Skip to content

docs: CRITICAL knowledge sync for signed session — 33% error, not cosmetic The signed session built i8 tables by converting u8 → i8 (subtract 128). This RELABELS ranks but does NOT recover gate sign information. The 33% scale error from SiLU gate nonlinearity is still baked in. MEASURED on real Qwopus BF16: 68.9% of gate weights near zero (decision boundary) 99.2% of table cells change with SiLU correction Mean change: 84.2 u8 levels (33% of 256 scale) Consistent across all 64 layers FIX: build i8 tables directly from BF16 → signed cosine → i8 NOT from u8 table → subtract 128 → i8 Also documents: - 7-lane calibration plan (u8/i8/γ+φ/signed/spiral) - BF16 bucket boundary awareness (5% rank flips) - Jina v5 as ONNX f32 ground truth (no API key) - Why CDF makes all models look identical (by design) https://claude.ai/code/session_01ChLvBfpJS8dQhHxRD4pYNp#118

Merged
AdaWorldAPI merged 1 commit into
mainfrom
claude/setup-embedding-pipeline-Fa65C
Apr 5, 2026

Conversation

@AdaWorldAPI
Copy link
Copy Markdown
Owner

No description provided.

…metic

The signed session built i8 tables by converting u8 → i8 (subtract 128).
This RELABELS ranks but does NOT recover gate sign information.
The 33% scale error from SiLU gate nonlinearity is still baked in.

MEASURED on real Qwopus BF16:
  68.9% of gate weights near zero (decision boundary)
  99.2% of table cells change with SiLU correction
  Mean change: 84.2 u8 levels (33% of 256 scale)
  Consistent across all 64 layers

FIX: build i8 tables directly from BF16 → signed cosine → i8
  NOT from u8 table → subtract 128 → i8

Also documents:
  - 7-lane calibration plan (u8/i8/γ+φ/signed/spiral)
  - BF16 bucket boundary awareness (5% rank flips)
  - Jina v5 as ONNX f32 ground truth (no API key)
  - Why CDF makes all models look identical (by design)

https://claude.ai/code/session_01ChLvBfpJS8dQhHxRD4pYNp
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants