Update README.md

README.md

@@ -4,7 +4,7 @@
 
 Outlier is an open-source ternary-quantized Mixture-of-Experts runtime and a small family of models trained by one founder on a single-developer pipeline. Routed experts are stored in {-1, 0, +1} ternary at ~1.6 bits per weight. A frozen full-precision Qwen2.5 base acts as the shared expert — our models are **overlays** that attach to an unmodified base, not standalone checkpoints. Top-2 routing per MoE layer. Apache 2.0.
 
-Built solo in 14 days on a Mac Studio M1 Ultra plus spot B200 GPUs. Total compute spend under $1,000. Three U.S. provisional patents filed.
+Built solo in 16 days on a Mac Studio M1 Ultra plus spot B200 GPUs. Total compute spend under $1,000. Three U.S. provisional patents filed.
 
 ## Honest benchmark status
 
@@ -17,34 +17,29 @@ Every number here has a sample size, a stderr, a harness version, and a status.
 | Outlier-150B V3.2 | **84.46%** | 0.29% | 14,042 | 0.4.9.1 | **Verified** — day 13 re-measurement |
 | Outlier-70B V3.3 (alpha-fixed) | **83.10%** | 0.30% | 14,042 | 0.4.9.1 | **Verified** — day 13 full run |
 | Outlier-40B V3.2 | **77.80%** | 0.33% | 14,042 | 0.4.11 | **Verified** — day 12 full run |
-| Outlier-10B V3.2 | 76.19% | — | limited | 0.4.11 | **Unverified** — source file is a smoke test, re-running |
+| Outlier-10B V3.3 | **70.87%** | — | 14,042 | 0.4.9.1 | **Verified** — day 13 V3.3 run |
 
-**Two caveats we're documenting publicly:**
+### Secondary benchmarks (V3.3, verified)
 
-1. **10B is not defended.** Its day 12 source file used `--limit 570` per subtask and does not meet our standard for a headline number. A full-sample re-run is scheduled for the current cluster sprint. The 76.19% is left on the table as a placeholder, clearly flagged.
+| Model | HellaSwag | ARC-C | ARC-E | Winogrande | TruthfulQA |
+| --- | --- | --- | --- | --- | --- |
+| 150B | 77.00% | 68.50% | 90.00% | 85.50% | 69.19% |
+| 70B | 85.95% | 73.46% | 91.62% | 81.29% | 67.12% |
+| 40B | 84.64% | 73.12% | 91.29% | 80.98% | 67.49% |
+| 10B | 78.30% | 62.88% | 85.98% | 73.80% | 62.11% |
 
-2. **Harness version drift.** The 150B re-measurement landed 1.30pp higher on v0.4.9.1 than an earlier v0.4.11 run on the same weights. We don't yet know whether the drift is systematic. We're locking v0.4.9.1 as our harness going forward and documenting both numbers in our ground-truth file so reviewers can reproduce either.
+**Caveat:** Harness version drift between v0.4.9.1 and v0.4.11 produced a 1.30pp delta on the same 150B weights. We've locked v0.4.9.1 as our reference harness and document both numbers in our ground-truth file so reviewers can reproduce either.
 
 For context: Llama 3.1 70B lands around 83.1% MMLU on full sample. Outlier-70B V3.3 alpha-fixed is in that neighborhood, on a model family trained solo on consumer hardware plus spot GPUs for under $1,000 total compute spend.
 
-## What's running right now
-
-As of Day 14 (April 14, 2026), a 2×B200 cluster sprint is executing in parallel:
-
-- **Alpha-fix reruns** on 10B, 40B, and 150B (the technique that gave 70B V3.3 its +1.61pp recovery)
-- **Full secondary benchmarks** on all scales: GSM8K, HellaSwag, ARC-C, ARC-E, Winogrande, TruthfulQA, HumanEval, MMLU-Pro
-- **Speed experiments:** EAGLE3 speculative decoding, SWIFT self-speculation, dead-expert pruning, FP4 Blackwell inference, verified paged-runtime tok/s benchmarks
-- **Long context:** LongRoPE swap targeting 256K on 70B and 150B (replacing the current YaRN 4x config that reaches 128K)
-- **Safety tier:** Llama Guard 3 wrapper integration, red team evaluation, DPO safety fine-tune on 70B Instruct
-
-No verified-before-published numbers appear on this page. New numbers land in the changelog when their source files land on disk.
-
 ## What actually exists today
 
 - **Open-source engine** at [github.com/Outlier-host/outlier](https://github.com/Outlier-host/outlier) — Apache 2.0. Ternary MoE loader, three-tier paged cache, MPS + CPU backends, lm-eval compatible, alpha overlay loader for post-training recovery.
-- **First-party inference verified on Apple Silicon.** 10B in paged mode on a 64 GB Mac Studio M1 Ultra. Non-paged 10B runs at ~13.5 tok/s on the same hardware. Full 70B V3.3 paged-runtime tok/s is being benched in the current sprint.
-- **GPU-resident expert dequantization.** A patched modeling file materializes ternary experts to bf16 at load time, ~56× speedup vs the original CPU→GPU path on a single B200. Port to 70B and 40B in progress.
-- **Alpha-fix technique.** 280 per-expert scalar gates trained in 18 minutes on one B200 recovered +1.61pp MMLU on 70B where a 68M-parameter LoRA fine-tune regressed. Overlay file is 15 KB — 250,000× fewer trainable parameters than the LoRA approach it outperformed. We believe this is a novel post-training recovery primitive for quantized MoE; a prior art search is running.
+- **Real inference on Apple Silicon.** 10B verified in paged mode on a 64 GB Mac Studio M1 Ultra. Real backend wired end-to-end through the web UI — not a stub.
+- **GPU-resident expert dequantization.** A patched modeling file materializes ternary experts to bf16 at load time, ~56× speedup vs the original CPU→GPU path on a single B200.
+- **Alpha-fix technique.** 280 per-expert scalar gates trained in 18 minutes on one B200 recovered +1.61pp MMLU on 70B where a 68M-parameter LoRA fine-tune regressed. Overlay file is 15 KB — 250,000× fewer trainable parameters than the LoRA approach it outperformed.
+- **Web UI** with research mode, code agent, and real model backend. Convergence integration in progress.
+- **5,300+ downloads** across 9 model repos in 16 days.
 - **Three U.S. provisional patents filed:** #64/026,886 (April 3) · #64/030,368 (April 6) · #64/034,028 (April 9). A fourth covering the alpha-fix technique is under novelty review.
 
 ## What we're not claiming
@@ -67,18 +62,9 @@ Our current `10B / 40B / 70B / 150B` labels count routed-expert parameters and u
 
 V3.2 repos under the old names remain available, marked as superseded, pointing to the renamed V3.3 repos.
 
-## Two-variant launch
-
-At public launch, each scale ships as two distinct SKUs:
-
-- **-Base** — research weights, no safety training, clearly labeled "research use only." Apache 2.0. This is what the founder uses.
-- **-Instruct** — same weights plus a Llama Guard 3 wrapper and (on 68B-A32B) a DPO safety LoRA adapter. Apache 2.0, with the safety layer distributed under its own compatible license. This is the default download and the one the Pro desktop app ships with.
-
-Both variants publish identical benchmark numbers. The difference is the safety layer and the stated intended use.
-
 ## Status
 
-**Pre-launch.** V3.3 public release targets late Day 17 or Day 18 (approximately April 17–18, 2026), immediately after the current cluster sprint closes and the Pro desktop app is signed and notarized. Engine is already public.
+**Pre-launch.** V3.3 public release targets Day 18 (April 18, 2026). Convergence integration sprint is assembling the real backend, web UI, research mode, and code agent into a single release branch. Website live at [outlier.host](https://outlier.host).
 
 ## Links
 
@@ -91,6 +77,7 @@ Both variants publish identical benchmark numbers. The difference is the safety
 
 ## Changelog
 
+- **April 16, 2026:** Day 16. 10B V3.3 secondary benchmarks verified (HellaSwag 78.30%, ARC-C 62.88%, ARC-E 85.98%, Winogrande 73.80%, TruthfulQA 62.11%). Secondary benchmark table added to org card. Real backend wired end-to-end through web UI. Website deployed at outlier.host. Convergence sprint assembling release branch for Day 18. 5,300+ total downloads across 9 repos.
 - **April 14, 2026:** Day 14 cluster sprint running — alpha-fix reruns on 10B/40B/150B, speed experiments (EAGLE3, SWIFT, dead-expert pruning, FP4), LongRoPE 256K context, Llama Guard 3 + DPO safety tier. Org card updated with verified Day 13 numbers and V3.3 naming transition.
 - **April 13, 2026:** Day 13 cluster sprint closed with five wins. 70B V3.3 alpha-fixed at 83.10% MMLU (n=14,042, stderr 0.30%). 150B V3.2 re-measured at 84.46% MMLU under lm_eval 0.4.9.1 (supersedes the earlier 83.16% value). YaRN 4x config validated for 128K context on 70B and 150B. V4 HESTIA + LoRA approach killed after regression, fully recovered via 280-scalar alpha-fix on a 15 KB overlay.
 - **April 11, 2026:** Removed an unverified four-row MMLU table that relied on a decommissioned cluster's unsaved source files. The forensic cleanup led directly to the Day 9 provenance rules that now govern every number on this card.