PR #81 — feat(frank-archon): raise the pinned host pool to 66,560; the campaign inherits MNBT 2,048
One shape, two files of numbers and the tests that pin them. Frank chose this shape
in the grilling round on the campaign's measuring geometry: both the host pool and
the step size move, as one deploy, because the pool is what widens the window and the
step is what the quantised expert file tiers can actually stream.
What changed
scripts/frank-archon.env — HOST_POOL_MIB 49,152 → 66,560 a rank (tp × 65 GiB
against the 246 GiB the box has) and the TTM pair 162,000 / 204,800 → 197,000 / 224,000 (the floor the bats ledger formula computes at the new pool: 2 × pool + 48,824 + prefix 8,192 + MNBT 2,048 + seqs × 10 × 112). The pool is the
KV lever under the budget resolver — every MiB pinned in host RAM lowers the expert
floor on the card one-for-one and the cache takes it.
scripts/run-qwen38-bf16.sh — the campaign's private FRANK_QWEN38_MNBT (default
1,024) is deleted; MNBT is now ${FRANK_ARCHON_MAX_BATCHED_TOKENS:?…},
production's 2,048, the widest step the quantised expert file tiers stream. The
1,024 was PR #80's answer to a starved cache; the pool raise is the fix and a narrow
step was the symptom's disguise. The audit law now states it: a mode that carries its
own step measures a shape nothing serves.
tests/frank_archon_unit.bats — the pool assertion follows the env (66560).
tests/qwen38_reference.bats — the SHAPE fixture carries the new pool, and the two FRANK_QWEN38_MNBT=1024 override tests become the inheritance assertion: every
mode's argv quotes the value the env file owns (ref_step/score_step/calib all
equal awk'd FRANK_ARCHON_MAX_BATCHED_TOKENS). The calib-shape paragraphs stay —
they are the 65,536-step OOM record.
serverAppConfigs-docs/LLM/frank-archon.md — shape block moves to the new numbers and
the Eras table gains the row for the era this replaces (2026-10-20 → 10-29,
payload unchanged, rollback = revert this commit). The 2026-10-20 boot-ledger line
stays as that era's record; the new boot's ledger replaces it in the follow-up commit,
the way PR #80 → #81 has been handled.
Verification
./tests/run.sh — 576 tests, 0 failures (the first run failed one test: the SHAPE fixture still pinned --host-pool-mib 49152; fixed in the same commit set).
Deploy has not run yet. Merging, the deploy is scripts/deploy.sh dankserver --service frank-archon; the campaign's live calib run
is killed first, and the new boot ledger arbitrates the cache line before the corpora
are re-cut.
Rollback
Revert 255796b, redeploy the service. The env returns to 49,152 / 162,000 / 204,800 and
the runner to its own FRANK_QWEN38_MNBT; the Eras row carries the same instruction.
# PR #81 — feat(frank-archon): raise the pinned host pool to 66,560; the campaign inherits MNBT 2,048
One shape, two files of numbers and the tests that pin them. Frank chose this shape
in the grilling round on the campaign's measuring geometry: **both** the host pool and
the step size move, as one deploy, because the pool is what widens the window and the
step is what the quantised expert file tiers can actually stream.
## What changed
- `scripts/frank-archon.env` — `HOST_POOL_MIB` 49,152 → **66,560** a rank (tp × 65 GiB
against the 246 GiB the box has) and the TTM pair 162,000 / 204,800 →
**197,000 / 224,000** (the floor the bats ledger formula computes at the new pool:
`2 × pool + 48,824 + prefix 8,192 + MNBT 2,048 + seqs × 10 × 112`). The pool is the
KV lever under the budget resolver — every MiB pinned in host RAM lowers the expert
floor on the card one-for-one and the cache takes it.
- `scripts/run-qwen38-bf16.sh` — the campaign's private `FRANK_QWEN38_MNBT` (default
1,024) is **deleted**; `MNBT` is now `${FRANK_ARCHON_MAX_BATCHED_TOKENS:?…}`,
production's 2,048, the widest step the quantised expert file tiers stream. The
1,024 was PR #80's answer to a starved cache; the pool raise is the fix and a narrow
step was the symptom's disguise. The audit law now states it: a mode that carries its
own step measures a shape nothing serves.
- `tests/frank_archon_unit.bats` — the pool assertion follows the env (66560).
- `tests/qwen38_reference.bats` — the `SHAPE` fixture carries the new pool, and the two
`FRANK_QWEN38_MNBT=1024` override tests become the **inheritance assertion**: every
mode's argv quotes the value the env file owns (`ref_step`/`score_step`/calib all
equal `awk`'d `FRANK_ARCHON_MAX_BATCHED_TOKENS`). The calib-shape paragraphs stay —
they are the 65,536-step OOM record.
- `serverAppConfigs-docs/LLM/frank-archon.md` — shape block moves to the new numbers and
the **Eras** table gains the row for the era this replaces (2026-10-20 → 10-29,
payload unchanged, rollback = revert this commit). The 2026-10-20 boot-ledger line
stays as that era's record; the new boot's ledger replaces it in the follow-up commit,
the way PR #80 → #81 has been handled.
## Verification
- `./tests/run.sh` — **576 tests, 0 failures** (the first run failed one test: the
`SHAPE` fixture still pinned `--host-pool-mib 49152`; fixed in the same commit set).
- Deploy has not run yet. Merging, the deploy is
`scripts/deploy.sh dankserver --service frank-archon`; the campaign's live calib run
is killed first, and the new boot ledger arbitrates the cache line before the corpora
are re-cut.
## Rollback
Revert `255796b`, redeploy the service. The env returns to 49,152 / 162,000 / 204,800 and
the runner to its own `FRANK_QWEN38_MNBT`; the Eras row carries the same instruction.
---
cwd: /home/blank/.omp/wt/wt-20261027-142335-6d1fc1f · session: 01a17f3f-3e2e-7509-aef8-be516c27cf98
frank-archon.env: HOST_POOL_MIB 49,152 -> 66,560 a rank (tp x 65 GiB against the
246 GiB the box has), and the TTM pair 162,000/204,800 -> 197,000/224,000 — the
floor the bats ledger formula computes at the new pool. The pool is the KV lever
under the budget resolver: every MiB pinned in host RAM lowers the expert floor on
the card one-for-one and the cache takes it.
scripts/run-qwen38-bf16.sh: the campaign's private FRANK_QWEN38_MNBT (1,024) is
gone — MNBT now comes from FRANK_ARCHON_MAX_BATCHED_TOKENS, production's 2,048 and
the widest step the quantised expert file tiers stream. That 1,024 was PR #80's
answer to a starved cache; the pool is the fix, a narrow step was the symptom's
disguise, and a per-mode literal measures a shape nothing serves.
tests: frank_archon_unit greps the new pool value; qwen38_reference's SHAPE fixture
carries it, and the two MNBT-override tests become the inheritance assertion (every
mode's argv quotes the value the env file owns) plus the calib assertions.
vault: the shape block moves to the new numbers and the Eras table gains the row for
the era this replaces. The 2026-10-20 boot-ledger line stays as the old era's record
until the new boot's ledger replaces it.
dank
merged commit f83c241668 into main2026-10-29 13:14:27 -07:00
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 #81 — feat(frank-archon): raise the pinned host pool to 66,560; the campaign inherits MNBT 2,048
One shape, two files of numbers and the tests that pin them. Frank chose this shape
in the grilling round on the campaign's measuring geometry: both the host pool and
the step size move, as one deploy, because the pool is what widens the window and the
step is what the quantised expert file tiers can actually stream.
What changed
scripts/frank-archon.env—HOST_POOL_MIB49,152 → 66,560 a rank (tp × 65 GiBagainst the 246 GiB the box has) and the TTM pair 162,000 / 204,800 →
197,000 / 224,000 (the floor the bats ledger formula computes at the new pool:
2 × pool + 48,824 + prefix 8,192 + MNBT 2,048 + seqs × 10 × 112). The pool is theKV lever under the budget resolver — every MiB pinned in host RAM lowers the expert
floor on the card one-for-one and the cache takes it.
scripts/run-qwen38-bf16.sh— the campaign's privateFRANK_QWEN38_MNBT(default1,024) is deleted;
MNBTis now${FRANK_ARCHON_MAX_BATCHED_TOKENS:?…},production's 2,048, the widest step the quantised expert file tiers stream. The
1,024 was PR #80's answer to a starved cache; the pool raise is the fix and a narrow
step was the symptom's disguise. The audit law now states it: a mode that carries its
own step measures a shape nothing serves.
tests/frank_archon_unit.bats— the pool assertion follows the env (66560).tests/qwen38_reference.bats— theSHAPEfixture carries the new pool, and the twoFRANK_QWEN38_MNBT=1024override tests become the inheritance assertion: everymode's argv quotes the value the env file owns (
ref_step/score_step/calib allequal
awk'dFRANK_ARCHON_MAX_BATCHED_TOKENS). The calib-shape paragraphs stay —they are the 65,536-step OOM record.
serverAppConfigs-docs/LLM/frank-archon.md— shape block moves to the new numbers andthe Eras table gains the row for the era this replaces (2026-10-20 → 10-29,
payload unchanged, rollback = revert this commit). The 2026-10-20 boot-ledger line
stays as that era's record; the new boot's ledger replaces it in the follow-up commit,
the way PR #80 → #81 has been handled.
Verification
./tests/run.sh— 576 tests, 0 failures (the first run failed one test: theSHAPEfixture still pinned--host-pool-mib 49152; fixed in the same commit set).scripts/deploy.sh dankserver --service frank-archon; the campaign's live calib runis killed first, and the new boot ledger arbitrates the cache line before the corpora
are re-cut.
Rollback
Revert
255796b, redeploy the service. The env returns to 49,152 / 162,000 / 204,800 andthe runner to its own
FRANK_QWEN38_MNBT; the Eras row carries the same instruction.cwd: /home/blank/.omp/wt/wt-20261027-142335-6d1fc1f · session: 01a17f3f-3e2e-7509-aef8-be516c27cf98