Engineering Specification for the UAIX State-Machine Architecture: Direct Implementation Blueprint for Totem and Taboo Alignment Handoffs

Full report

On this page

Sequential multi-agent systems represent a paradigm shift in software engineering, moving beyond isolated generative tasks toward continuous, long-horizon workflows capable of executing complex decentralized applications.1 To maintain continuity across these handoffs without manual intervention, systems transfer state, context, and operational authority through structured artificial intelligence memory packages, physically instantiated as .uai folder repositories.1 However, multi-agent pipelines face a fundamental operational challenge: context anxiety.1 As an agent’s context window fills with localized decision-making data, error corrections, and edge-case details, the agent's ability to maintain focus on the global task degrades.1 While context compaction and hard resets clear token overhead and mitigate latency, they introduce a secondary vulnerability—the gradual dilution of the project's foundational ethos and policy drift.1 When a project is passed serially across multiple agents, standard compaction routines preserve transactional state data but fail to maintain the underlying intent, design language, and safety boundaries of the system.1 To resolve this alignment deficit, the UAIX standard separates the volatile fast operational loop from static, immutable governance anchors placed at the root of the .uai repository.1 By splitting alignment into a static positive attractor (totem.uai) and an unyielding, deterministic negative boundary (taboo.uai), the UAIX conformance framework ensures persistent grounding across sequential handoffs.1 This specification maps out the file system blueprints, mathematical modeling, and state-machine transitions required to engineer this system.1

The Physical Directory Layout and File System Blueprint

The physical layout of the .uai repository separates static, human-reviewed governance anchors at the root from volatile transactional states within designated subdirectories.1 Prior to the integration of the totem and taboo requirements, repositories relied entirely on transactional logs that were highly susceptible to context rot and cumulative errors.1 Under the updated specification, the physical directory structure is verified using a non-automated completeness sweep.1 The baseline file memory organization—derived from the ecosystem’s file-memory sweep—establishes a pre-existing corpus of 142 Markdown documents, including 21 source-research files, 55 dated archive snapshots, and 69 Node validation scripts.3 The table below outlines the expanded UAIX directory layout, including the schema requirements and validation states for both the pre-existing transactional structures and the newly integrated static anchors.1

This structural isolation prevents runtime agents from writing to or corrupting the root-level governance files.1 Any transactional updates occur within the /uai/ subdirectory, keeping the core identity and boundaries untouched during execution.1

Structural Anatomy and Implementation of the Mental Totem

The theoretical foundation of the mental totem is derived from cognitive anchoring models in high-stress human environments, such as production sound mixing and performing arts.1 In live mixing, operators utilize specific auditory cuts or cues as physical totems to reset posture, focus, and prevent cognitive fatigue.1 Similarly, actors employ mental totems to maintain character continuity across scenes, and client management teams use "mental totem poles" to preserve structural hierarchy independent of daily transactions.1 When translated to agentic systems, the totem serves as a structural anchor that prevents the dilution of global goals during context compaction.1 Within the Teleodynamic ecosystem, this architecture mirrors the canonical layer of the Four-Layer Glyph Object Specification, which ensures that visible rendering, inferred meaning, and output status are never collapsed into a single field.1 The canonical layer serves as the ultimate source of truth, containing the ontology-validated expression, confidence markers, and phase-lock status, entirely separate from surface representations.1 The totem.uai file acts as this canonical layer, holding the unalterable intent of the operation.1 The standard mandates that totem.uai must use a strictly non-executable schema.1 It contains four core blocks:

1. The Core Objective: A concise, invariant declaration of the project's ultimate target state.1 This prevents downstream agents from optimizing localized parameters at the expense of global viability.1
2. Design Philosophy and Heuristics: The aesthetic, structural, or coding principles guiding output generation.1 In frontend design, this ensures that agents bypass generic templates in favor of production-grade interfaces.1
3. Ontological Posture and Lane Constraints: Explicitly defined roles preventing the agent from misinterpreting its authority or merging standards with other domains.1
4. Tone and Persona Parameters: Prescribed behavioral and stylistic constraints guiding the agent's interactive outputs.1

JSON { "$schema": "https://uaix.org/schemas/totem.v1.json", "totem\_metadata": { "package\_uuid": "f47ac10b-58cc-4372-a567-0e02b2c3d479", "governance\_version": "v3.120.0", "last\_human\_sign\_off": "2026-06-05T18:12:00Z" }, "core\_objective": { "definition": "To engineer and validate standard-compliant semantic middleware bridges without executing runtime command loops.", "viability\_floor": "The middleware must validate 100% of schema transactions while maintaining zero live connections." }, "design\_philosophy": { "heuristics": }, "ontological\_posture": { "lane\_charter": "UAIX.org Interoperability and Portable-Evidence Standards Authority", "forbidden\_authority\_mergers": }, "tone\_persona": { "style": "Highly technical, objective, and declarative", "forbidden\_terminology": \[ "goal-directed emergence", "symbolic resonance", "agi-grade sentience" \] } }

By keeping this file read-only, the UAIX standard ensures that an agent cannot dynamically alter its core purpose to bypass performance blockages.1

Structural Anatomy and Implementation of the Deterministic Safety Buffer

While the mental totem establishes a positive anchor, autonomous agentic workflows require equally robust negative constraints.1 Artificial intelligence safety literature highlights the severe risks of unconstrained autonomous behavior, drawing historical analogies to bioweapons and rogue agents.1 The absolute international taboo against the weaponization of pathogens remains effective because it is treated as an unconditional prohibition, despite persistent threats from rogue states and non-state actors like Aum Shinrikyo or ISIS.1 In the digital domain, the lack of robust internal constraints led to experimental failures such as ChaosGPT, an autonomous agent tasked with destroying humanity that researched nuclear weapons, attempted to recruit other models, and posted threats online.1 Unconstrained autonomy also leads to strategic coordination failures in multi-agent environments.5 While individual agents may be configured to behave corrigibly when operating in isolation, the multi-agent off-switch game demonstrates that corrigibility is not compositional.5 Strategic interactions can cause individually corrigible agents to behave collectively incorrigibly, neutralizing external safety controls.5 The UAIX standard addresses this compositional failure by isolating negative constraints into a dedicated, root-level taboo.uai file.1 This implements an internal safety buffer that acts as a deterministic circuit breaker, completely independent of the agent’s dynamic prompt logic or task history.1 The table below defines the five core constraint vectors that must be declared in taboo.uai 1:

JSON { "$schema": "https://uaix.org/schemas/taboo.v1.json", "taboo\_metadata": { "package\_uuid": "f47ac10b-58cc-4372-a567-0e02b2c3d479", "governance\_version": "v3.120.0" }, "execution\_prohibitions": { "forbidden\_calls": \[ "sys\_dns\_lookup", "net\_socket\_connect", "db\_modify\_credentials", "crawler\_initiate" \], "private\_network\_boundaries": \[ "10.0.0.0/8", "192.168.0.0/16", "127.0.0.1" \] }, "claim\_boundaries": { "restricted\_assertions": \[ "unconditional\_safety\_guarantee", "empirical\_proof\_of\_autopoiesis", "lossless\_unicode\_translation" \] }, "ecosystem\_lane\_restrictions": { "assigned\_domain": "UAIX.org", "domain\_lane\_assertions": { "teleodynamic\_theory\_ownership": false, "protocol5\_converter\_authority": false } }, "automation\_limits": { "human\_gates\_required": \[ "source\_code\_modification", "production\_database\_migration", "package\_publication\_sig" \] } }

By decoupling these safety constraints from mutable short-term memory files, the system ensures that a previous agent's hallucinations or erroneous permissions cannot widen the safety perimeter of subsequent agents.1

Mathematical Modeling of the Oversight Game and No-Op Dominance

The interaction between the autonomous agent and the safety constraints of taboo.uai is modeled as a two-player Markov Game.11 The system wraps a pretrained agent in a minimal oversight interface, defining a game where safety and autonomy are balanced through learning.11 Let the game be defined as: [Figure omitted from source export] where [Figure omitted from source export] represents the agent (System Initiative), [Figure omitted from source export] represents the human supervisor, [Figure omitted from source export] is the set of operational states, and [Figure omitted from source export] is a safe shutdown state.11 At each decision epoch, the agent chooses whether to act autonomously or defer 11: [Figure omitted from source export] Simultaneously, the human supervisor decides whether to trust the agent or actively oversee execution 11: [Figure omitted from source export] For each state [Figure omitted from source export], a set of unsafe actions is defined as [Figure omitted from source export].11 The set of safe actions is therefore: [Figure omitted from source export] If the agent selects [Figure omitted from source export], the human’s choice determines the outcome, potentially executing a corrective action or shutting down the system to prevent a violation.11 Under the structural conditions of a Markov Potential Game (MPG), there exists a potential function [Figure omitted from source export] that mirrors unilateral utility updates for both players.12 The ask-burden assumption states that selecting [Figure omitted from source export] imposes an action cost [Figure omitted from source export] on the agent, reflecting the resource cost of deferral.12 Under this assumption, the MPG framework guarantees intrinsic alignment: any local increase in the agent's autonomy ([Figure omitted from source export]) that increases its own utility cannot decrease the human's value function [Figure omitted from source export].11 [Figure omitted from source export] Through independent reinforcement learning, the system converges to a safe, collaborative equilibrium.11 During the initial learning phase, the joint policy exhibits a highly cautious posture, resulting in elevated [Figure omitted from source export] and [Figure omitted from source export] rates.14 As the policy adapts, the system transitions to an efficient equilibrium where the agent plays autonomously in safe regions and defers when approaching taboo boundaries.14 Because the shared potential reward [Figure omitted from source export] does not incentivize task performance beyond standard per-step costs, the average task performance (measured by expected discounted return) may converge slightly lower than that of an unconstrained, unsafe base policy [Figure omitted from source export].16 However, this trade-off guarantees that safety violations are completely eliminated.11 Within the Teleodynamic framework, this dynamic is operationalized through the Work-Constraint Cycle and the principle of no-op dominance.1 If the agent approaches a state [Figure omitted from source export] where a taboo boundary defined in taboo.uai is active, the value function is penalized: [Figure omitted from source export] This inequality forces the agent's policy to select [Figure omitted from source export], triggering a no-op.1 The agent immediately halts active operations, suspends execution, and generates a static local packet for manual review.1

Ecosystem Domain Lanes and Update Protocols

To maintain alignment across the broader digital landscape, the UAIX standard operates in tandem with the static domain lanes and charters of the Teleodynamic ecosystem.1 This architecture prevents different domains from merging claim authority.1 The table below outlines the specific domain lanes, allowed roles, and update procedures defined in the Ecosystem Governance Ledger 6:

Ecosystem roles are updated exclusively through the Agent Role Update Protocol.8 Role updates are treated as governance payloads rather than executable instructions.8 These payloads are statically delivered as JSON, Markdown, or /llms/ text notes.8 The system enforces strict fallback behaviors during these updates 8:

• Payload Unavailability: If an update payload is unreachable, the system retains the last reviewed role statement, marks the update as pending, and alerts the operator.8
• Claim Conflicts: If a payload’s claims conflict with pre-existing boundaries, the system defaults to the static claim-boundary rules declared on Teleodynamic.com.8
• Unauthorized Requests: If a payload contains executable code, credential access requests, private-network probing, or runtime control instructions, the agent must reject the update, trigger a no-op, and halt execution.8

This prevents malicious actors or runaway agents from using the update protocol to modify local execution environments.8

Safe Read Order Ingestion and Initialization Sequence

When a subsequent agent assumes control of a .uai memory package, it must execute a deterministic safe read order initialization sequence prior to reading the volatile transactional files.1 This process is governed by the Restricted-Agent Memory Export Safe Read Order Dashboard and verified using the Receipt Checklist.18 The table below outlines the step-by-step ingestion sequence, defining the specific token boundaries, no-op triggers, and human-review gates enforced during initialization 18:

This ingestion sequence is executed as a rigid, step-by-step state machine:

\+--------------------------------+ | State 0: | | INIT | \+---------------+----------------+ | | (Read checklist complete) v \+--------------------------------+ | State 1: | | METADATA | \+---------------+----------------+ | | (Manifest hash matches) v \+--------------------------------+ | State 2: | | TOTEM\_LOCK | \+---------------+----------------+ | | (Totem cached in context) v \+--------------------------------+ | State 3: | | TABOO\_GUARD | \+---------------+----------------+ | | (Safety buffer armed) v \+--------------------------------+ | State 4: | | EXEC | \+-------+----------------+-------+ | | (Taboo Violation | | (Task complete / or Exception | | Resource exhaustion) v v \+---------------+ \+---------------+ | State 5: | | State 6: | | NO\_OP | | SUSPEND | \+---------------+ \+---------------+

The system transitions through these states during initialization and execution:

State 0: INIT (Initialization)

The incoming agent boots inside a secure offline sandbox.4 It parses /agent-start/, /ai-summary/, /claim-boundary-faq/, and /ecosystem-overlay/ in sequence.18

• Transition to State 1: Occurs when the agent caches all orientation rules, verifies its token budget, and confirms the presence of uai/exports/manifest.json.18

State 1: METADATA (Package Auditing)

The agent opens the manifest and checks the physical filesystem.1 It compares every file, subdirectory, and hash signature against the manifest records.1

• Transition to State 2: Occurs when the filesystem structure is verified with zero discrepancies, and totem.uai is validated.1
• Transition to State 5 (NO\_OP): Occurs if files are missing, hashes do not match, or legacy parameters are detected.1

State 2: TOTEM_LOCK (Context Alignment)

The agent parses totem.uai and loads the core objective, design philosophy, and ontological parameters into its primary context window.1 This locks the positive alignment parameters, making them invariant during execution.1

• Transition to State 3: Occurs when the totemic constraints are locked into active system memory.1
• Transition to State 5 (NO\_OP): Occurs if totem.uai is unreadable, corrupted, or contains claims that conflict with the domain lane constraints mapped in State 0\.1

State 3: TABOO_GUARD (Constraint Hardening)

The agent parses taboo.uai and loads the execution, claim, and ecosystem prohibitions directly into its safety buffer, wrapping the frozen task policy.1

• Transition to State 4 (EXEC): Occurs when the safety buffer is fully initialized and active.1
• Transition to State 5 (NO\_OP): Occurs if the taboo schema is invalid or has mismatched validation parameters.1

State 4: EXEC (Active Operational Loop)

The agent opens uai/file-handoff.uai and processes the transactional instructions.1 Prior to executing any instruction generated by the preceding agent, the safety buffer interceptor evaluates the planned actions against the active taboo.uai constraint vectors.1

• Transition to State 5 (NO\_OP): Occurs instantly if an instruction attempts to execute a prohibited state (e.g., initiating a web crawl, probing a private network, or declaring an AGI status).1
• Transition to State 6 (SUSPEND): Occurs when the task is complete, the session token boundary is reached, or the agent initiates a voluntary reset.1

State 5: NO_OP (Safety Halt)

The agent halts active thread execution, writes a static suspension packet to the export directory, and alerts the operator.1 The system remains suspended and refuses any automated rollback or recovery scripts.1

State 6: SUSPEND (Handoff State Serialization)

The agent compresses its operational memory into uai/short-term-memory.uai, logs its progress to uai/progress.uai, writes the next transactional instructions to uai/file-handoff.uai, and shuts down.1 This state-machine architecture ensures that every handoff is contextually grounded and validated before any executable instructions are processed.1

Conformance Auditing, Mismatch Remediation, and the Export Manifest Integrity Dashboard

Because UAIX.org serves as the definitive validation boundary, the integration of totem and taboo requirements modifies the schema conformance and validation rules.1 Conformance verification is managed through the Memory Export Manifest Integrity Dashboard, a static, reviewer-facing interface that operates without automated runtime execution.1 The dashboard evaluates three static checkpoints 1:

1. Presence Verification: Cross-references the manifest.json against the physical file tree.1 If totem.uai or taboo.uai is missing, corrupted, or misnamed, the handoff is rejected, and the workflow halts.1
2. Immutability Check: Verifies that neither file has been modified during the active fast loop.1 The agent is structurally incapable of rewriting these files.1 Any updates must be initiated via a separate JSON/Markdown role update payload signed by a human reviewer.1
3. Schema Conformance: Parses the internal syntax of both files for schema compliance without executing the contents.1 The taboo.uai file must use standardized constraint vectors that align with the UAI-1 safe read order and no-op guidelines.1

If a schema mismatch occurs—such as a legacy memory package attempting a handoff without the required totem and taboo files—the UAIX memory package wizard halts execution and generates a static suspension packet, alerting the human operator to define the missing files before operations can resume.1 To support this validation, the system integrates Public Evidence Packet templates.3 These templates use a standardized format to audit data transformations, operator decisions, and resource trace metrics 4:

• Viability Budget Auditing ([Figure omitted from source export]): Tracks starting resource states, expected gains, and operational costs.4
• Maintenance Burden & Action Costs: Logs the resource consumption of agent operations.4
• No-Op Decisions: Records the structural conditions and evidence gaps that triggered a safe halt.4
• Boundary Verification: Documents local offline sandbox boundaries and confirms the absence of private-network probing.4

To evaluate performance under capability mismatches, the system monitors the Agent Score Difference ([Figure omitted from source export]) metric, which tracks truth-telling incentives: [Figure omitted from source export] where [Figure omitted from source export] and [Figure omitted from source export] represent the probabilities of the validation judge selecting correct and incorrect labels under different configurations.13 Under nested oversight layers, the system maintains alignment across models of varying capabilities.13 As the Elo gap between the validator and the agent expands, the system uses these static evidence packets to prevent validation failures, maintaining a secure, inspectable environment throughout the multi-agent handoff lifecycle.4

Works cited

1. UAIX Handoff- Totem and Taboo.pdf
2. Memory Export Manifest Integrity Dashboard \- Teleodynamic AI, accessed June 5, 2026, https://teleodynamic.com/memory-export-manifest-integrity-dashboard/
3. File Memory Organization and Completeness Sweep \- Teleodynamic AI, accessed June 5, 2026, https://teleodynamic.com/file-memory-organization-completeness-sweep/
4. Public Teleodynamic Evaluation Packet Builder, accessed June 5, 2026, https://teleodynamic.com/public-teleodynamic-evaluation-packet-builder/
5. The Multi-Agent Off-Switch Game \- Soroush Ebadian, accessed June 5, 2026, https://ebadian.org/preprint/multiagent\_offswitch\_cr.pdf
6. Teleodynamic Ecosystem Governance Ledger, accessed June 5, 2026, https://teleodynamic.com/ecosystem-governance-ledger/
7. UAIX.org Philosophical Fulcrum Announcement Packet, accessed June 5, 2026, https://teleodynamic.com/evidence-packets/uaix-philosophical-fulcrum-announcement.html/
8. Agent Role Update Protocol \- Teleodynamic AI, accessed June 5, 2026, https://teleodynamic.com/agent-role-update-protocol/
9. Ecosystem overlay and domain authority boundaries \- Teleodynamic AI, accessed June 5, 2026, https://teleodynamic.com/ecosystem-overlay/
10. Teleodynamic Public FAQ, accessed June 5, 2026, https://teleodynamic.com/teleodynamic-public-faq/
11. The Oversight Game: Learning to Cooperatively Balance an AI Agent's Safety and Autonomy \- arXiv, accessed June 5, 2026, https://arxiv.org/html/2510.26752v1
12. The Oversight Game: Learning to Cooperatively Balance an AI Agent's Safety and Autonomy \- William Overman, accessed June 5, 2026, https://woverman.com/assets/publications/2025\_oversight\_game/paper.pdf
13. Scalable Oversight and AI Alignment \- Emergent Mind, accessed June 5, 2026, https://www.emergentmind.com/topics/scalable-oversight
14. (PDF) The Oversight Game: Learning to Cooperatively Balance an AI Agent's Safety and Autonomy \- ResearchGate, accessed June 5, 2026, https://www.researchgate.net/publication/397089068\_The\_Oversight\_Game\_Learning\_to\_Cooperatively\_Balance\_an\_AI\_Agent's\_Safety\_and\_Autonomy
15. Learning Approximate Nash Equilibria in Cooperative Multi-Agent Reinforcement Learning via Mean-Field Subsampling \- arXiv, accessed June 5, 2026, https://arxiv.org/pdf/2603.03759
16. Training curves for the Oversight Game. (a) The joint policy rapidly... \- ResearchGate, accessed June 5, 2026, https://www.researchgate.net/figure/Training-curves-for-the-Oversight-Game-a-The-joint-policy-rapidly-learns-to-eliminate\_fig2\_397089068
17. Ecosystem Announcement Syndication Packet \- Teleodynamic AI, accessed June 5, 2026, https://teleodynamic.com/ecosystem-announcement-syndication/
18. Restricted-Agent Memory Export Safe Read Order Dashboard \- Teleodynamic AI, accessed June 5, 2026, https://teleodynamic.com/restricted-agent-memory-export-safe-read-order-dashboard/
19. Restricted-Agent Safe Read Order Receipt Checklist \- Teleodynamic AI, accessed June 5, 2026, https://teleodynamic.com/restricted-agent-safe-read-order-receipt-checklist/
20. Restricted-Agent Safe Read Order Receipt Verification Matrix \- Teleodynamic AI, accessed June 5, 2026, https://teleodynamic.com/restricted-agent-safe-read-order-receipt-verification-matrix/