@@ -141,3 +141,36 @@ Initial sampling, draft prefill and catch-up after plain rounds can cost more
 than speculation saves. No speed improvement is guaranteed. No NVIDIA run,
 MTP-family support, broad corpus qualification, statistical paired speed
 claim, commit/push or deployment is established by this checkpoint.
+
+## Additional 7B diagnostic and hosted delivery
+
+Executable revision `501e35d` is pushed; full hosted verification and public
+JSON publication succeed (see `CiPublicationAudit-2026-10-05.md`). Hosted
+Metal skips are real: Apple's paravirtual device reports GPU family zero.
+Local real Metal coverage above is separate from that hosted evidence.
+
+An additional existing Qwen2.5-7B-Instruct-1M Q8_0 source was losslessly
+prepared with the C# compiler, source SHA-256
+`d81ca8ca04442ed5ff7d036fa546183170b8269281380bed977431a5449c84b2` and package
+identity `9c1e2aca1f23bcc51f06802cac5e9e25e2835e0f9a6988faf92971a2c3986a16`
+(the identity is not a whole-file hash). External draft: prepared 0.5B Q8_0,
+without layer drop. Metal/f16 KV, eight threads, context 512, the same
+37-token chat prompt as the earlier 7B diagnostic, and 256 generated tokens.
+One warm-up per mode and three alternating measured fresh-process pairs.
+Final compiler/helper/12 generation processes and the C# reporter exit 0;
+every fixed and auto output has all 256 tokens in the same order as its paired
+dense output. Raw prompt, tokens, JSON/stderr, C# reporter, orchestration,
+summary and fuller limits are under `artifacts/adaptive-speculation/seven-billion/`.
+
+| Mode | Median generation wall ms | Median reported decode tokens/s |
+|---|---:|---:|
+| Dense | 13085.350 | 20.313 |
+| Fixed 3 | 10466.081 | 26.105 |
+| Auto cap 3 | 12371.546 | 24.239 |
+
+Fixed 3 remains fastest here, and remains the default. Adaptive choices vary
+by round, including zero in the third. Other user-owned CPU work exists on
+the shared host; identical load/thermal conditions are not established.
+Only three pairs on one prompt/model pair, with no semantic answer-quality
+or statistical speedup verdict. This does not close `TASK-SPC-005` or add
+these local diagnostic rows to the public hosted benchmark stream.
@@ -79,3 +79,36 @@ unsafe content injection, schema or data ownership change. Pinned Actions,
 run/attempt freshness, hash validation and bounded root-member extraction stay
 in place. New hosted completion and public source-revision confirmation are
 required before this task is marked verified.
+
+## Delivered executable/workflow revision `501e35d`
+
+[verify #28](https://github.com/managedcode/Synapse/actions/runs/37290201607),
+[performance #16](https://github.com/managedcode/Synapse/actions/runs/37290201688)
+and the final [Pages deployment](https://github.com/managedcode/Synapse/actions/runs/37293092158)
+all succeed. .NET, Rust checks and JSON publication succeed on every runner.
+
+| .NET runner | Total | Passed | Failed | Skipped |
+|---|---:|---:|---:|---:|
+| osx-arm64 | 577 | 530 | 0 | 47 |
+| linux-x64 | 577 | 529 | 0 | 48 |
+| win-x64 | 577 | 529 | 0 | 48 |
+
+The new performance report contains 22/22 expected artifacts, no missing or
+invalid artifacts, and 124 rows: 12 matched and 112 unreviewed variants.
+The canonical HTTPS JSON equals the uploaded Pages artifact byte-independent
+JSON projection exactly. Its generation time is
+`2026-10-05T09:56:20.2981639+00:00`; verification and performance both retain
+source SHA `501e35dbeab34468205f3e5f34679daa050f9c09`. Every original benchmark
+row field/value is unchanged by site publication (optional model enrichment
+is excluded from that equality comparison). No missing test artifacts.
+The live browser renders these new counts, scopes, run numbers and SHA.
+
+The performance-triggered publication initially kept the still-running
+verification stream's previous completed SHA, explicitly shown separately.
+Verification completion automatically caused a second publication with both
+new streams. Its deploy job queued without an allocated runner or pending
+approval, then completed successfully; it was not cancelled or bypassed.
+The red/green source regression checks failure conditions; no hosted failure
+or timeout branch was deliberately fault-injected. Raw final JSON, producer
+artifacts, Pages artifact and watch logs remain in the ignored audit directory.
+This post-delivery record changes documentation only.
@@ -5,7 +5,9 @@
 - [x] Inspect hosted runs, actual test JSON and the live metrics endpoint.
 - [x] Red contract test, independent Rust checks and bounded .NET step timeout.
 - [x] Website 25/25, format/build and workflow lint; preserve actual skips.
-- [ ] New hosted delivery evidence (see `website.plan.md`).
+- [x] Hosted `verify` 37290201607, performance 37290201688 and Pages
+  37293092158 succeed for executable/workflow revision `501e35d`; canonical
+  public JSON and rendered counts match the uploaded artifact exactly.
 
 ## 2026-10-05 measured speculative depth (`TASK-SPC-005`)
 
@@ -15,6 +17,9 @@
 - [x] Format/build/analyzers, 18 focused tests, coverage, Rust and manual
   security/architecture review. Full .NET gate remains unverified: the local
   35-minute run was stopped after a reproduced evaluation child timeout.
+  Subsequent hosted full suites succeed: 577 discovered tests per OS, zero
+  failures, 47/48 genuine skips; this does not resolve the local timeout or
+  confer unavailable GPU qualification.
 - [ ] Paired real-model speed/quality qualification including short requests,
   warm-up and drift; broader runtime and benchmark tasks stay open.
 
 
@@ -52,4 +52,8 @@ This does not close the statistical benchmark or broader runtime plans.
   independent of earlier failures after toolchain installation.
 - [x] Focused Website 25/25, format/build/workflow lint and evidence review;
   `docs/Development/CiPublicationAudit-2026-10-05.md` records exact scope.
-- [ ] Deliver the workflow and confirm a new hosted verification/publication.
+- [x] Deliver the workflow at `501e35d` and confirm hosted verification,
+  performance JSON and Pages; the canonical endpoint and browser display the
+  exact new artifact, including 577 tests per OS and 124 benchmark rows.
+  Failure conditions have a red/green source contract; hosted failure/timeout
+  branches were not deliberately fault-injected. Broader qualification stays open.