PR #82 — fix(frank-archon): fold the pinned host pool back to 49,152; keep the 2,048 step
PR #81 shipped two changes as one shape: a 66,560 MiB pinned host pool a rank and the
campaign's 2,048-token step. The pool half measured inert the same day on both payloads
that inherit this env, so this folds it back and keeps the step.
The measurement that decided it
payload
boot
vram_kv
engine's own line
frank-archon (iq4r-moe, fp8)
49,152
5.64 GiB
747,492 tokens, 93 % backing
frank-archon (same payload)
66,560
5.64 GiB — identical
pool opened 0 B used; elastic 26.73 / experts 20.72 / vram_weights 24.30 unchanged
qwen38-w5nl64 rig (bf16)
49,152
269.99 MiB
window 7,436 tokens
qwen38-w5nl64 rig (bf16)
66,560
271.25 MiB
a rounding step; window still 7,436 tokens
The rig's window is set by its own 95.37 GiB n-gram table claim and expert plane, not by the
host pool, so the 682 refused calibration documents are fixed by cutting them to the window
(scripts/rechunk-kld-corpus.py) — which the runner's own header law already states.
What changed
scripts/frank-archon.env — HOST_POOL_MIB 66,560 → 49,152 and the TTM pair
197,000 / 224,000 → 162,000 / 204,800 (the bats floor formula at this pool), with the A/B
recorded in the comment. MAX_BATCHED_TOKENS=2048 stays.
scripts/run-qwen38-bf16.sh — the step comment no longer claims the pool raise is the fix; it
states the measured 1.26 MiB and that the corpora follow the boot ledger.
tests/frank_archon_unit.bats, tests/qwen38_reference.bats — the pool assertion and the SHAPE fixture follow the env.
serverAppConfigs-docs/LLM/frank-archon.md — shape block back to 49,152; the live-shape ledger
keeps both boots (identical above host_pool); the Eras row for 2026-10-20 → 10-29 records the
fold-back; the pool evidence table marks 66,560 as the A/B row.
Verification
./tests/run.sh — 576 tests, 0 failures.
Rollback
Revert this commit: the pair returns to 66,560 / 197,000 / 224,000 and the Eras row's rollback
cell says so.
# PR #82 — fix(frank-archon): fold the pinned host pool back to 49,152; keep the 2,048 step
PR #81 shipped two changes as one shape: a 66,560 MiB pinned host pool a rank and the
campaign's 2,048-token step. The pool half measured **inert the same day** on both payloads
that inherit this env, so this folds it back and keeps the step.
## The measurement that decided it
| payload | boot | `vram_kv` | engine's own line |
|---|---|---|---|
| frank-archon (iq4r-moe, fp8) | 49,152 | 5.64 GiB | `747,492 tokens`, 93 % backing |
| frank-archon (same payload) | 66,560 | **5.64 GiB — identical** | pool opened `0 B used`; `elastic 26.73 / experts 20.72 / vram_weights 24.30` unchanged |
| qwen38-w5nl64 rig (bf16) | 49,152 | 269.99 MiB | window 7,436 tokens |
| qwen38-w5nl64 rig (bf16) | 66,560 | 271.25 MiB | a rounding step; window still 7,436 tokens |
The rig's window is set by its own 95.37 GiB n-gram table claim and expert plane, not by the
host pool, so the 682 refused calibration documents are fixed by cutting them to the window
(`scripts/rechunk-kld-corpus.py`) — which the runner's own header law already states.
## What changed
- `scripts/frank-archon.env` — `HOST_POOL_MIB` 66,560 → **49,152** and the TTM pair
197,000 / 224,000 → **162,000 / 204,800** (the bats floor formula at this pool), with the A/B
recorded in the comment. `MAX_BATCHED_TOKENS=2048` stays.
- `scripts/run-qwen38-bf16.sh` — the step comment no longer claims the pool raise is the fix; it
states the measured 1.26 MiB and that the corpora follow the boot ledger.
- `tests/frank_archon_unit.bats`, `tests/qwen38_reference.bats` — the pool assertion and the
`SHAPE` fixture follow the env.
- `serverAppConfigs-docs/LLM/frank-archon.md` — shape block back to 49,152; the live-shape ledger
keeps both boots (identical above `host_pool`); the Eras row for 2026-10-20 → 10-29 records the
fold-back; the pool evidence table marks 66,560 as the A/B row.
## Verification
- `./tests/run.sh` — **576 tests, 0 failures**.
## Rollback
Revert this commit: the pair returns to 66,560 / 197,000 / 224,000 and the Eras row's rollback
cell says so.
---
cwd: /home/blank/.omp/wt/wt-20261027-142335-6d1fc1f · session: 01a17f3f-3e2e-7509-aef8-be516c27cf98
PR #81's pool half measured inert the day it deployed: the 66,560 boot printed the
same elastic 26.73 GiB / experts 20.72 / kv cache 5.64 / vram_weights 24.30 as the
49,152 boot and opened the pool with 0 B used, and the qwen38-w5nl64 rig moved
vram_kv 269.99 -> 271.25 MiB (window still 7,436 tokens: its 95.37 GiB n-gram table
claim decides, not the pool). The step half stays. Tests: 576 ok / 0 fail.
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
PR #82 — fix(frank-archon): fold the pinned host pool back to 49,152; keep the 2,048 step
PR #81 shipped two changes as one shape: a 66,560 MiB pinned host pool a rank and the
campaign's 2,048-token step. The pool half measured inert the same day on both payloads
that inherit this env, so this folds it back and keeps the step.
The measurement that decided it
vram_kv747,492 tokens, 93 % backing0 B used;elastic 26.73 / experts 20.72 / vram_weights 24.30unchangedThe rig's window is set by its own 95.37 GiB n-gram table claim and expert plane, not by the
host pool, so the 682 refused calibration documents are fixed by cutting them to the window
(
scripts/rechunk-kld-corpus.py) — which the runner's own header law already states.What changed
scripts/frank-archon.env—HOST_POOL_MIB66,560 → 49,152 and the TTM pair197,000 / 224,000 → 162,000 / 204,800 (the bats floor formula at this pool), with the A/B
recorded in the comment.
MAX_BATCHED_TOKENS=2048stays.scripts/run-qwen38-bf16.sh— the step comment no longer claims the pool raise is the fix; itstates the measured 1.26 MiB and that the corpora follow the boot ledger.
tests/frank_archon_unit.bats,tests/qwen38_reference.bats— the pool assertion and theSHAPEfixture follow the env.serverAppConfigs-docs/LLM/frank-archon.md— shape block back to 49,152; the live-shape ledgerkeeps both boots (identical above
host_pool); the Eras row for 2026-10-20 → 10-29 records thefold-back; the pool evidence table marks 66,560 as the A/B row.
Verification
./tests/run.sh— 576 tests, 0 failures.Rollback
Revert this commit: the pair returns to 66,560 / 197,000 / 224,000 and the Eras row's rollback
cell says so.
cwd: /home/blank/.omp/wt/wt-20261027-142335-6d1fc1f · session: 01a17f3f-3e2e-7509-aef8-be516c27cf98