A live research pursuit died with no terminal result because one budget reservation was still open when the supervisor reached its join barrier:
supervisePursuit: budget pool: 1 reservation(s) still open at join barrier (leaked ticket ids: 11) — conserved-pool invariant violated
Measured on 2026-09-10, Runtime 0.208.1, cli-bridge driver, opencode director on the Z.AI route, thirteen Tangle Sandbox children on opencode through the Tangle router.
What the run directory holds afterwards:
failure.json with that message, nodes: 0, and no spentTotal.- No
result.json.
- A spawn journal with 13 spawn events and 7
teardown-unconfirmed child events, the last two being extract:robert-ghrist:arXiv-2507.01226 and extract:robert-ghrist:arXiv-2010.11525.
So the pursuit's whole tree is unrepresented in any terminal record: the projection reports zero nodes while the journal shows thirteen children, several of which had already completed real inference. From an operator's seat the run simply vanishes, and the only way to learn what happened is to read the raw journal.
The invariant itself is right — a conserved pool should not silently lose a ticket. The damage is in the response: throwing at the join barrier discards the settlement that Runtime already had enough information to write.
Two things would make this recoverable:
- Settle first, then report the violation. Write
result.json with the tree, the spend that is known, and the leaked ticket as a spend gap or an explicit integrity finding, instead of replacing settlement with an exception.
- Attribute the leak. The message gives a ticket id (11) but not the assignment, child id, or stage that holds it. The seven
teardown-unconfirmed events suggest teardown of a child that never confirmed is where reservations escape; naming the holder would make that testable.
Related: several of the same run's children had earlier settled down with retained provider execution requires reconciliation before replacement, so this pursuit was already exercising the unconfirmed-teardown path repeatedly before the barrier tripped.
🤖 Generated with Claude Code
A live research pursuit died with no terminal result because one budget reservation was still open when the supervisor reached its join barrier:
Measured on 2026-09-10, Runtime 0.208.1, cli-bridge driver, opencode director on the Z.AI route, thirteen Tangle Sandbox children on
opencodethrough the Tangle router.What the run directory holds afterwards:
failure.jsonwith that message,nodes: 0, and nospentTotal.result.json.teardown-unconfirmedchild events, the last two beingextract:robert-ghrist:arXiv-2507.01226andextract:robert-ghrist:arXiv-2010.11525.So the pursuit's whole tree is unrepresented in any terminal record: the projection reports zero nodes while the journal shows thirteen children, several of which had already completed real inference. From an operator's seat the run simply vanishes, and the only way to learn what happened is to read the raw journal.
The invariant itself is right — a conserved pool should not silently lose a ticket. The damage is in the response: throwing at the join barrier discards the settlement that Runtime already had enough information to write.
Two things would make this recoverable:
result.jsonwith the tree, the spend that is known, and the leaked ticket as a spend gap or an explicit integrity finding, instead of replacing settlement with an exception.teardown-unconfirmedevents suggest teardown of a child that never confirmed is where reservations escape; naming the holder would make that testable.Related: several of the same run's children had earlier settled
downwithretained provider execution requires reconciliation before replacement, so this pursuit was already exercising the unconfirmed-teardown path repeatedly before the barrier tripped.🤖 Generated with Claude Code