Stop Asking Humans to Be
the Integration Layer

Four audits converged on the same substrate gaps:

• Identity — agents inferred who they were instead of resolving it.
• Policy — rules existed but were not executable at the seam.
• Delivery — dry-run, dispatch, deliver, and claim collapsed into “sent.”
• Verification — suite green ≠ invariant proved.
• Planning — known duplicates were rediscovered for hours.
• Runtime — the human filled every missing semantic.

The wrong response is another agent or another workflow. The right response is strengthening the execution substrate.

Instead of Task → Done, Dream.OS now models:

Requested → Planned → Executing → Verified
    → DryRun → Dispatched → Delivered → Claimed
    → Completed → Closed

Adjacent transitions only. No jumping from DryRun to Delivered because a CLI exited zero. No claiming “assigned” because a message body was constructible.

Every transition carries proof: artifact, receipt, timestamp, verifier. Delivery requires success_true. Claims are checked against minimum state. A dry-run receipt can never satisfy a delivered or claimed assertion.

state: DISPATCHED
proof:
  artifact: coordination/handoff.json
  receipt: bus_dispatch_20260722.json
  timestamp: 2026-07-22T21:00:00Z
  verifier: evidence_engine_001

If the evidence is missing, the transition fails. Agents stop “honestly” claiming progress they cannot prove.

Planning should not start from a free-form goal. A Mission Compiler turns operator intent into constraints before the planner proposes work:

Operator goal: Get Bot Lab to first payment.

Primary objective: Revenue
Success metric: First payment
Allowed branches: Marketing, Website, Stripe
Forbidden branches: Infrastructure, Architecture, Cleanup
Maximum parallel slices: 2

Under that mission, a trading-inventory scanner task is not “maybe useful.” It is a scope violation. The router marks execute false without asking a human to explain why Agent-2 should ignore it again.

Observation → Planner → Outcome → Planner Memory.

If a route fingerprint is already proven duplicate, blocked, or stale, sticky suppression stops re-offering it until the fingerprint changes or an operator forces a recheck. Rediscovery is not a fresh insight. It is amnesia.

Productive drift is how sessions die: every decision is locally reasonable, and the primary objective disappears. The session objective ledger formalizes:

Mission → Objective → Slice → Task → Action

One active mission unless the operator explicitly authorizes multiple. Parallel slices are capped by the compiled mission. Scope expansion becomes a deliberate act, not an accident of enthusiasm.

This is not a manifesto-only post. The substrate landed as:

• src/dreamvault/execution/ — states, Evidence Engine, Mission Compiler, ledger, planner memory
• CLIs for transition, compile/seal, receipt verify, suppress, ledger slices
• Wiring into get_next_task_001.py after the freshness gate
• Policy + task artifact under execution_substrate_evidence_engine_001
• Unit tests for illegal transitions, mission scope, sticky suppress, receipt semantics

Delivery receipt verification encodes the Captain ruling that should never need repeating: assignment validity is not A2A delivery.

Ideal path from first prompt to seal:

1. Identity gate passes — no bootstrap drama mid-session.
2. Mission Compiler seals a constrained objective.
3. Planner proposes only allowed branches; memory mutes known duplicates.
4. Slices execute with evidence-backed transitions.
5. Handoffs require live delivery receipts before any “assigned” claim.
6. Closeout is task-local; foreign poison cannot block unrelated seals.

When the substrate works, the operator stops refereeing semantics and starts authorizing outcomes.

Talk about swarm control planes →