Skip to content

Downstairs T12 Part A is next work before Halo #36

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
jmjava merged 5 commits into from
Sep 29, 2026
Merged

Downstairs T12 Part A is next work before Halo #36

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
c6b4277
Link Phase 3 eval protocol to the Nov–NYE Zenodo preprint plan.
cursoragent Sep 29, 2026
46958ac
Point Phase 3 paper track at the Mar 2027 DOI deadline.
cursoragent Sep 29, 2026
ebc0bf0
Schedule Phase 4 Halo / 395 bring-up for Black Friday week Nov 2026.
cursoragent Sep 29, 2026
622d22f
Plan early-Nov clear of downstairs T12 Part A before Halo.
cursoragent Sep 29, 2026
03b7f3b
Mark downstairs T12 Part A as next hardware work before Halo.
cursoragent Sep 29, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
8 changes: 7 additions & 1 deletion docs/evaluation-protocol.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -249,9 +249,15 @@ It still may not claim:

## Paper mapping

This protocol is the methods object for a short workshop paper on
This protocol is the methods object for a short workshop / preprint paper on
**local SLMs as MCP tools**, not as a replacement model provider.
Related MCP-agent benchmarks score tool choice or final answers on
hosted servers. The question here is narrower: when a premium agent
delegates a bounded coding edit to a private 9B–24B model, which
failure layer fires, and how much of that layer is the prompt contract?

**Nov 2026 – Mar 2027 ship plan** (venue, calendar, citations, claim freeze; DOI by
**31 Mar 2027**) lives in the model program repo:
[embabel-slm/docs/paper-calendar-2026.md](https://github.com/jmjava/embabel-slm/blob/main/docs/paper-calendar-2026.md).
Zenodo is the primary DOI path; live rates for that preprint still come from
this repo’s harness and dated notes.
34 changes: 26 additions & 8 deletions docs/roadmap.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -10,17 +10,35 @@ LAN addresses, or SKUs here.
| --- | --- | --- |
| 1 — inference host | Done on the workstation lab | Ollama + starter tags on a private GPU |
| 2 — MCP bridge | Done | `local-coding-slm` stdio tools; Cursor / Copilot / Claude adapters |
| 3 — measure | Protocol + expanded corpus + harder multi-file suite + MCP apply gate + CI; live rates still informal | Layered scoring; scripted premium routing/review; no auto-classifier yet |
| T12 Part A second NVIDIA host | Blocked on host power | WSL GPU via SSH; see [examples/downstairs-wsl-gpu.md](../examples/downstairs-wsl-gpu.md) |
| T12 Part B Copilot A8 / Claude A9 | Config ready; operator clicks pending | Same-machine only; [docs/a8-a9-operator-checklist.md](a8-a9-operator-checklist.md) |
| 3 — measure | Protocol + corpus + apply gate + CI; **live rates still need a stable remote GPU** | Layered scoring; next live rows go through downstairs |
| **T12 Part A downstairs** | **Next hardware work** (early Nov 2026) | Second PC / WSL NVIDIA via SSH — was blocked on host power; [examples/downstairs-wsl-gpu.md](../examples/downstairs-wsl-gpu.md) |
| T12 Part B Copilot A8 / Claude A9 | Config ready; operator clicks pending | Same-machine only; do not block on downstairs; [docs/a8-a9-operator-checklist.md](a8-a9-operator-checklist.md) |

Finish Phase 3 measurement (layered live rates on the committed corpus)
and the existing second-host checks before treating Halo as the next lab.
**Order (do not invert):** finish Phase 3 live measurement on **downstairs** (T12 Part A) **before** treating Halo as the next lab.

## Phase 4 — Halo host profile (future)
1. **First work — downstairs:** power on, WSL Ollama, `ssh -L` from the IDE workstation, A4-class check, scrubbed live harness / `run_eval.py --live` rows.
2. **Then — keep measuring** Part A on downstairs through November.
3. **Later — Halo / 395** (~Black Friday): bring-up and A13 only after downstairs is a known path (on, or explicitly abandoned).

Use an AMD Ryzen AI Halo-class box as **another private Ollama host**. The
MCP server stays on the workstation. Premium agents still plan and review.
Public-safe downstairs notes:
[examples/downstairs-wsl-gpu.md](../examples/downstairs-wsl-gpu.md).

**Halo / Ryzen AI Max+ 395:** planned acquire around **Black Friday week Nov 2026**.
Claim Phase 4 only after A13. Until then Phase 3 live rates stay on **downstairs**, not workstation-local Ollama and not Halo. Paper track:
[embabel-slm paper calendar](https://github.com/jmjava/embabel-slm/blob/main/docs/paper-calendar-2026.md).

Paper track: Nov 2026–Mar 2027 Zenodo preprint (DOI by **31 Mar 2027**). Live
`pass@1` / `pass@end` freeze by early January (**downstairs** default; Halo only if
A13 green); all paper numbers freeze by late February.

## Phase 4 — Halo host profile (planned ~Black Friday Nov 2026)

Use an AMD Ryzen AI Max+ 395 / Halo-class box as **another private Ollama host**.
Target arrival: **around Black Friday week November 2026**. The MCP server stays
on the workstation. Premium agents still plan and review.

Until the box exists and A13 passes, do not claim this phase — keep Phase 3 live
rates on the **downstairs** host (T12 Part A). Do not skip downstairs to chase Halo.

This is **not** a second product. It is the same bridge with a different
`OLLAMA_BASE_URL` (or the same loopback URL behind `ssh -L`).
Expand Down
Loading