Two asset slices have now burned their full 200-turn ceiling without committing, and the worker that
returned named the cause.
Measured
| slice |
turns |
committed |
outcome |
| #641 |
200 |
nothing |
orchestrator committed 14 room GLBs and the bake fix |
| #683 |
200 |
nothing |
orchestrator committed 12 humanoid GLBs, 4 skin maps, a manifest |
Both ran asset-pipeline-lead on deepseek-v4-flash. Both produced correct, contract-passing work.
Neither reached its own commit or report step.
What consumed the turns, in #683's own words
- The aisha height-solve non-convergence — the largest single item. The first re-bake refused at
the height solve.
- Bake-identity archaeology — the exact per-actor flags had to be reconstructed across the
#687 / #688 / #651 / #576 history.
The second is the interesting one. Per-actor bake flags are not recorded in one readable place, so
every asset slice that regenerates a humanoid pays to rediscover them from issue history. That cost
is proportional to how many actors the slice touches, and both ceiling-hitting slices touched all of
them.
Why this is not simply "raise maxTurns"
Raising the ceiling would let the archaeology finish rather than removing it, and the work would
still arrive uncommitted if the ceiling is hit again. The measured cost is a missing SSOT for
per-actor bake identity, not a budget shortfall.
It also has a second, cheaper half: a worker that spends its last turn on the fix has none left to
commit or report. Both slices needed an orchestrator commit and a resume for the report.
claimScope / notEvidenceFor
- claimScope: two slices, both
asset-pipeline-lead on deepseek-v4-flash, both regenerating
the full humanoid set, and one worker's own account of where its turns went.
- notEvidenceFor: that the ceiling is wrong for other roles or models; that archaeology is the
larger of the two costs (the worker ranked the height solve first); or where the per-actor flags
should live, which is a design decision this card does not make.
No ## done_when yet — the remedy is a judgement, and naming it here would be inventing a contract
from a title.
Two asset slices have now burned their full 200-turn ceiling without committing, and the worker that
returned named the cause.
Measured
Both ran
asset-pipeline-leadondeepseek-v4-flash. Both produced correct, contract-passing work.Neither reached its own commit or report step.
What consumed the turns, in #683's own words
the height solve.
#687/#688/#651/#576history.The second is the interesting one. Per-actor bake flags are not recorded in one readable place, so
every asset slice that regenerates a humanoid pays to rediscover them from issue history. That cost
is proportional to how many actors the slice touches, and both ceiling-hitting slices touched all of
them.
Why this is not simply "raise maxTurns"
Raising the ceiling would let the archaeology finish rather than removing it, and the work would
still arrive uncommitted if the ceiling is hit again. The measured cost is a missing SSOT for
per-actor bake identity, not a budget shortfall.
It also has a second, cheaper half: a worker that spends its last turn on the fix has none left to
commit or report. Both slices needed an orchestrator commit and a resume for the report.
claimScope / notEvidenceFor
asset-pipeline-leadondeepseek-v4-flash, both regeneratingthe full humanoid set, and one worker's own account of where its turns went.
larger of the two costs (the worker ranked the height solve first); or where the per-actor flags
should live, which is a design decision this card does not make.
No
## done_whenyet — the remedy is a judgement, and naming it here would be inventing a contractfrom a title.