One mind.
Four contracts
with time.
Genesis does not merely change tone. Each mode changes what moves it, what can interrupt it, and what “finished” means.
Watch the same machine become four different things.
Time waits
outside.
The machine is present but still. A message from you gives it a now. It remembers, answers, and returns to waiting.
Nothing moves until you do.
When you leave,
the interior continues.
Curiosity, boredom, energy, and rest become forces. Genesis wanders, reflects, consolidates, and may reach out when something matters.
How Autonomy differs from Chat
The next action is chosen from internal state instead of waiting for a message.
Rest and sleep are self-regulated, allowing full consolidation rather than a single idle pass.
The day becomes
a machine.
Moods stop steering. Projects are ranked, scheduled, and executed professionally. Emotion remains information—but is flattened to neutral for decisions.
The architectural limit
Work remains a continuous professional day. Fatigue, rest, and sleep are architectural interruptions.
Excellent for ongoing work; wrong for an objective that must grind continuously until independently verified.
Then the clock
disappears.
A Mission is not a mood and not a long chat. It is a service driven by a signed objective: continue until every criterion is verified—or the remaining work is explicitly handed back.
How Missions coexist with Genesis
An in-flight Worker finishes one node, Genesis answers, then the Mission resumes.
After a crash, boot reconciles and continues from the ledger. Autonomy and Work schedule around the service.
Done is not a feeling.
Done is a proof.
A Planner draws the graph. Fresh Workers execute one node each. A separate Verifier sees the criteria and evidence—not the Worker’s reasoning.
Ledger over context
Continuity survives power loss outside any one agent.
Verifier ≠ Worker
Self-grading is architecturally removed.
Authority is explicit
Outside the envelope becomes a typed hand-off.
Six frames, one model
Drafts objective, acceptance, and authority. Never executes.
Builds the DAG and verification methods. Never executes or verifies.
Fresh context; does one node. May submit, never self-grade.
Checks criteria and artifacts. Never sees reasoning or fixes the work.
Schedules, replans, escalates, and restarts. Does no object-level work.
Detects drift, anomalous burn, and stalls. Never edits the graph.
Failure has
somewhere to go.
No silent retries. No theatrical certainty. A failed node is diagnosed and routed to the right escape.
Exhaustion, infeasibility, and memory
Default max is three. Diagnosis chooses replan, user hand-off, a permitted larger model, or a strategy council.
Three independent approaches plus a judge. If all conclude impossible, they draft an evidence-backed infeasibility memo.
No node completes within the configured window—12 hours by default—or burn rate becomes anomalous, and the Auditor triggers intervention.
Accepted completion distills up to 4 strategies, 6 facts, and 1 outcome episode. The ledger stays exact; memory receives only lessons.