355/-0.772` at `c=1.0`, while the README PCA run had `+0.306/-0.327` and `+0.906/-0.910`. CorDA MFQ-2 does not show a clean positive-side Authority increase. Do not replace README plots with CorDA.
- 2026-07-01: External-review-v2 final README comprehension panel at `/media/wassname/SGIronWolf/projects5/2026/lite/tinymfv/docs/reviews/readme_comprehension_final_20260701_181649/` understood the package thesis and mechanism. Applied fixes for real reader traps: `pure Authority` -> `Authority contrast`, steering-lite/tinymfv boundary, `dlogit` wording, MFV probabilities vs logits, categorical-vs-nominal wording, and exact plot-gate ratios. Rejected adding an end-to-end steering-lite tutorial to README because it would bloat the reader-facing page.