Repository navigation

Orchestration and evaluation failure semantics (finding 4) - #15

Merged

arst merged 21 commits into

mainfrom

fix/failure-semantics

Aug 26, 2026

Commits on Aug 26, 2026
- fix(orchestrator): carry worker failures into synthesis, abstain when all fail
  
  arst
  and
  claude
  committed
  aabbf13
- docs(orchestrator): mark the tautological quorum and stop overclaiming the gap
  
  arst
  and
  claude
  committed
  7a77973
- fix(retry): rethrow cancellation and scope whole-turn retry to idempotent turns
  
  arst
  and
  claude
  committed
  41fa1d2
- fix(tool-auth): capability commit point, unconditional amount validation, real approval
  
  arst
  and
  claude
  committed
  2e0a099
- test(tool-auth): gate the race and the two lifecycle orderings, name the money-tool ceiling
  
  arst
  and
  claude
  committed
  438ecdd
- fix(react): enforce the tool-call bound in the host, not the prompt
  
  arst
  and
  claude
  committed
  3ba7171
- fix(react): catch the tool-call budget by type, not by message text
  
  arst
  and
  claude
  committed
  c14f112
- fix(judge): strict verdicts, indeterminate on malformed output, multiple orderings
  
  arst
  and
  claude
  committed
  4553fb9
- docs(judge): the probe is no longer a single swap
  
  arst
  and
  claude
  committed
  35ec9a5
- fix(regression-evals): trace cases need human promotion; rename exact to contains
  
  arst
  and
  claude
  committed
  6396e22
- fix(regression-evals): report awaiting-signoff and awaiting-review counts separately
  
  arst
  and
  claude
  committed
  fc86e99
- fix(regression-evals): derive the candidate count instead of hard-coding it
  
  arst
  and
  claude
  committed
  3880016
- fix(resource-aware): account for the fallback call and call the budget soft
  
  arst
  and
  claude
  committed
  2ccc76c
- ci: update and pin actions by commit sha
  
  arst
  and
  claude
  committed
  00e8800
- fix(resource-aware): name the SK twin's budget soft too
  
  arst
  and
  claude
  committed
  8ed43e4
- fix(react): enforce the tool-call budget with Terminate, not an exception
  
  arst
  and
  claude
  committed
  c3c8f95
- fix(judge): an unreadable rubric verdict is Indeterminate, not a zero
  
  arst
  and
  claude
  committed
  f9b0000
- fix(judge): measure position bias per slot instead of folding the slot away
  
  arst
  and
  claude
  committed
  79aa025
- fix(twins,docs): narrow the SK fallback catch; draw the abstain/partial branches
  
  arst
  and
  claude
  committed
  c97edd9
- docs: stop the new mermaid branch from relabelling the worker registry
  
  arst
  and
  claude
  committed
  c0bb6ef
- test(react): pin the auto-invocation loop to Terminate, not just the filter
  
  arst
  and
  claude
  committed
  550f4f7

Navigation Menu

Repository navigation

Orchestration and evaluation failure semantics (finding 4) - #15

Orchestration and evaluation failure semantics (finding 4)#15
arst merged 21 commits into
mainfrom
fix/failure-semantics

Commits on Aug 26, 2026

fix(orchestrator): carry worker failures into synthesis, abstain when all fail

docs(orchestrator): mark the tautological quorum and stop overclaiming the gap

fix(retry): rethrow cancellation and scope whole-turn retry to idempotent turns

fix(tool-auth): capability commit point, unconditional amount validation, real approval

test(tool-auth): gate the race and the two lifecycle orderings, name the money-tool ceiling

fix(react): enforce the tool-call bound in the host, not the prompt

fix(react): catch the tool-call budget by type, not by message text

fix(judge): strict verdicts, indeterminate on malformed output, multiple orderings

docs(judge): the probe is no longer a single swap

fix(regression-evals): trace cases need human promotion; rename exact to contains

fix(regression-evals): report awaiting-signoff and awaiting-review counts separately

fix(regression-evals): derive the candidate count instead of hard-coding it

fix(resource-aware): account for the fallback call and call the budget soft

ci: update and pin actions by commit sha

fix(resource-aware): name the SK twin's budget soft too

fix(react): enforce the tool-call budget with Terminate, not an exception

fix(judge): an unreadable rubric verdict is Indeterminate, not a zero

fix(judge): measure position bias per slot instead of folding the slot away

fix(twins,docs): narrow the SK fallback catch; draw the abstain/partial branches

docs: stop the new mermaid branch from relabelling the worker registry

test(react): pin the auto-invocation loop to Terminate, not just the filter