Wet-Ink

Interactive decoder the real schedule of the v5 kernel — tune W and watch revision opportunities appear and vanish

committed & dry — frozen forever wet window W — T2T gate may rewrite still masked frontier

W 16

speed 1×

step 0

NFE 0

opened 0

T2T rewrites 0

How it works every mechanism is reused verbatim from the base model — the schedule is the only new variable

➡️

One-token frontier

Each diffusion step opens exactly the leftmost remaining mask and commits it AR-style (greedy at T=0). Progress is guaranteed: one mask fewer per step — the AR end of the BD3-LM / Set Diffusion spectrum.

💧

Wet window W

Positions [f−W, f) stay revisable through LLaDA2.2's trained gate p0 > θ & old ≠ x0 — with its SPLIT/DELETE edit ops and anti-loop resampler, all verbatim (L-EBPO).

🔒

Hard dry edge

Outside the window editing_confidence = −∞: a token that leaves the window is never touched again. Ink dries with distance — exactly the owner's §2.1 idea.

Results GSM8K, first 45 tasks · greedy · gen_length 160 · 2×T4 · 4-bit NF4 · paired vs the completed v4b run

Arm	n	Accuracy	Wilson 95% CI	Mean NFE	Rewrites/task
wetink_w16 (the method, W=16)	45	24.4%	[14.2, 38.7]	206.6	6.9
noedit_w16 (frontier only)	22	13.6%	[4.7, 33.3]	201.1	0
wetink_w8 (narrow window)	10	20.0%	[5.7, 51.0]	202.7	5.6
stock_b32 (v4b baseline)	80	22.5%	[14.7, 32.8]	42.5	—
noedit_b32 (v4b, editing off)	80	23.8%	[15.8, 34.1]	50.1	—

p = 1.0

paired exact McNemar
wetink_w16 vs stock (11/45 = 11/45)

×4.88

NFE cost (42 → 207 paired)
structural, no churn explosion

0 / 22

outcomes flipped by window edits
on common tasks

6.9

real rewrites per task
(churn 5.6) — the gate does fire

Honest finding: the raw +10.8 pp gap vs the frontier-only control is a task-composition artifact — the control stopped on the hard tasks 0–21 (every arm: 13.6% there; easy tasks 22–44: 34.8%). Window rewrites flipped no paired outcome. Third consecutive round where inference-time revision is quality-neutral on this family — published as a negative result.

Journey five versions, one idea — full data in the GitHub repo

Version	Base	Approach	Outcome
v1	Nemotron‑Diffusion‑3B	confidence-gate re-masking each stock step	quality ↑, 13× NFE churn (broke termination)
v2	Nemotron‑Diffusion‑3B	full published cycles (RemeDi / CoDR)	termination guaranteed; RemeDi harmful
v3	Nemotron‑Diffusion‑3B	full benchmark, n=200+164	RemeDi −19.5..−31.5 pp (p<0.001); CoDR neutral
v4b	LLaDA2.2‑mini	stock generate() study, 12 arms, n=403	stock gate neutral; window doesn't slide
v5	LLaDA2.2‑mini	own wet-ink frontier decoder	paired-identical to stock (p=1.0) at ×4.88 NFE

Direction for v6: a k>1-token frontier (NFE ÷ k), rewrite-distance telemetry (a direct measurement of the ink-drying profile), n ≥ 500 per arm.

🌊Wet‑Ink

Interactive decoder the real schedule of the v5 kernel — tune W and watch revision opportunities appear and vanish

How it works every mechanism is reused verbatim from the base model — the schedule is the only new variable

One-token frontier

Wet window W

Hard dry edge

Results GSM8K, first 45 tasks · greedy · gen_length 160 · 2×T4 · 4-bit NF4 · paired vs the completed v4b run

Journey five versions, one idea — full data in the GitHub repo