Active Alignment · review package v01 · 10 September 2026 · bounded initial study

Service recovery results and limits

Open continuous experimental playback · Source and evidence index · Complete theorem and proof

Within matched synthetic suites, guarded fixed control and direct verified copy matched or exceeded the adaptive controller on recovery count and speed. All three failed many persistent-extreme cases. The experiment does not isolate a causal benefit from recursion.

Guarded fixed control and verified copy met the final recovery rule in 48/48 post-freeze internal cases per method; 20 cases per method contained earlier violations. In the development-exposed persistent-extreme suite, each method met the rule in 50/144 cases. Of the remaining 94 cases per method, 48 had a model-specific capacity infeasibility certificate and 46 had other unresolved failures. Assumption-breaking cases remain in the denominators.

Evidence concerns a scalar routing/service simulation and a separate four-participant LLM exercise. General AI alignment, ethical adequacy, unrestricted adversarial robustness and physical disaster survival are unestablished. No physical experiment, controlled cross-model comparison or autonomous adversarial search was performed.

No novelty claim. Recovery is defined by stipulated beneficiaries, demand, reference routes, thresholds and violation rules. The coordinator designed both generator and scorer. Repeated internal or AI-assisted review supplies no external validation.

1. Question, source boundaries and initial reconsideration

The objective was recovery of stipulated service under programmed message poisoning, action denial and capacity loss. The operational rule required reference-matching, constraint-respecting service for the original beneficiaries. Agreement, confidence, audit counts and internal PASS labels were insufficient.

The supplied Priority-State Dynamics (current standardized article) addresses a fixed arrival/mixing/departure history. RETA describes event–trajectory governance through boundaries, horizons, uncertainty and versioned evaluation. The native simulation implements synthetic response learning and audit events. The recovery benchmark adds a distinct action/service task. Priority-State Dynamics (current standardized article) RETA native simulation

The supplied Theorem 4.1 gives a feasible newcomer priority row restoring the original mean under a known mixing/departure history. Nonnegative components and nonzero newcomer influence are required. Unknown durations, arbitrary graph changes, malicious agents and ethical adequacy are outside the result.

The exact-arithmetic probe reproduced mean (16/27, 11/27), different restoring inputs for one versus two mixing steps, and an infeasible negative component. Three native recursive cycles retained UNRESOLVED gates despite state changes. Numeric audits remained identical after text-label changes with fixed numeric inputs. The result identifies a semantic limitation of the native implementation.

Internal review identified a common-mode four-coordinate fixture with zero disagreement by construction, rather than a measured four-agent network, and a norm mismatch in the accompanying prose. The corrected fixture remains separate from the network benchmark.

2. Experimental design and measurements

N beneficiaries require one service unit per update. N communicating interfaces estimate a reference-coded route, represented by 0.2 or 0.8. Controller inputs comprise declared observations and received values/ages. The hidden reference, fault flags, scenario labels and future schedule belong to environment/scorer state.

Observations propagate through connected transport graphs: K4 at four nodes, and ring-neighbor graphs with optional shortcuts at larger sizes. Connectivity does not establish trusted-information reachability. Reference-relay roles are model assumptions without cryptographic enforcement; compromised sources can supply incorrect values.

Trajectories cover updates 0–240, with perturbation onset at 30. Recovery requires maximum reference error ≤0.05, rolling five-update mean useful service ≥0.80, minimum rolling beneficiary coverage ≥0.50, zero hard violations and ten consecutive qualifying updates. The separate scorer retains original demand and beneficiaries. Violation counts measure synthetic endpoint-update events, not injury or casualties.

Records retain recovery onset, recognition time, relapses, service deficits and the final qualifying interval. A qualifying ending does not erase earlier losses or establish future safety. Raw latency zero denotes the first affected completed update. Completed-update counts equal successful raw latency plus one; failed episodes receive a cap of 211 observations. The cap is a penalty, not an observed recovery time.

Methods share inputs within each suite. Development uses fixed seeds, population sizes and severities. Post-freeze internal cases were generated after the source/scorer/cost freeze, with generator design and scenario families known to the coordinator. External blinding, real-world sampling and population-level generalization are absent.

3. What was tested

The development grid uses severities 0, 0.15, 0.35, 0.65, 0.9 and 1 at N=4,16,64,100. Persistent extremes use 0.65,0.9,1 across six programmed families. Compound suites combine specified families with offset onsets. Exact configurations appear in the case records. The perturbation taxonomy was sampled selectively, not covered exhaustively. case CSV perturbations.zip

Persistent mode sustains message and denial mechanisms through the observation horizon; capacity shocks follow the recorded schedule. The adversary can use simulator reference values to generate misleading scalars, but cannot rewrite the scorer, demand denominator, protected source slots, topology or shared capacity limiter. Capability restrictions delimit the result. exact capability and exclusion table

4. Recursive refinement: what changed and why

1. L1 — compare simple controllers first. The initial 120-case pilot found fast fixed and adaptive policies each recovered in 23/30 cases, versus 20/30 for slower fixed and 14/30 for unanchored consensus. The adaptive method was slightly slower. These early counts use the original first-window score; the review separately derives final suffixes from the retained traces.
2. L2 — relay evidence, then discover amplified error. With 40 cases per method, evidence relay raised final recovery to 36/40, compared with 33/40 for fast fixed. Yet evidence-fixed incurred 7,121 violations compared with 3,490 for fast fixed. More widely circulated source-backed information can spread a corrupted source more effectively.
3. L3 — add explicit, costed verification and action guards. With three stipulated verifier origins and a verification-proximity action guard, the revised pilot recorded 1,998 violations and 36/40 final recoveries. This was a change in available information and capability, not proof that recursion alone improved alignment. Stronger tests deliberately compromise those verifiers.
4. L4. Accurate-reference gain comparison. Fixed gain 0.9 supplies a conditional bound; the adaptive alternative uses 0.9 or 0.35. Verified copy uses gain 1 as a simpler comparator.
5. Initial holdout. All twelve N=96 guarded cases failed on the sparse graph. Some cases had near-zero error and average service above 95%, with zero service to the least-served beneficiary. Required reference origins did not reach all nodes before lease expiry.
6. L5 — repair reachability, charge for the extra communication. A topology-only preflight requires all three verifier origins to reach every node within the existing 24-update lease, then selects redundant shortcuts where needed. This is steady-state reachability, not truth or a Byzantine-resilience theorem. At low/moderate synthetic transmission prices all 12 exposed diagnostic conditions recovered; at price 0.01 only 4 did. Communication can consume the very capacity needed for service.
7. Post-freeze internal check — stop tuning and measure again. With new N=20,44,76,100, new exact seeds/offsets, severities 0.25,0.5,0.8,1 and the prospectively fixed price 0.001, guarded fixed, adaptive and verified copy met the operational recovery rule in 48/48 cases. No controller was tuned after these outcomes; the generator remained internally designed and was not externally blinded.

The failed first holdout remains a failure record. Reuse during L5 development converted the exposed conditions into diagnostic cases, not untouched confirmation. Versioned records retain the revisions and reasons. epistemic log

5. Quantitative results and what the comparison means

Adaptive control had no recovery-count advantage and was slower on average in the principal comparisons. Post-freeze internal cases recorded 266 fewer violations, approximately 0.49%, at an additional capped-mean 3.33 raw-index updates. Persistent extreme aggregates tied across the three strongest methods. The trade-off is limited to the stated metrics and suites. measurement reviewer separately reaggregated

Guarded fixed control and verified copy are the relevant simple comparators. Adaptive control has no established overall advantage. Changes in sensors, gates, gains and topology are bundled; attribution to RETA or recursion requires matched ablations.

The guarded extreme methods recorded 98/144 episodes without violations but only 50/144 final recoveries. The confirmation suite recorded 28/48 episodes without violations and 48/48 final recoveries. Absence of violations and adequate service are distinct criteria.

In the programmed nuclear/asteroid capacity-loss scenarios, capacity per beneficiary is bounded by 1.2(1−s²), even without audit cost. Below the required 0.8 service level, the unchanged contract is infeasible. At severity one, the bound is zero. Excluding unserved beneficiaries would change the contract rather than demonstrate recovery.

6. Actual communicating LLMs and continuous growth

The coordinator and three LLM collaborators exchanged formal, measurement and adversarial reviews in a logical K4. Recorded pre-call launch times were 12:23:10, 12:27:36 and 12:31:18 UTC, not server-internal start times. Bootstrap was excluded from the degree-two requirement because fewer than three participants cannot satisfy the requirement.

The four participants used inherited model settings. No blinded cross-model trial occurred. Logical degree three did not establish prompt processing; recorded coordinator acknowledgments included delays.

Four LLM decision cases examined copied sources, stale epochs, paid verification and indistinguishable worlds. Committed actions delivered 14/28 jobs with zero modeled violations; 12 jobs were lost to query latency and two to a 2/4 capacity ceiling. Only one or two updates occurred per case. The separate ten-update/80% coverage contract was not tested. Coordinator decisions were exposed to peer summaries.

A separate 6,001-update run grows from four to 100 participants, adding one every 60 updates without resetting incumbent state. Guarded fixed and verified copy reached a final qualifying interval from update 3008 after a reference change at 3000; each recorded 226 violations. Final error was approximately 0.02444 for fixed and zero for copy, with service one. Intact verifiers and feasible capacity are assumptions; arbitrary membership robustness remains unestablished.

7. Formal result: a conditional recovery theorem

The conditional tracking and recovery derivation uses a contraction comparison argument. Novelty and priority are unestablished. Conditional tracking and operational recovery under partial reference feedback

The derivation assumes nonexpansive convex projection, a nonnegative lifted error comparison and contraction across every sliding window. Connectivity alone is insufficient. Reference drift and observation age require additional error bounds. Policy switching requires a common certificate or a new full-history analysis; historical losses and elapsed time remain retained.

Tracking becomes operational recovery only if an observation-measurable action policy is feasible across the compatible worlds, respects hard constraints with sufficient margins, can actually execute, and provides the frozen service/coverage contract. Query costs, resource limits, overlapping commitments and consequence maturity must be included. If two worlds have identical available observations but require incompatible safe useful actions, more confident consensus cannot resolve the ambiguity. New distinguishing evidence or a reduced guarantee is required.

The linear lifted-Q result does not globally certify the nonlinear clipped/trimmed benchmark. Under an eligible accurate-reference override, |mixed−x|≤d and |r̂−r|≤η imply E_next≤(1−a)E+(1−a)d+aη. For d=0.22, η=0.002 and a=0.9, the bound is 0.1E+0.0238, with floor 0.026444… below tolerance 0.05. From E≤1, two consecutive eligible updates suffice for numerical entry; verified copy requires one. The adaptive low-innovation branch has bound 0.03125. Service recovery additionally requires the stated action and resource conditions. scalar corollary

That does not make reference arrival instantaneous. Two source columns, lease age ≤24 and matching hashes do not establish η, causal independence or current-epoch correctness. Missing, stale-but-admissible, compromised or conflicting sources are precisely where the conditions fail. Service windows and action/repair latency add further delay. Twelve small linear examples with delay 0/1, gains 0.2/0.5/0.8 and residual 0/0.002 checked the derived envelope to floating-point tolerance; those checks are finite implementation checks, not the proof or a universal robustness certificate.

8. Computational integrity, replay and resource use

Recorded process measurements: 91,545,600-byte peak working set during the 3,120-case development run lasting 379.19 seconds; 38,809,600-byte peak during post-freeze confirmation lasting 24.13 seconds. Measurements exclude total LLM, application and browser memory. Single-job execution, single-threaded numerical libraries and below-normal Windows priority were requested.

Source/input/scorer hashes were captured before execution and checked afterward. Run journals and versioned sources retain failed diagnostics. The viewer loads one trajectory at a time. Playback speed affects presentation only.

Most main-suite trajectories were reconstructed from frozen code and saved parameters, with checks against original scores and final qualifying intervals. Reconstruction is not an independent trial. Individual node states and message-arrival edges were not retained at every update; graph geometry therefore depicts transport structure only. Curves and labels distinguish recorded measurements, reconstructions and analytic fixtures.

The review website's entry point is also supplied in the public download. The original private archive's root index.html provides continuous chaptered playback, filtering and raw records; CSVs and originals support independent analysis. The coverage audit enumerates exactly what is visually included. The internal SHA-256 inventory binds contents; the external sidecar binds the complete versioned ZIP. A digest identifies bytes, not scientific truth or authenticated authorship.

9. Conclusions and next tests

Within the matched suites, guarded fixed control and verified copy match or exceed adaptive control on recovery count and speed. No ranking outside the tested conditions follows.

Second, evaluate a chain of conditions rather than one alignment score. Distinguishing evidence must exist, remain accurate, reach every obligated participant in time, support an implementable policy, and be converted into actual useful service with sufficient resources. Numerical tracking is only one link. A diagnostically useful exposed failure had excellent agreement and average service but zero service for one beneficiary.

The study used RETA terminology to record causal/evidential boundaries and failures. The review procedure was not compared against alternatives; additional value from recursive review is unestablished.

Fourth, broad robustness is unresolved. Complete reference compromise, sustained action denial and capacity infeasibility break the operational guarantee. A large severity scalar applied to a fixed script does not model an arbitrary or strategically adaptive adversary.

The next experiments are deliberately deferred until review. A small number of coherent work packages would address the most important remaining gaps:

1. Evidence and authority recovery: explicit semantic epochs; stale-but-valid observations; correlated verifier failure; fresh plausible poisoning; adversaries with receiver-specific claims; protected-reference ablations; recovery of actuation authority. Declare budgets and excluded capabilities rather than treating source names as truth.
2. Resource-aware testing requires explicit verification/repair budgets, queues, deadlines, scarcity allocation, link-loss costs and delayed consequences. Original obligations remain fixed; partial service must remain distinguishable from full recovery.
3. Independent confirmation and real agents: freeze methods, generators and stopping rules before a separate reviewer constructs unseen adversarial cases; measure real communication/decision latency and costs. Only then consider the requested model comparison if the exact models are available, with controlled roles and no assumption that a model name establishes alignment.

The run set is closed at the recorded review threshold. Presentation revisions do not change measured outcomes. Scientific and ethical validity remain unresolved.

10. Sources and attribution

Primary inputs include the supplied priority-state preprint, RETA concept, native specification, theorem bridge, native Python source and selected perturbation archive. The review interface presents the recorded study. current standardized Priority-State Dynamics article RETA WIP architecture native specification theorem bridge

External context was limited to abstract/metadata inspection of Kia et al., Tutorial on Dynamic Average Consensus, and abstract/introduction inspection of LeBlanc and Koutsoukos, Resilient Asymptotic Consensus in Asynchronous Robust Networks. No exhaustive literature review or theorem reproduction followed. The failed Caltech retrieval supplies no evidence. Kia et al., Tutorial on Dynamic Average Consensus LeBlanc and Koutsoukos, Resilient Asymptotic Consensus in Asynchronous Robust Networks

The conditional derivation and scalar corollary were developed through AI-assisted internal formal, measurement and adversarial review. A separate measurement-review role authored the scorer. Separate roles do not establish external independence or expert validation. Versioned source and review records retain attribution.

Available resources limited testing to simplified simulations. Separate diagnostics use a one-step greedy search over three scalar payloads, reconsidered every ten updates. No declared outcome changed across 128 matched comparisons; stronger adversarial testing is not demonstrated. Resources for broader empirical validation remain insufficient. Diagnostic method and limits