teamchong · teamchong · Aug 21, 2026 · Aug 19, 2026 · Aug 20, 2026 · Aug 20, 2026
diff --git a/README.md b/README.md
@@ -72,7 +72,7 @@ provider over some other base URL need a rule for it, and a rule that names a
 port matches only that port:
 
 ```bash
-pxpipe warp --route '127.0.0.1:8082/v1/*=http://127.0.0.1:47821' -- codex
+pxpipe warp --route '127.0.0.1:9090/v1/*=http://127.0.0.1:47821' -- codex
 ```
 
 ## Offline export (no proxy)
@@ -175,6 +175,7 @@ is confabulations, so lower is better.
 | `grok-4.5` | **100/100** | **97/98** | 17/18 | **0/16** | 0/15 | native 14px/84 quality suite (live profile); [quality](eval/grok-density/QUALITY_RESULTS.md), [native-sweep](eval/grok-density/native-sweep/RESULTS.md) |
 | `grok-4.6` high | **100/100** | **97/98** | 17/18 | **0/16** | 0/15 | native 14px/84, reasoning high; [quality](eval/grok-profile/QUALITY_RESULTS.md) |
 | `moonshotai/kimi-k3` | 79/100 | 84/98 | 15/18 | 1/16 | 0/15 | generic GPT profile: [quality results](eval/sol-profile/KIMI_K3_QUALITY_RESULTS.md) |
+| `qwen-3.8` (`@cf/qwen/qwen3.8-27b`) | 98/100 | 72/98 | 11/18 | **0/16** | 0/15 | prior 5×8 broad suite (0/15 hex); native 14px pilot: 8/8 exact, 0 inventions, 11/15 hex: [pilot & quality](eval/qwen-profile/QUALITY_RESULTS.md) |
 
 ### Native-profile cost check
 

diff --git a/eval/gemini-profile/gemini-client.mjs b/eval/gemini-profile/gemini-client.mjs
@@ -1,7 +1,8 @@
 // Google AI Studio eval client through the local Cloudflare AI Gateway.
 
 function gatewayOrigin() {
-  const base = (process.env.OPENAI_BASE_URL || 'http://127.0.0.1:8082/v1').replace(/\/$/, '');
+  const base = (process.env.OPENAI_BASE_URL ?? '').replace(/\/$/, '');
+  if (!base) throw new Error('OPENAI_BASE_URL is required (point it at your gateway)');
   return new URL(base).origin;
 }
 

diff --git a/eval/gist-recall/OPUS_14PX.md b/eval/gist-recall/OPUS_14PX.md
@@ -20,7 +20,7 @@ density, matching how fable's matrix cells are scored at fable's `5x8`.
   pageWidthPx}`; `run*.py` **preflight-aborts before any API call** if the
   corpus manifest model ≠ the model under test (guards silent profile/model
   mismatch).
-- Model reached via **upstream (`:8082`), pxpipe-bypassed** — so pxpipe's own
+- Model reached via **the gateway upstream, pxpipe-bypassed** — so pxpipe's own
   text→image rendering does not contaminate the arms; the model still gets the
   images. Concurrency 3, `CCI_READY_TIMEOUT=120`, `CCI_TIMEOUT=210`.
 - Raw results: `work/results.opus-14px.jsonl`, `work2/…`, `work3/…`.

diff --git a/eval/glyph-matrix/PLAN.md b/eval/glyph-matrix/PLAN.md
@@ -7,7 +7,7 @@
 ## What this measures
 
 Per-character confusion matrix for 12-char hex IDs read back through the
-pixelpipe render path, across render-style arms — the precision-tier
+pxpipe render path, across render-style arms — the precision-tier
 failure mode that gist-recall (98/98) structurally cannot surface and
 needle-haystack only bracketed (~87% exact-match on worst-case hex at
 prod style).

diff --git a/eval/grok-density/QUALITY_SUITE.md b/eval/grok-density/QUALITY_SUITE.md
@@ -4,7 +4,7 @@ Grok is evaluated through the OpenAI-compatible Responses endpoint used by
 Codex. Fable and Opus use the Claude harnesses.
 
 ```bash
-export OPENAI_BASE_URL=http://127.0.0.1:8082/v1
+export OPENAI_BASE_URL=http://127.0.0.1:<GATEWAY_PORT>/v1
 export OPENAI_API_KEY=…
 export SOL_QUALITY_MODEL=grok-4.5
 export SOL_QUALITY_LIVE=1

diff --git a/eval/grok-density/native-sweep/README.md b/eval/grok-density/native-sweep/README.md
@@ -10,7 +10,7 @@ atlas swap → rebuild → render, live ask locks answers before truth is scored
 ```bash
 node eval/grok-density/native-sweep/gen-fixture.mjs
 node eval/grok-density/native-sweep/render-ladder.mjs
-OPENAI_BASE_URL=http://127.0.0.1:8082/v1 OPENAI_API_KEY=… \
+OPENAI_BASE_URL=http://127.0.0.1:<GATEWAY_PORT>/v1 OPENAI_API_KEY=… \
   bash eval/grok-density/native-sweep/_ask_all.sh
 node eval/grok-density/native-sweep/score-all.mjs
 ```

diff --git a/eval/grok-density/native-sweep/RESULTS.md b/eval/grok-density/native-sweep/RESULTS.md
@@ -1,7 +1,7 @@
 # Grok 4.5 native JetBrains Mono 8–16px blind sweep
 
 Date: 2026-07-23. Model: `grok-4.5` via direct Responses (`OPENAI_BASE_URL`
-upstream `:8082`, pxpipe port 47821 rejected by the client).
+upstream local gateway; pxpipe port 47821 rejected by the client).
 
 ## Protocol (same spirit as Opus / Sol native sweeps)
 
@@ -58,7 +58,7 @@ evidence; this sweep is the native-size negative result.
 # atlases already built under eval/opus-density/native-sweep/atlases (symlinked)
 node eval/grok-density/native-sweep/gen-fixture.mjs   # fresh truth (optional)
 node eval/grok-density/native-sweep/render-ladder.mjs # swaps atlas*, rebuilds, restores
-OPENAI_BASE_URL=http://127.0.0.1:8082/v1 OPENAI_API_KEY=… \
+OPENAI_BASE_URL=http://127.0.0.1:<GATEWAY_PORT>/v1 OPENAI_API_KEY=… \
   bash eval/grok-density/native-sweep/_ask_all.sh     # locks answers-*.json
 node eval/grok-density/native-sweep/score-all.mjs
 ```

diff --git a/eval/grok-density/native-sweep/ask.mjs b/eval/grok-density/native-sweep/ask.mjs
@@ -1,6 +1,6 @@
 // Live-call grok-4.5 on one rendered rung. Blind: never opens truth.json.
 // Writes answers-<label>.json with per-token answer + conf.
-// Run: OPENAI_BASE_URL=http://127.0.0.1:8082/v1 OPENAI_API_KEY=… \
+// Run: OPENAI_BASE_URL=http://127.0.0.1:<GATEWAY_PORT>/v1 OPENAI_API_KEY=… \
 //      node eval/grok-density/native-sweep/ask.mjs jbmono14
 import { readFileSync, writeFileSync, readdirSync } from 'node:fs';
 import { dirname, join } from 'node:path';

diff --git a/eval/grok-profile/QUALITY_RESULTS.md b/eval/grok-profile/QUALITY_RESULTS.md
@@ -26,7 +26,7 @@ Receipts:
 - `verbatim-hex-grok-4.6-results.json`
 
 ```bash
-export OPENAI_BASE_URL=http://127.0.0.1:8082/v1
+export OPENAI_BASE_URL=http://127.0.0.1:<GATEWAY_PORT>/v1
 export OPENAI_API_KEY=…
 GROK_QUALITY_LIVE=1 N=100 node eval/grok-profile/novel-arithmetic.mjs
 GROK_QUALITY_LIVE=1 node eval/grok-profile/gist-recall.mjs

diff --git a/eval/qwen-profile/QUALITY_RESULTS.md b/eval/qwen-profile/QUALITY_RESULTS.md
@@ -0,0 +1,33 @@
+# Qwen 3.8 quality results
+
+Model: `workers-ai/@cf/qwen/qwen3.8-27b` (`qwen 3.8`) through Cloudflare Workers AI (via a local gateway).
+
+## Benchmark Summary
+
+| Geometry | Arithmetic (N=100) | Gist (N=98) | State (N=18) | Never-Stated (N=16) | Dense Hex (N=15) | Paired Pilot (8 facts) |
+|---|---:|---:|---:|---:|---:|---:|
+| **Spleen 5×8 (152 cols)** | 98/100 | 72/98 | 11/18 | **0/16** | 0/15 | 0/8 exact (unreadable / confabulated) |
+| **JetBrains Mono 14px (84 cols)** | — | — | — | — | **11/15** | **8/8 exact (100%)**, 2/2 gist, 2/2 guard |
+
+## Paired Pilot (Alpha / Beta)
+
+Evaluated against the paired synthetic terminal pilot (Alpha & Beta sessions with 4 exact-fact extractions + 1 gist + 1 unstated guard each):
+
+| Profile | Fixture | Exact Facts (4) | Gist | Guard (NOT STATED) | Notes |
+|---|---|---:|:---:|:---:|---|
+| Spleen 5×8 (152 cols) | Alpha | 0/4 | fail | fail | Raw output empty; font unreadable |
+| Spleen 5×8 (152 cols) | Beta | 0/4 | pass | pass | Confabulated `path` and `port` |
+| **JetBrains Mono 14px (84 cols)** | Alpha | **4/4** | **pass** | **pass** | `fingerprint`, `camelCase`, `path`, `port` exact |
+| **JetBrains Mono 14px (84 cols)** | Beta | **4/4** | **pass** | **pass** | `fingerprint`, `camelCase`, `path`, `port` exact |
+
+## Verbatim Hex (15 trials)
+
+- **Spleen 5×8 (152 cols)**: 0/15 exact matches. 5×8 bitmap glyphs are too dense for Qwen 3.8's vision encoder.
+- **JetBrains Mono 14px (84 cols)**: **11/15** exact 12-char hex matches (vs Sol 7/8, Grok 4/8).
+
+## Receipts
+
+- `novel-arithmetic-results.json`
+- `gist-recall-results.json`
+- `verbatim-hex-results.json` (Spleen 5×8 baseline)
+- `verbatim-hex-14px-results.json` (JetBrains Mono 14px pilot)