Paper Results, Supersession and Evidence Limits

This page summarizes the results in the Governed Agent Morphogenesis v1.0 preprint (opens in a new tab) and the limits on what they establish. Morphogenesis coordinates organizational changes across independently authoritative subsystems. Its morphology head records one accepted successor; it does not make effects across those subsystems globally atomic.

Bounded mission results

The mission study compared four conditions across four demand shapes and three fault schedules. It ran 32 paired seeds for each demand/fault cell: 1,536 runs across 48 design cells.

The two durable controllers matched on correct reports and reservation counts. Both blocked affected work while reconciliation was unavailable, then completed after recovery. Minimal adaptation retried with fresh operation identities, creating duplicate reservations; under shrinking demand, 64 runs exhausted the role budget and missed required work.

These results support durable coordination over the tested minimal adaptation. They do not show unique superiority over a competent durable workflow. Reservation benefits depended on demand: a false specialization forecast could cost more than keeping the organization fixed.

Resolve a losing proposal's effects

In the paper's competing-proposal trace, two approved proposals activate Teams and obtain Work effects before attempting to advance from the same morphology epoch. Only one wins the head compare-and-swap. The losing proposal's effects still exist and must be resolved with their owners:

1. Stop scheduling new effects for the losing proposal.
2. Reconcile prepared operations with each owner using the original stable operation identity. A lost acknowledgement or timeout does not show that the effect failed.
3. If an outcome remains unknown, block advancement instead of retrying under a new identity.
4. Fence new Work authority and account for already admitted jobs. Drain or complete those jobs before detaching or retiring resources.
5. Release the reservation when its owner confirms it is no longer needed.
6. Record a separate terminal superseded receipt. Do not issue a normal activation receipt for the losing proposal.

Cleanup can be partial; compensation cannot make arbitrary external effects disappear. One accepted head does not prove that competing proposals had no effects.

Evidence limits and assumptions

• The mission matrix is a local deterministic finite-grid study with a small role catalog, fixture authorities and prescribed faults. Its 1,536 runs are not independent general-safety trials.
• The authority matrix tested nine prescribed cases against four controllers. Its rejections do not estimate attack probabilities, and the proposals were not generated by an LLM.
• The TLA+ model covered two proposals, one predecessor, one stable effect identity and at most one admitted job per proposal. It explored 260 distinct states under stated fairness assumptions. It is not a refinement proof of the TypeScript runtime or an unrestricted proof.
• Conditional safety depends on each effect owner authenticating and checking scope, epoch, mandate, resource state and fences at use time; stable durable operation IDs; conditional storage writes; and an explicit trusted-clock/expiry contract. An unavailable owner or indeterminate result can block progress.
• The study did not evaluate LLM proposal quality, multi-tenant deployments, independent physical failure domains, long-running production traffic or typical production economics. It does not establish organizational improvement, global optimality or general production readiness.

For implementation and operational procedures, see Operations, Security and Evidence. The preprint's frozen datasets and verifiers are indexed in the v0.5 evidence artifact (opens in a new tab).