The calibration floor goes back to 12,000,000 tokens, and the repo default now matches the run.
The served iq4r this arm competes with was calibrated on a 10 M-token set, so the arm must be
calibrated at least as deeply. The 1.5 M value came from the vendor's note that a pooled Gram from
10 M tokens beat one from 1 M by under 1% — that measures the vendor's own plate, not ours: cutting
our calibration 6x to save wall clock buys a worse position on the only comparison that decides
whether the arm ships.
scripts/run-qwen38-bf16.sh: RADIANCE_CALIB_TOKENS default 1500000 -> 12000000, with the law
in the header; also corrects the pointed-at corpus to the re-chunked calib-7168.jsonl.
tests/qwen38_reference.bats: the calib dry run asserts the 12 M floor (and the latest pin).
tests/qwen38_campaign.bats: the calibration-window dry run now passes Q38_CALIB_MIN_FREE_MIB=1, because its fixture lands on /tmp (a 16 GiB tmpfs) where the 60 GiB
seal preflight can never pass. That test was asserting environment luck, and it fails the same way
on main today.
Consequences, stated plainly
One round of the re-chunked corpus is ~9.31 M tokens, so a 12 M floor seals in round 1 plus part of
round 2. A round that seals nothing is normal — the floor is pooled and round_verdict answers
"continue", never failure (PR #85). At the rig's measured ~70 tok/s the window holds both cards for ~48 h; that is the price of the deeper calibration.
Testing
./tests/run.sh -> ALL TESTS PASSED (589). The live window on dankserver already carries FRANK_QWEN38_CALIB_TOKENS=12000000 in its environment, so the run and the code agree.
## What
The calibration floor goes back to **12,000,000 tokens**, and the repo default now matches the run.
The served iq4r this arm competes with was calibrated on a 10 M-token set, so the arm must be
calibrated at least as deeply. The 1.5 M value came from the vendor's note that *a pooled Gram from
10 M tokens beat one from 1 M by under 1%* — that measures the vendor's own plate, not ours: cutting
our calibration 6x to save wall clock buys a worse position on the only comparison that decides
whether the arm ships.
- `scripts/run-qwen38-bf16.sh`: `RADIANCE_CALIB_TOKENS` default `1500000` -> `12000000`, with the law
in the header; also corrects the pointed-at corpus to the re-chunked `calib-7168.jsonl`.
- `tests/qwen38_reference.bats`: the calib dry run asserts the 12 M floor (and the latest pin).
- `tests/qwen38_campaign.bats`: the calibration-window dry run now passes
`Q38_CALIB_MIN_FREE_MIB=1`, because its fixture lands on `/tmp` (a 16 GiB tmpfs) where the 60 GiB
seal preflight can never pass. That test was asserting environment luck, and it fails the same way
on main today.
## Consequences, stated plainly
One round of the re-chunked corpus is ~9.31 M tokens, so a 12 M floor seals in **round 1 plus part of
round 2**. A round that seals nothing is normal — the floor is pooled and `round_verdict` answers
"continue", never failure (PR #85). At the rig's measured ~70 tok/s the window holds both cards for
**~48 h**; that is the price of the deeper calibration.
## Testing
`./tests/run.sh` -> **ALL TESTS PASSED** (589). The live window on dankserver already carries
`FRANK_QWEN38_CALIB_TOKENS=12000000` in its environment, so the run and the code agree.
---
cwd: /home/blank/.omp/wt/wt-20261027-142335-6d1fc1f · session: 01a17f3f-3e2e-7509-aef8-be516c27cf98
Frank's rule: the served iq4r this arm competes with was calibrated on a
10M-token set, so the arm is calibrated at least as deeply. The floor was
lowered to 1.5M on the vendor's '10M beats 1M by under 1%' note; that
measures the vendor's own plate, not ours — the arm is judged against a
10M-calibrated reference, so cutting the calibration 6x to save wall clock
buys a worse position on the only comparison that matters.
One round of the re-chunked corpus is ~9.31M tokens, so a 12M floor seals
in round 1 plus part of round 2. A round that seals nothing is normal (the
floor is pooled) and round_verdict treats it as continue, never failure.
Also fixes tests/qwen38_campaign.bats' window dry run: its fixture lands on
/tmp (a 16 GiB tmpfs) where the 60 GiB seal preflight can never pass, so the
test now uses the documented Q38_CALIB_MIN_FREE_MIB override instead of
asserting environment luck.
Frank's rule: the served iq4r this arm competes with was calibrated on a
10M-token set, so the arm is calibrated at least as deeply. The floor was
lowered to 1.5M on the vendor's '10M beats 1M by under 1%' note; that
measures the vendor's own plate, not ours — the arm is judged against a
10M-calibrated reference, so cutting the calibration 6x to save wall clock
buys a worse position on the only comparison that matters.
One round of the re-chunked corpus is ~9.31M tokens, so a 12M floor seals
in round 1 plus part of round 2. A round that seals nothing is normal (the
floor is pooled) and round_verdict treats it as continue, never failure.
Also fixes tests/qwen38_campaign.bats' window dry run: its fixture lands on
/tmp (a 16 GiB tmpfs) where the 60 GiB seal preflight can never pass, so the
test now uses the documented Q38_CALIB_MIN_FREE_MIB override instead of
asserting environment luck.
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.
What
The calibration floor goes back to 12,000,000 tokens, and the repo default now matches the run.
The served iq4r this arm competes with was calibrated on a 10 M-token set, so the arm must be
calibrated at least as deeply. The 1.5 M value came from the vendor's note that a pooled Gram from
10 M tokens beat one from 1 M by under 1% — that measures the vendor's own plate, not ours: cutting
our calibration 6x to save wall clock buys a worse position on the only comparison that decides
whether the arm ships.
scripts/run-qwen38-bf16.sh:RADIANCE_CALIB_TOKENSdefault1500000->12000000, with the lawin the header; also corrects the pointed-at corpus to the re-chunked
calib-7168.jsonl.tests/qwen38_reference.bats: the calib dry run asserts the 12 M floor (and the latest pin).tests/qwen38_campaign.bats: the calibration-window dry run now passesQ38_CALIB_MIN_FREE_MIB=1, because its fixture lands on/tmp(a 16 GiB tmpfs) where the 60 GiBseal preflight can never pass. That test was asserting environment luck, and it fails the same way
on main today.
Consequences, stated plainly
One round of the re-chunked corpus is ~9.31 M tokens, so a 12 M floor seals in round 1 plus part of
round 2. A round that seals nothing is normal — the floor is pooled and
round_verdictanswers"continue", never failure (PR #85). At the rig's measured ~70 tok/s the window holds both cards for
~48 h; that is the price of the deeper calibration.
Testing
./tests/run.sh-> ALL TESTS PASSED (589). The live window on dankserver already carriesFRANK_QWEN38_CALIB_TOKENS=12000000in its environment, so the run and the code agree.cwd: /home/blank/.omp/wt/wt-20261027-142335-6d1fc1f · session: 01a17f3f-3e2e-7509-aef8-be516c27cf98