Ty_Tech/SlabUploader
6
0
Fork 0
Code Issues 15 Pull Requests Actions Packages Projects Releases Wiki Activity

docs(m1): calibrated worker delegation policy (prompt framing, verify-before-land)

Browse Source
This commit is contained in:
Ty
2026-09-13 22:48:48 -07:00
parent 92d1a7ec5e
commit ba8e5f72f6
1 changed files with 10 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs/M1-POR.md
+10 -6
View File
@@ -48,15 +48,19 @@ See `deploy/README.md` §Milestone 1 UAT — deployment POR: one container, host
Two local inference endpoints are available as dispatchable workers for test / dev / QA:
- **`.202:8081`** — `gemma-4-12B-it-q4_k_m` (12B, Q4) — **completion only, no vision.** Use for text/code tasks: drafting, code review, test-writing help.
- **`.202:8082`** — `Qwen3.8-27B-UD-IQ3_XXS` (27B, IQ3) — **completion + multimodal (vision).** This is the **vision worker**: any task that must look at a slab photo (bg-removal tuning, mask/axis calibration review, fixture sanity checks) goes here.
- **`.202:8081`** — `gemma-4-12B-it-q4_k_m` (12B, Q4) — **completion only, no vision.** Use for lighter text: drafting, code review, summaries.
- **`.202:8082`** — `Qwen3.8-27B-UD-IQ3_XXS` (27B, IQ3) — **completion + multimodal (vision).** The **vision worker** (any task that must look at a slab photo) and the stronger general-reasoning endpoint.
Coordinator: Hermes (me) — plans, decomposes, dispatches, verifies, and does work directly when a card is small or judgment-heavy.
**Rules:**
- **Vision tasks → `.202:8082` only.**
- **Never more than one task on an endpoint at a time.** Two concurrent jobs on the same endpoint back up and overheat. Serialize dispatch; one endpoint = one active task.
- Both run low-quant (Q4 / IQ3). Fine for dispatch-style assist; deep sustained reasoning is better done by the coordinator directly or on a higher-quant model.
### Calibrated delegation policy (2026-09-14)
- **Frame every prompt**: system role + "show each step" + explicit constraints. Small/low-quant models drop multi-step arithmetic in one-shot "reply JSON only" prompts but get it exactly right with scratch-space framing. Do not judge capability from under-framed probes.
- **The deterministic core stays the coordinator's** (TDD'd, unit-tested). A model that's "usually right" on math is not good enough for the spec — math is cheap to test and must be exact.
- **Verify before land**: any load-bearing model output is checked against tests/spec before entering the repo. Workers are assist, not autonomous builders.
- **Output quirks**: gemma wraps JSON in ``` fences — strip before parsing. Vision calls run ~18s/image (text ~1s).
- **One task per endpoint at a time** (no parallel on the same endpoint — backs up and overheats).
- Both run low-quant (Q4/IQ3). Fine for framed assist; deep sustained reasoning is best done by the coordinator directly.
## Test plan (M1)