Evidence for report v1.1 and ICE 2.0

Adds appendix/ice-v1.1-ice2-evidence.tar.gz: the raw llama-perplexity KLD output behind every number in report v1.1 (U1-U3) and the ICE 2.0 report (3.1-3.3), with an INDEX and per-file SHA-256; links from both reports; SHA256SUMS updated. Corrects 19G code to -17.5% and states the U1 session precisely (abliterated CyberTiel checkpoint, 163 chunks).

README.md

@@ -31,7 +31,7 @@ checkpoints.
 |---|---|---|
 | ICE v1, report 1.0 | August 2026 | error-propagation classes, five rules, four laws; Ornith-1.5-35B-A3B |
 | ICE v1, report 1.1 | October 2026 | Law 3 depth profile measured per checkpoint (U-shaped); Law 4 fitted on the per-block error (χ² 14.67 → 0.29); measurements unchanged |
-| ICE 2.0, report 1.0 | October 2026 | Sangreal domain-matched imatrix, ICE 1.6 recipes, 12G–25G ladder, format-preserving expert repair; paired against ICE 1.6, code KLD −4.5% to −17.6% at every tier |
+| ICE 2.0, report 1.0 | October 2026 | Sangreal domain-matched imatrix, ICE 1.6 recipes, 12G–25G ladder, format-preserving expert repair; paired against ICE 1.6, code KLD −4.5% to −17.5% at every tier |
 
 ## What ICE is
 
@@ -93,6 +93,7 @@ paper/ICE_Technical_Report.md      the report
 appendix/original-ICE/             PRINCIPLE.md and RECIPE.md, the method as first written
 appendix/recipes/                  every tensor-type-file cited in the paper
 appendix/measurements/             raw llama-perplexity output for every KLD quoted
+appendix/ice-v1.1-ice2-evidence.tar.gz   raw KLD logs for report v1.1 and the ICE 2.0 report (INDEX.md inside)
 appendix/SHA256SUMS.txt            checksum for every file above
 ```
 

appendix/SHA256SUMS.txt

@@ -56,3 +56,4 @@ efc76732cf05eb6fdf750336a0a9ac574c495b87b4bddc0bdb5c321a3b5295c1 *./recipes/cfg_
 e88b620a8c80cab33661be765ed9f01ec9a56d1ee9ae239d429a1f19e0aa5f2d *./recipes/cfg_UD-IQ4_XS_comparison.txt
 d4f78a46b67f7105edf66f8c6a0a8a99dafa4337b6dbe707128908d837618cb5 *./recipes/cfg_UD-Q5_K_S_comparison.txt
 cf561e0bbb74d0fdf3ea49cb85006c0221a180cfbeb180c91fb3ca160b7dc7d1 *./recipes/cfg_UD-Q6_K_comparison.txt
+43cdb2c80ddffaa86897ac3bcec0fb4639ae075d8fd63047498a1492495a549b *./ice-v1.1-ice2-evidence.tar.gz

appendix/ice-v1.1-ice2-evidence.tar.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:43cdb2c80ddffaa86897ac3bcec0fb4639ae075d8fd63047498a1492495a549b
+size 231689

paper/ICE2_Technical_Report.md

@@ -20,7 +20,7 @@ ICE 2.0 also differs from v1 in three further respects: a domain-matched calibra
 ICE 1.6 recipes built on refined versions of Laws 3 and 4, and a five-tier ladder from 12G to 25G.
 
 In a paired, single-session comparison against its own ICE 1.6 parent files on MiniCPM-V-4.7-35B-A3B, ICE 2.0
-lowers code KLD by **4.5–17.6%** and code @16K KLD by **4.1–18.6%** at every tier, with the strongest effect at
+lowers code KLD by **4.5–17.5%** and code @16K KLD by **4.1–18.6%** at every tier, with the strongest effect at
 19G (−14% to −19% on all three columns, paired t below −7). On Ornith-1.5-35B-A3B, the checkpoint ICE v1 was
 developed on, KLD at equal size falls at each step from ICE 1.5 to ICE 1.6 to ICE 2.0.
 
@@ -79,18 +79,20 @@ quantity this report measures.
 
 Mean KL divergence against the model's own bf16 logits (lower is better). "wiki" and "code" are 64 chunks of
 2,048 tokens; "code @16K" is 8 chunks of 16,384 tokens. Sizes in 3.1–3.2 exclude the one-layer MTP draft head
-(0.29 GB), as the rival files carry none; sizes in 3.3 are file sizes as published.
+(0.29 GB), as the rival files carry none; sizes in 3.3 are file sizes as published. The raw `llama-perplexity`
+output behind every number in this section is in [`appendix/ice-v1.1-ice2-evidence.tar.gz`](../appendix/ice-v1.1-ice2-evidence.tar.gz)
+(contents described in its `INDEX.md`).
 
 ### 3.1 ICE 2.0 against its ICE 1.6 parent (one session, paired)
 
 MiniCPM-V-4.7-35B-A3B Uncensored (Qwen3.6-35B-A3B architecture), 14 October 2026. Imatrix: Sangreal 10k,
-computed on a sibling Qwen3.6 fine-tune (Occamy-1.0).
+computed on a sibling Qwen3.6 fine-tune (Occamy-1.0). Raw logs: `ice2-3.1-3.2_minicpm-v47-35b-a3b_2026-10-14/`.
 
 | tier | wiki | code | code @16K |
 |---|---|---|---|
 | 12G | 0.1495 → 0.1574 (+5.3%) | 0.1064 → **0.1014 (−4.7%)** | 0.0847 → **0.0812 (−4.1%)** |
 | 15G | 0.0716 → 0.0717 (+0.2%, n.s.) | 0.0509 → **0.0486 (−4.5%)** | 0.0418 → **0.0381 (−8.8%, t −6.9)** |
-| 19G | 0.0295 → **0.0253 (−14.1%, t −10.5)** | 0.0189 → **0.0156 (−17.6%, t −7.8)** | 0.0151 → **0.0123 (−18.6%, t −7.4)** |
+| 19G | 0.0295 → **0.0253 (−14.1%, t −10.5)** | 0.0189 → **0.0156 (−17.5%, t −7.8)** | 0.0151 → **0.0123 (−18.6%, t −7.4)** |
 | 23G | 0.0120 → 0.0121 (+0.3%, n.s.) | 0.0080 → **0.0076 (−5.0%)** | 0.0065 → **0.0061 (−6.7%)** |
 | 25G | 0.0081 → 0.0082 (+1.0%, n.s.) | 0.0053 → **0.0049 (−6.7%)** | 0.0042 → **0.0040 (−6.7%)** |
 
@@ -121,7 +123,8 @@ checkpoint. UD's ladder has no file at the sizes of 12G and 15G: it ends at 12.2
 ICE v1 was developed on Ornith-1.5-35B-A3B; its abliterated derivative (huihui-ai) was rebuilt with ICE 1.5,
 ICE 1.6 and ICE 2.0.
 
-**ICE 1.5 → ICE 1.6, with Unsloth Dynamic 2.0, one session (25 September 2026):**
+**ICE 1.5 → ICE 1.6, with Unsloth Dynamic 2.0, one session (25 September 2026).** Raw logs:
+`ice2-3.3_ornith15-huihui_ice15-ice16-ud_2026-09-25/`.
 
 | file | size | wiki | code | code @16K |
 |---|---:|---:|---:|---:|
@@ -139,7 +142,8 @@ ICE 1.6 lowers KLD over ICE 1.5 at every shipped tier (paired; at 15G by 4.4–4
 −6.4). At matched size ICE 1.6 23G is below `UD-Q4_K_XL` on every column (−11.8% wiki, −6.4% code), and
 ICE 1.6 19G is 21% below `UD-IQ4_XS` on code. `UD-Q5_K_XL` sits on the floor side of Law 4, 2.16 GB above 25G.
 
-**ICE 1.6 → ICE 2.0** (separate sessions; the bf16 reference PPL agrees within 0.4%):
+**ICE 1.6 → ICE 2.0** (separate sessions; the bf16 reference PPL agrees within 0.4%). ICE 2.0 raw logs:
+`ice2-3.3_ornith15-huihui_ice20_2026-10-28/`.
 
 | tier | code | code @16K | wiki |
 |---|---|---|---|

paper/ICE_Technical_Report.md

@@ -80,16 +80,18 @@ had looked like the largest remaining direction.
 
 Every number in this report stands as measured. Later measurements on the same checkpoint refined two of the
 four laws; the recipes from ICE 1.6 on are built on the refined forms. The method that followed this report, ICE 2.0, is
-described separately in [`ICE2_Technical_Report.md`](ICE2_Technical_Report.md).
+described separately in [`ICE2_Technical_Report.md`](ICE2_Technical_Report.md). Raw logs for every number in these
+updates: [`appendix/ice-v1.1-ice2-evidence.tar.gz`](../appendix/ice-v1.1-ice2-evidence.tar.gz) (folders `v1.1-U1_…`,
+`v1.1-U2_…`, and `ice2-3.3_…_2026-09-25` for U3).
 
 ### U1. Law 3 (placement): the depth profile is U-shaped
 
 Section 4.3 fits the per-block value of a bit as `g(t) ≈ exp(−t/9.95)` and Rule 5 (Section 3.6) places the
-high type shallow-first. A direct measurement on Ornith-1.5-35B-A3B (13 September 2026) shows the per-block
-value is **U-shaped**: it reaches its minimum around blocks 25–30 and rises again toward the deep tail.
+high type shallow-first. A direct measurement on the abliterated, CyberTiel-calibrated Ornith-1.5-35B-A3B checkpoint
+(13 September 2026) shows the per-block value is **U-shaped**: it reaches its minimum around blocks 25–30 and rises again toward the deep tail.
 
-Six blocks demoted at the 25G budget, four byte-identical files except for *which* blocks hold the low type,
-one session, 64 chunks:
+Six blocks demoted at the 25G budget, four files identical in size (within 32 bytes) and differing only in *which*
+blocks hold the low type, one session, 163 chunks of WikiText-2:
 
 | blocks demoted | mean KLD |
 |---|---:|
@@ -98,12 +100,14 @@ one session, 64 chunks:
 | 34–39 (shallow-first, as Rule 5 prescribes) | 0.030007 |
 | 20–25 | 0.030237 |
 
-Demoting the trough is 2.5% better at identical size, active bpw and file bpw, and also
+In this session the trough placement is 1.6% better than shallow-first (paired over the 145 common chunks:
+t = −3.25, p = 0.00057, better on 93 of 145). The shipped 25G tier with the trough demoted measured 2.5% better
+than its shallow-first predecessor on the 64-chunk harness, at identical size, active bpw and file bpw, and also
 better on the 99%, 99.9% and same-top-1 columns. The comparison needs the **paired** standard error (per-chunk
 differences): the unpaired `√(σ₀²+σ₁²)` overstates the uncertainty about 3× and hides a result at p ≈ 0.0006.
 
 **Update.** `g(t) ≈ exp(−t/9.95)` and "shallow-first" are replaced by a measured profile. The depth profile is
-a property of the checkpoint: measure it per model (a handful of byte-identical files at the most generous
+a property of the checkpoint: measure it per model (a handful of size-identical files at the most generous
 tier's budget, one session) and demote the trough. A wrong depth prior only costs bytes when a tier's high/low
 boundary falls shallower than the trough, so the generous, high-bpw tiers are the ones to check first.