DEPRECATE specs 060G and 060H — refuted by PR openai#1898 measurement

PR openai#1898 (X-Abhishek-X) ran Partial SpinQuant + EMBED_BITS=6 reinvest on
the same chain and reported val_bpb 1.06614 vs their base openai#1851's 1.06128
= +0.00486 REGRESSION. Their PR framed it as -0.01486 vs the 2-week-old
merged SOTA openai#1493 (1.0810) instead of vs their actual parent.

Implications:
- 060G (Partial SpinQuant): empirically null/negative on this chain.
- 060H (EMBED_BITS=6 alone or with LQER reinvest): even riskier without
  SpinQuant's rotation protection.

Both specs marked as DEPRECATED at the top. Not deleted (kept as
documentation for if conditions change later, e.g., deploy-time repair
specifically targeting tok_emb precision).

Author: leon2k2k2k

research/specs/060G-partial-spinquant.md

@@ -1,5 +1,20 @@
 # Spec 060G — Partial SpinQuant from PR #1898 on 060A baseline
 
+**Status: DEPRECATED 2026-04-29 — empirically refuted by PR #1898 itself.**
+
+PR #1898 ran this exact lever (Partial SpinQuant + EMBED_BITS=6 reinvest) on
+its own base (#1851 at 1.06128) and got **1.06614** — a **regression of
++0.00486 BPB**. Their framing of "−0.01486 vs merged SOTA #1493 (1.0810)"
+is misleading; the like-for-like comparison vs their actual parent shows
+the lever doesn't help.
+
+Reason to keep this spec on file: documentation of why we're NOT pursuing
+SpinQuant on the 060A line. If the lever ever becomes promising in the
+future (e.g., paired with deploy-time repair or different bit allocations),
+the spec is here as a starting point. Do not run.
+
+---
+
 **Date:** 2026-04-29
 **Branch:** `exp/060G-partial-spinquant` (forked from research)
 **Parent:** 060A + #1898 SpinQuant code (port from `X-Abhishek-X/parameter-golf` PR #1898).

research/specs/060H-embed-bits-6.md

@@ -1,5 +1,21 @@
 # Spec 060H — EMBED_BITS=6 with LQER recovery on 060A baseline (eval-only)
 
+**Status: DEPRECATED 2026-04-29 — implied refutation by PR #1898.**
+
+PR #1898 ran EMBED_BITS=6 *with* SpinQuant rotation as protection against
+INT6 noise and got a **+0.00486 BPB regression** vs their base. EMBED_BITS=6
+*without* SpinQuant (this spec's H1, H2, H3 arms) has *more* INT6 noise on
+`tok_emb`, so all arms here would likely regress further. Don't run on its
+own — the pessimistic scenario in our prediction table is the most likely
+outcome.
+
+Document kept for reference. If we ever build deploy-time repair (060C)
+that specifically targets `tok_emb` precision recovery, this spec becomes
+worth re-examining as a stack candidate — but only after that's measured
+and shown to work.
+
+---
+
 **Date:** 2026-04-29
 **Branch:** `research` (config-only)
 **Parent:** 060A `final_model.pt` + `RESUME_FROM_CKPT` infrastructure (commit `a7c0ed8`).