arXiv (Oct 2 surface; MPI-SWS / MIT / Purdue) — LLM workflows often park human approval outside workflow state (UI click / audit log). The orchestrator then lacks a commit-time check that every risky path hit an approval gate. Guarded-commit makes approval a workflow transaction: an append-only ledger records a resolution + referenced evidence; a credential-confined ACT adapter (only the adapter holds the irreversible-action credential) checks that record against artifact hash, governance-policy version, ledger prefix, and executed-path witness before invoking the external API. Four-condition commit predicate: (1) barrier coverage with path obligations, (2) policy/knowledge pinning, (3) artifact identity (EDIT binds pre / Δ / post hashes), (4) safe reuse/automation (fresh derived record; transitive root to HUMAN or declared automation). Validator on 999 synthetic DAGs (0 false accept/reject); replay-hash equality on 252,386/252,386 artifact-present traces across 271,035 public traces. Builder takeaway: confine irreversible credentials to a commit adapter; bind approvals to content-addressed evidence + executed path + policy version; treat reuse as a new authorization event.

arXiv (Oct 1 / Oct 2 surface) — Multi-agent work on shared code/docs/tools often looks locally correct while the combined result is wrong — the global coherence problem (state, not intelligence). Two failure modes: (1) observation aliasing — Observation-Aliasing Impossibility Theorem: a policy can guarantee a valid action from its observation iff every fiber shares a common admissible action; best randomized success 1/k when k look-alike worlds have pairwise-disjoint admissible sets; more agents/messages/reasoning cannot invent missing distinctions. (2) pieces that do not glue — pairwise-valid spends break a shared budget. Design rule: models propose; harness owns state and commit. Nine studies: aliasing probe 40/40 with deciding event visible vs chance-compatible when hidden; TeamBench shared 20-call budget overspent 5/5 ordinary / 4/5 with live count / 0/5 with commit ledger or escrow; τ²-bench silent-revert stock 0.07 vs harness 1.00. Builder takeaway: name an owner for every cross-agent constraint; enforce at commit or split into escrow quotas; put deciding distinctions into context (identifiability, not verbosity).