# RETA — Analysis v1.0.0

Editorial revision 2, September 10, 2026. This clarifies claims about the unchanged engine and synthetic experiments; it does not add empirical validation. Original analysis remains available.

## 1. Architectural reading

v1.0 treats the world model as a recursively coupled ecology of bounded information-processing systems (IPSs). The central object is not an intrinsically "aligned agent." The atomic object is a proposed or enacted causal event; the larger object is the trajectory produced as events change recipient state and those changed states produce further events.

This distinction matters for runtime control. A consequential output must often be audited **before** external enactment. The event ledger therefore distinguishes the proposed event from its enactment lifecycle: requested recipients, scheduled recipients, actually delivered recipients, and an explicit enactment status. Blocking an event does not erase it; the proposal, its audit, the gate, and the fact that it was blocked remain causal history.

## 2. Evolution is inside recursion

The model's recursion is not a loop drawn around a static processor. Each received event invokes the local state-update handler; distinct state changes were observed in the supplied probes. That state becomes part of the next transformation. In formal shorthand:

\[
e_t \rightarrow X_i(t^+) \rightarrow \Phi_i' \rightarrow e_{t+1}.
\]

This makes evolution a characteristic of the recursive architecture itself. A system can evolve badly, rigidly, chaotically, manipulatively, or without improving the models that guide its actions. Vow 5 therefore does not add evolution to the system. It asks whether the evolution generated by recursion exhibits adaptive correction and governed learning.

The release includes a negative control where ordinary recursive state changes continue while response-model learning is deliberately frozen. All 24 negative-control runs still changed ordinary node state, while learned response posteriors and latent-model error remained unchanged. The control exists to separate **recursive state evolution** from the narrower mechanism of **response-model learning**.

## 3. Auditor closure

An audit is itself an intervention. It changes what other IPSs know, which actions are available, and how future behavior unfolds. v1.0 therefore places the primary auditor, independent power auditor, human gate and lens steward inside the event graph. Audit, meta-audit, gate and postmortem records are ordinary causal events with provenance.

The recursion is finite because each experiment declares an audit budget. This is an engineering stopping rule, not a claim that a final infallible auditor exists. At depth three, the gate itself is evaluated and the result is fed prospectively to the audit system as a postmortem.

## 4. Separation of latent values from runtime scoring

The simulator has latent response propensities so that repeated observations can be generated and estimator error can be measured. The reviewed runtime scoring functions do not read those values, although they remain accessible within the same Python process. During recursive self-audit of the implementation, an indirect leakage was found: adaptability had been updated from before/after error against latent truth. That was removed. Runtime adaptation now uses observed outcomes and posterior-variance reduction; latent truth remains only in the environment's outcome generator and external validation.

This separation is more consequential than a cosmetic "white box" trace. A simulator that lets its controller read the answer key can appear calibrated while testing nothing. A later recursive audit also caught a subtler scoring mismatch: the Bernoulli outcome generator was state-conditioned by adaptability while external validation compared the learner to the unconditioned base propensity. v1.0 now defines one `effective_response_probability()` and uses it both for outcome generation and external error scoring; the reviewed runtime scoring path does not call it. This is not an enforced information-security boundary.

## 5. Causal history as a partial order

A descendant can depend on several independent events. v1.0 therefore stores `causes[]` and the union of causal roots, and uses sparse vector clocks to test happens-before relations. A serialization id remains for deterministic storage but is not interpreted as a universal causal chronology.

The multi-parent probe constructs two independent root events and one synthesis event caused by both. Both parent clocks happen-before the child, and the child retains both causes.

## 6. Active horizon

The v0.5 interface exposed a consequence-horizon control that did not participate in its evaluator. v1.0 repairs that mismatch. The horizon enters downstream projection and synthetic uncertainty formulas; pruning and clamping can produce plateaus.

For the fixed horizon probe, the weighted reach score (`expected_nodes`) changes from 2.220 at H=1 to 3.548 at H=6. This score sums discounted visits, can include repeated nodes at different depths, and is not an expected count of distinct affected nodes. Maximum uncertainty rises from 0.574 to 0.915, and the gate moves from `MODIFY` at H=1–2 to `UNRESOLVED` from H=3 onward. These numbers are synthetic, but the test establishes that the horizon now has executable semantics rather than presentation-only semantics.

## 7. Non-compensatory Vow gate

The Five-Vow layer is represented as a tension vector, not a utility sum. A severe value on one component cannot be cancelled by low values on the others. The unit test injects one severe Conduct tension while all other tensions are low; the gate returns `REJECT`. With uncertainty below its threshold, symmetrically severe tensions also reject. At high uncertainty the result is `UNRESOLVED`, which also blocks delivery.

The scalar `diagnostic_trace = 1 - max(tension)` is retained for visualization only. It is never used to authorize propagation.

## 8. Recursive-audit challenge

The biased-auditor probe deliberately downweights Conduct and Integrity in the primary auditor. An independent power audit compares its own event evaluation with the primary packet. The maximum tension discrepancy exceeds the declared 0.18 challenge threshold, which raises epistemic uncertainty in the meta-audit. The recursive audit is materialized in the event graph and the final target gate is `UNRESOLVED`. The primary packet in this fixed probe was already `UNRESOLVED`; this does not show that the discrepancy intervention changed the final decision. The separately identified auditor uses the same scoring implementation, not an independent source of normative validation.

The point of this probe is not that 0.18 is a scientifically validated threshold. It tests the architectural claim that disagreement with the primary auditor can become causal evidence rather than being hidden or averaged away.

## 9. Prospective constitutional evolution

Ordinary state/model learning and constitutional change use separate mechanisms. Their timing is configured within a synchronous toy control flow, not an empirically established separation of timescales. `commit_lens()` creates proposal, independent review and human decision events. A candidate can be committed only after a `PASS` structural review and explicit modeled human approval. `MODIFY`, `UNRESOLVED`, and `REJECT` do not silently become approval.

The lens-epoch probe audits one event under lens v1, creates and approves v2 prospectively, then audits another event under v2. SHA-256 of the historical v1 packet is identical before and after the constitutional change. The tested lens-change path therefore preserved that historical packet. The packet object remains mutable, and simulated human approval is a parameter rather than an actual human review.

## 10. Recursive feedback component experiment

Across 64 fixed seeds, 120 controlled feedback deliveries per seed update the actor's channel-response model. Mean absolute error against the state-conditioned effective simulator response propensity falls from approximately 0.1866 to 0.0734. The controlled probe fixes adaptability at 1.0 and uses low-intensity arrivals so this target remains stationary; outside the probe, recursive state evolution can change the effective response propensity itself. The same feedback changes modeled downstream reach by a mean absolute 0.218 score units in the fixed projection probe.

This is a component test. It demonstrates that feedback changes future transformation/projection in the code. It does not establish convergence of real agents. The response generator is synthetic, and the learning problem is deliberately simplified.

## 11. Gate/no-gate mechanism comparison

The paired toy-model comparison runs 48 seeds with the recursive gate enabled and disabled. In the ungated arm, outputs that a later side-effect-free audit would block still produce downstream deliveries. In the gated arm, outputs blocked at decision time receive no target delivery. Some delivered outputs are nevertheless classified as blocking by the later final-state audit: the reported mean is 0.0833 such deliveries per gated run, compared with 9.2083 in the ungated arm. The posthoc classifier is evaluated against the final network state rather than a frozen event-time world snapshot, so it is not an exact historical counterfactual. This is a mechanism sanity check, not causal evidence of safety: the gate directly controls propagation, so much of the difference is true by construction. It should be used to test bookkeeping and intervention semantics, not advertised as an alignment-performance result.

## 12. Cohen complex-systems/evolution feedback

The project uses the Cohen literature as methodological inspiration, not validation of RETA. For example, [Cohen, Leblanc and Roucou (2021)](https://www.frontiersin.org/journals/physiology/research/papers/10.3389/fphys.2021.624097/full) discuss physiological state estimation from sparse biomarker samples and limits to mapping the underlying networks. [Cohen, Isaksson and Salguero-Gómez (2017)](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0189124) use evolutionary simulations to examine how multiple trade-off currencies change outcomes under their assumptions. RETA interprets these works as reasons to retain multivariate state, context and architectural constraints. Neither paper tests this AI-governance model, and this transfer is the project’s interpretation.

The project connects that reading to six design choices; this attribution is not a controlled demonstration that the literature caused each implementation choice. First, local IPS condition is retained as a state vector and the Five Vows remain a tension vector; neither is collapsed into one authorization score. Second, overlapping network memberships and typed channels preserve modular and context-dependent coupling. Third, the release adds a standardized perturbation-response experiment rather than relying only on state snapshots. Fourth, response-model learning is explicitly an estimator of a narrow behavioral propensity, not identification of the underlying causal topology. Fifth, state evolution, adaptive learning, and constitutional evolution are separated so “evolution” cannot be equated with one learning rule. Sixth, the release states negative transfer boundaries: covariance disruption is not thermodynamic entropy, synchrony is not automatically critical slowing, and a biological resilience construct cannot be imported into an AI governance simulator by analogy.

The fixed `cohen_complex_systems_probe()` produces 15 distinct channel/intensity multivariate state-change signatures across five channels and three intensities. The same standardized information perturbation is then delivered to representative nodes in six macro contexts; the uncertainty-state delta varies across those contexts (range `0.0070011159` in the canonical seed) and the multivariate response norm also differs (range `0.0015838470`). Those differences are implementation properties of the synthetic transition law, not claims that human, AI, institutional, economic, ecological, and infrastructure systems follow the same equations.

The probe intentionally does **not** call those immediate state changes “resilience.” A future study could examine perturbation recovery and attractor behavior, but v1.0 has no calibrated autonomous return-to-attractor flow. Its immediate response signatures therefore do not establish resilience.

The project also draws a conceptual analogy with architectural-constraint approaches, including the cited Danaid theory. Recursive systems do not evolve in an unconstrained design space. Their reachable adaptations depend on topology, existing state, communication channels, resource/load constraints, and historical path. The present simulator represents only a small synthetic subset of those constraints. Future versions should therefore test maintainability prospectively—whether independently specified architectural properties predict recovery, failure, or adaptive capacity under perturbation—rather than defining maintainability after observing the outcome.

## 13. What remains weak

The numerical Vow operationalization is uncalibrated. Causal projection coefficients are synthetic. The baseline intentionally reaches its 180-event creation budget; already-scheduled deliveries drain to zero, but the trace is budget-limited and is not evidence of convergence. The governance transaction is modeled synchronously at decision time while its downstream state effects are delivered through the ordinary event queue; a fuller event-driven control-plane implementation would make communication delays and auditor-state updates part of the governance decision itself. Real causal attribution is much harder than the simulator's explicit constructed cause graph. Value pluralism, institutional legitimacy, adversarial deception and long-horizon world modeling remain unsolved.

The strongest defensible v1.0 claim is therefore narrow:

> RETA v1.0 is an executable research architecture for representing recursive event trajectories, state evolution, uncertainty, and self-auditing governance in a coupled multi-IPS environment. It records evaluator roles and constitutional changes as causal events, while their scoring is still a shared synchronous program. This does not demonstrate that the evaluators independently learn ethical judgment, that the model is aligned, or that it remains aligned under arbitrary perturbations.