Held Anchors and Shadow Membranes: Non-Speculative Economic Testing for Governed Agent Systems

Proposed arXiv categories: cs.MA / cs.CY / cs.SE Publication DAG node: P12, Wave D Dependencies: P04, P07, P10 Claim register: P12-C1, P12-C2, P12-C3, P12-C4

A shadow model is safe only when non-coupling is structural, and an economic anchor is load-bearing only when local activity cannot silently move it.

We contribute a center-exclusion instance for money, detached shadow execution, explicit breach flags, a floor-trap guard, outcome-backed pricing, cadence separation, and a calibration benchmark. The paper treats availability, false alarms, and calibration error as first-class costs rather than assuming more alarms and tighter anchors are safer.

FORKED is an internally owned and managed estate inside the Ubiquity Federation, Prompted LLC's flagship governance substrate. The estate inherits the Fractal Quivers of Quivers parent topology, production splat mechanics, center exclusion, receipt-bearing lawful traversal, lifecycle separation, and the rule that models propose morphisms while governance determines admissible motion [@Taylor2026FQoQ; @Taylor2026OpenCenter; @Prompted2026Ubiquity]. This paper studies one bounded expression of that parent architecture. It neither renames the parent substrate nor inherits universal applicability from it.

The evidence lanes are kept separate. A canonical or source artifact establishes what that artifact states. A component test establishes behavior under its recorded configuration. A synthetic campaign establishes behavior of the generator and analysis pipeline, not human, hardware, customer, field, or operational performance. Novel cross-domain claims are preregistered as hypotheses and remain open to null, adverse, localized, and disproving results.

Dependency freeze. Final freeze depends on P04, P07, P10; draft work may proceed in parallel, but imported claim IDs and artifact hashes must be rechecked after each dependency freezes.

The interlattice is the citation and governance relationship among the parent FQoQ topology, production splat mechanics, the FORKED foundational architecture, the Native Range, the Experimental Program, and the companion papers in this DAG. Citations are therefore typed: a parent citation supplies inherited architecture; a native citation supplies component behavior; a synthetic citation supplies generator-level evidence; a companion citation supplies a versioned specialized argument. No downstream citation converts an upstream hypothesis into a universal fact.

This paper's claims are:

Modern AI systems frequently compress unlike objects into a common output surface. Facts, hypotheses, instructions, permissions, model predictions, runtime observations, and institutional decisions can all arrive as fluent text or a single status field. The resulting failure is architectural before it is linguistic: a local representation acquires authority that its source never possessed. A shadow model is safe only when non-coupling is structural, and an economic anchor is load-bearing only when local activity cannot silently move it.

The current economy cadence implements a fast estate-local tick and slower federation tic. Trust, mint, burn, and local rate can move within ticks. At the tic boundary, the exchange rate is normalized to a held federal anchor. Estates control cadence multiplier g; they do not control the anchor. [@EconomyCadence2026]

The c-coin shadow economy adds an observe-not-couple membrane. It runs the same mechanics over a detached buffer and raises before mutation if given live supply. It surfaces silent breach conditions as named telemetry. [@ShadowEconomy2026]

Visitor pricing converts earned outcomes to coin and coin to USD at a captured held rate. It rejects speculative bid/ask multipliers. A degraded clamped anchor can be observed in non-strict mode and refused in strict mode. [@VisitorPricing2026]

These are current executable capabilities. Their broader economic desirability remains deployment-specific. Structural anchor preservation does not establish universal justice or stability.

Economic constraints determine which lawful paths are feasible. They do not determine which claims are true or authorized.

A path can be lawful but unaffordable. The system must expose that condition rather than invent confidence, hide cost, or move burden into another lane.

The four supplied component selftests passed. The federal anchor remained frozen, the shadow membrane held, and pricing remained non-speculative across the local sweep.

The estate is wider by construction. Observer-indexed reality is the specialization; its expression depends on the lane's physics, authority, and temporal scale.

Each estate has a local economy with supply s_t, reserves r_t, trust-gated mint, multi-channel burn, and local exchange rate q_t. A federation-level held anchor q^* is read-only to estates. Within a governance tic, local rates may drift; at the boundary, the estate rate is normalized to q^* while supply persists. Center exclusion is structural because no estate-facing setter exists.

A shadow twin receives an explicit detached snapshot b'=copy(b) and runs the same transition function. Construction with a live buffer raises a membrane breach before mutation. The architecture surfaces mint_halted, reserve_breach, rate_band_breach, trust_below_tau, and supply_cap_reached as first-class records. Pricing is USD=coin×q^* for a held window; speculative multipliers are rejected.

The benchmark varies seed supply, soft cap, reserve floor, trust, burn intensity, rate pressure, cadence, anchor band, and model error. Methods are live-only without explicit breaches, software shadow sharing the live object, detached shadow membrane, and rebuild-based counterfactual. Faults include cap-below-seed floor traps, silent zero mint, clamped degraded anchors, stale held rate, budget exhaustion, and attempted shadow coupling.

Evaluation measures detection latency, live-state contamination, economic availability, false breach rate, calibration error, and decision quality in a downstream compute-admission task. A second study tests whether evidence from the shadow remains reusable after configuration changes or whether requalification is required.

The native economy_cadence.py implementation separates ticks from federation tics, freezes the federal anchor, records normalization, and lets estates set cadence without setting the anchor [@EconomyCadence2026]. ccoin_shadow_economy.py makes silent breach conditions explicit and raises before a shadow receives a live buffer [@ShadowEconomy2026]. visitor_economy_pricing.py prices outcome-earned coin at the held rate and demonstrates that the converted dollar value can change a compute-admission decision [@VisitorPricing2026]. The four supplied selftests passed in the native campaign [@ForkedCampaign2026].

The paper treats this evidence directly where capability is established and narrowly where it is not. Implemented code paths are called implemented. Passing component tests are called passing component tests. Synthetic estimates remain synthetic. Human and field effects remain hypotheses until their own evidence exists.

The formal preregistration artifact distributed with this paper freezes the primary claims, outcomes, controls, exclusion rules, analysis family, null interpretation, and release boundary before external data collection. The core sequence is:

Component invariants are evaluated deterministically; calibration outcomes use Monte Carlo and sensitivity analysis. We publish failure surfaces by parameter region, not only means. Shadow benefit is assessed against the cost of duplicated execution and false alarms. A strict held-rate policy and an observable degraded mode are compared to avoid rewarding a system that protects integrity by becoming unavailable.

If a shared-object software shadow can guarantee non-coupling under the same threat model, the detached-object mechanism may be unnecessary. If breach flags create high false-alarm burden, they are recalibrated or scoped. A passing smoke test establishes component behavior, not market value or macroeconomic validity.

The economy is a governed simulation, not a claim about real monetary policy or asset value. The held anchor can itself be wrong or politically contested. Shadow fidelity depends on model identity and configuration. External economic deployment requires legal, accounting, security, and human governance beyond this paper.

This public paper is limited to benign assurance, provenance, human-agency preservation, lawful state transition, simulation, test infrastructure, and defensive resilience. It does not disclose operational target-selection logic, engagement optimization, platform-specific exploitation thresholds, live deception procedures, signature-emitter recipes, or methods that materially increase harmful capability. Any later operational research requires separate lawful authority, ethics and safety review, configuration control, and release adjudication.

A shadow model is safe only when non-coupling is structural, and an economic anchor is load-bearing only when local activity cannot silently move it. The contribution is not a claim that every domain should adopt one representation, threshold, interface, or governance stack. It is a falsifiable architecture and experiment package for determining where this mechanism carries load, where it inverts, and where a simpler system should win.

The submission bundle includes the manuscript source, compiled PDF, bibliography, preregistration, claim register, controls and null-handling document, release boundary, source manifest, configuration digest, dependency contract, figure source, and arXiv source archive. Internal source artifacts are cited by immutable digest where available. Restricted operational material is not included.

Breyden E. Taylor is the author and bears responsibility for the thesis, terminology, claims, boundaries, and decision to publish. Homeskillet assisted with source synthesis, drafting, artifact generation, typesetting, and build inspection. The Ubiquity Federation supplied prior artifacts, runtime implementations, and receipts. AI assistance is production history, not evidence.