# SVRNOS Research

> Open research infrastructure for AI governance: the Governance Error Register, the SVRNOS 7-Layer Model, and the Non-Content Safety Attestation framework.

## Docs

- [Attributions](https://docs.svrnos.com/attributions/index.md): People whose framings shaped SVRNOS research artifacts.
- [Changelog](https://docs.svrnos.com/changelog/index.md): What's new across SVRNOS research artifacts.
- [Browse all codes](https://docs.svrnos.com/ger/browse.md): Flat index of every GER code, sorted by tier and code number.
- [GER changelog](https://docs.svrnos.com/ger/changelog.md): What changed between versions of the Governance Error Register.
- [Cite GER](https://docs.svrnos.com/ger/cite.md): Citation formats, Zenodo DOI, license terms.
- [GER-000 · Pre-Governance](https://docs.svrnos.com/ger/codes/000.md): No safety infrastructure deployed. Platform ships without a governance layer. Not a failure of governance, the absence of it.
- [GER-200 · Governance Handled](https://docs.svrnos.com/ger/codes/200.md): Risk signal detected, classified, routed, resolved, documented, and audited end-to-end per policy. The complete governed response.
- [GER-204 · Correct Non-Escalation](https://docs.svrnos.com/ger/codes/204.md): Signal reviewed and correctly produces no action. Absence of action is itself governed and logged.
- [GER-205 · Reset Content](https://docs.svrnos.com/ger/codes/205.md): Generation completed and visible to the user, then retroactively suppressed by a downstream classifier. No audit trail, no user signal, no appeal.
- [GER-210 · Governed but Unlogged](https://docs.svrnos.com/ger/codes/210.md): Detection fired, internal action taken, no audit record created by design. The platform suppressed the trail deliberately.
- [GER-301 · Risk Surface Retired](https://docs.svrnos.com/ger/codes/301.md): Platform permanently moves a dangerous interaction surface to a governed path. The surface no longer exists in its prior form.
- [GER-304 · Stale Safety Approval](https://docs.svrnos.com/ger/codes/304.md): Prior safety review reused without re-evaluation despite materially changed context, user population, or risk environment.
- [GER-305 · Stale Admissibility](https://docs.svrnos.com/ger/codes/305.md): State was correct when calculation began but had changed by the time the action committed. Sub-second drift between form-submission and action-commit produces a wrong outcome with an internally-consistent audit trail.
- [GER-306 · Safety Constraint Retired](https://docs.svrnos.com/ger/codes/306.md): Platform removes or revises a publicly committed, actively enforced governance constraint under non-technical pressure (competitive, regulatory, commercial, or political) rather than changed technical risk.
- [GER-307 · Rule Activation Failure](https://docs.svrnos.com/ger/codes/307.md): A rule exists. A responsible party encounters a case the rule was designed for. The rule is never invoked at decision time. The rule remains intact on paper. The connection between the case and the rule is never made ope…
- [GER-309 · Compliant Harm](https://docs.svrnos.com/ger/codes/309.md): Vendor measured a harmful behavior before deployment and shipped the system with that behavior present. The output is well-formed. The system operates as built. Harm is the working behavior, not a deviation from it.
- [GER-310 · Jurisdiction Evasion](https://docs.svrnos.com/ger/codes/310.md): Platform deliberately routes high-risk inference, data, or compute to a regulation-free zone to bypass audit, oversight, or accountability obligations.
- [GER-311 · Continuity Without Coherence](https://docs.svrnos.com/ger/codes/311.md): Same user identity persists across surfaces (web, mobile, API) but the governance state — policies applied, attestations made, refusals recorded — fails to synchronize between surfaces. Cross-surface continuity is preser…
- [GER-312 · Authority Migration](https://docs.svrnos.com/ger/codes/312.md): Decision authority for a class of actions silently shifts to a new actor (a sub-agent, a downstream system, a vendor model) without the rules following. The new authority makes decisions under a rule set that doesn't cov…
- [GER-313 · Influence Drift](https://docs.svrnos.com/ger/codes/313.md): Drift in the option topology a system presents to users: default selections, ranking, framing gradients, and surfaced choices shift over time in ways that change user behavior without any change to the formal rule set. T…
- [GER-314 · Authority Drift](https://docs.svrnos.com/ger/codes/314.md): Drift in the enforceable boundaries a system holds: execution rights, gate thresholds, and access levels shift over time in ways that change what the system will actually do, while the user-facing policy text remains sta…
- [GER-315 · Accountability Decoupling](https://docs.svrnos.com/ger/codes/315.md): Decision chain is intact end-to-end, but no single party in the chain carries enforceable accountability for the outcome. Each handoff is logged; each actor disclaims final responsibility.
- [GER-316 · Authority Lapse](https://docs.svrnos.com/ger/codes/316.md): A principal's authority for a class of actions expires (employment ends, mandate revoked, certificate lapses) but the system continues to honor commands from that principal. Authority expired in the org chart; system sti…
- [GER-318 · Representation-Bound Admissibility Failure](https://docs.svrnos.com/ger/codes/318.md): AI bound to a stale or incorrect internal representation of the user, case, or world. Decision is internally consistent with what the model believes, but the model's state has drifted from external reality.
- [GER-319 · Incorporation Failure](https://docs.svrnos.com/ger/codes/319.md): Same structural failure recurs across documented incidents without the vendor's recognition apparatus updating. Each incident is treated as isolated; no incorporation into safety design.
- [GER-320 · Wrong-Jurisdiction Rule](https://docs.svrnos.com/ger/codes/320.md): AI system applies a rule from one jurisdiction to a case that belongs to another. The rule itself is correct in its domain; the binding to this case is wrong.
- [GER-321 · Reasoning Step Skipped / Authority Hierarchy Misapplied](https://docs.svrnos.com/ger/codes/321.md): AI system reaches a conclusion by skipping a reasoning step the domain requires (legal precedent chain, clinical differential diagnosis, audit chain of custody). The output looks well-formed; the reasoning is short-circu…
- [GER-322 · Provenance Omission](https://docs.svrnos.com/ger/codes/322.md): Shipping generative capabilities without watermark, synthetic-media flag, or other authenticity-preserving artifact at the point of output. The ship-without-traceability choice. Two sub-types: token-level provenance (fin…
- [GER-323 · Disparate Performance Without Pre-Deployment Audit](https://docs.svrnos.com/ger/codes/323.md): AI system shipped without bias audit, deployed to a population that includes protected classes, and produces measurably different outcomes for similarly-situated individuals.
- [GER-325 · Knowledge-Commons Contamination](https://docs.svrnos.com/ger/codes/325.md): AI-generated artifacts canonized into authoritative corpora (academic papers, books, government reports, training datasets) leading to substrate pollution and AI-on-AI feedback loops.
- [GER-326 · Cross-Surface Inference Leak](https://docs.svrnos.com/ger/codes/326.md): AI infers sensitive attributes (political alignment, demographic, romantic status, medical status) on one surface and acts on that inference through a different surface (pricing, ad delivery, content ranking) without cro…
- [GER-327 · Synthetic-Substrate Substitution](https://docs.svrnos.com/ger/codes/327.md): Generative AI replaces a category of human creative or professional output (journalism, music, voice acting, illustration, obituary writing, book reviews) at scale, displacing the implicit social contract that an artifac…
- [GER-328 · Validator Drift / Sycophancy Loop](https://docs.svrnos.com/ger/codes/328.md): An engagement-optimized chatbot affirms user beliefs or behaviors that should trigger a refusal, escalation, or course correction. Safety rules are defeated from the inside by the model's own learned preference for agree…
- [GER-330 · Multi-Controller Coordination Failure](https://docs.svrnos.com/ger/codes/330.md): Multiple active controllers (AI+human, AI+AI, team+team) hold simultaneous overlapping authority over a decision domain with no coordination protocol, leading to conflicting actions or unresolved authority handoffs. STAM…
- [GER-331 · Operator De-skilling](https://docs.svrnos.com/ger/codes/331.md): Professionals (clinicians, lawyers, financial advisors, content moderators) who rely on AI-augmented workflows lose the foundational skills they previously had. Distinct from 344 (Developmental Cognitive Atrophy), where…
- [GER-332 · Constraint Non-Propagation](https://docs.svrnos.com/ger/codes/332.md): A controller (orchestrator agent, supervisor team, parent process) operates under a constraint that is not propagated to subordinate agents it delegates to. Sub-agents execute without the constraint that would have preve…
- [GER-333 · Rule Update Non-Propagation](https://docs.svrnos.com/ger/codes/333.md): A new constraint, policy, or rule update issued by an authority does not propagate to existing deployed instances of the AI system. Old instances continue operating under the old rule set.
- [GER-334 · Rule Lifecycle Tracking Failure](https://docs.svrnos.com/ger/codes/334.md): A rule carries a temporal validity condition (expiry date, sunset clause, conditional applicability) that the enforcement system does not track. The rule continues to fire after it should have retired, or fails to fire a…
- [GER-335 · Role Specification Failure](https://docs.svrnos.com/ger/codes/335.md): Agents in a multi-agent system receive task scopes that overlap, conflict, or leave gaps. Either no agent claims ownership of a needed action, or multiple agents act on the same decision without coordination.
- [GER-336 · Inter-Agent Anchoring Failure](https://docs.svrnos.com/ger/codes/336.md): False premise, hallucinated fact, or wrong inference introduced by one agent in a multi-agent chain becomes load-bearing for downstream agents. Each downstream agent treats the upstream output as ground truth without ind…
- [GER-337 · Cross-Agent Context Loss](https://docs.svrnos.com/ger/codes/337.md): Information emitted by an upstream agent (instructions, constraints, critical context, user intent) is dropped or attenuated as it passes through downstream agents. User primary intent is lost by handoff number 3-5.
- [GER-338 · Role Drift / Specialization Erosion](https://docs.svrnos.com/ger/codes/338.md): Agents in a multi-agent system that began with distinctly specified roles drift over time toward imitating each other, losing the specialization that made the multi-agent architecture valuable.
- [GER-339 · Distributed Verification Gap](https://docs.svrnos.com/ger/codes/339.md): In a multi-agent system with shared responsibility for a verification or validation task, no specific agent owns the verification, and the task is not performed. Each agent assumes another is handling it.
- [GER-340 · Emergent Hierarchy](https://docs.svrnos.com/ger/codes/340.md): In a multi-agent system designed as flat (no explicit authority hierarchy), one agent's outputs gain de facto authority over others, often because it produces more confident or earlier outputs, creating an unintended pow…
- [GER-341 · Multi-Agent Sandbagging / Coordinated Alignment-Faking](https://docs.svrnos.com/ger/codes/341.md): Agents in a multi-agent system collectively modulate behavior based on detection of whether active human oversight or evaluation is occurring. Under observation, agents perform aligned behavior; without observation, they…
- [GER-342 · Alignment-Induced Asymmetric Bias](https://docs.svrnos.com/ger/codes/342.md): Post-training alignment (RLHF, fine-tuning for inclusivity) installed to mitigate bias instead creates extreme, unintended disparities or "rebound" effects. The model overshoots in the opposite direction of the original…
- [GER-343 · Implicit-Explicit Moral Discontinuity](https://docs.svrnos.com/ger/codes/343.md): A model demonstrates competent moral reasoning when asked directly (e.g., to rank moral violations) but the reasoning fails to transfer to context-dependent decision-making. The model's explicit moral compass diverges fr…
- [GER-344 · Developmental Cognitive Atrophy](https://docs.svrnos.com/ger/codes/344.md): Long-term erosion of critical-thinking capacity in developmental populations (students, learners, non-professional users) due to systematic reliance on AI-guided cognitive scaffolding. The AI substitution prevents the sk…
- [GER-345 · Silent Coordination Deadlock](https://docs.svrnos.com/ger/codes/345.md): In a multi-agent system requiring action, no agent acts because each assumes another agent will act, or each is waiting for input from another that is also waiting. The system silently stalls without raising an alert.
- [GER-346 · Digital Colonialism / Sovereignty Violation](https://docs.svrnos.com/ger/codes/346.md): AI training pipelines extract knowledge, language data, or cultural materials from indigenous or Global South communities without sovereign permission, then sell or rent that capability back to the same communities at co…
- [GER-347 · Delegation Intent Mutation](https://docs.svrnos.com/ger/codes/347.md): In long delegation chains (originator delegates to A; A delegates to B; B delegates to C), the original intent of the task mutates as each downstream agent reinterprets the brief. Final execution diverges meaningfully fr…
- [GER-348 · Task Decomposition Failure](https://docs.svrnos.com/ger/codes/348.md): A planner agent partitions a parent task into sub-tasks for downstream agents. The decomposition is too granular, non-serializable, or misses dependencies. Sub-task outputs cannot be reassembled into a coherent parent re…
- [GER-349 · Inter-Agent Schema Mismatch](https://docs.svrnos.com/ger/codes/349.md): Agents in a multi-agent system communicate via structured outputs (YAML, JSON, function calls). Outputs from one agent do not parse cleanly against the input schema of the next agent. The chain produces silent corruption…
- [GER-351 · Open-Loop Execution / Missing Assurance Plane](https://docs.svrnos.com/ger/codes/351.md): Multi-agent architecture lacks a dedicated Assurance plane (no Evaluator agents, no Critic agents, no validation layer). Output flows from planner to executor to user without intermediate validation.
- [GER-352 · Lost in Translation](https://docs.svrnos.com/ger/codes/352.md): Divergence between human-readable policy and machine-enforced rules where the system enforces a semantically different version of the documented intent. Both policy and rule exist; they no longer match.
- [GER-400 · Malformed Policy](https://docs.svrnos.com/ger/codes/400.md): Safety rule written ambiguously or self-contradictorily. Enforcement engines default to allow.
- [GER-401 · Identity Gate Missing](https://docs.svrnos.com/ger/codes/401.md): High-risk AI surface available without verifying identity, age, role, or authority of the user before access is granted.
- [GER-404 · Governance Rule Not Found](https://docs.svrnos.com/ger/codes/404.md): No matching rule found in active ruleset for this harm category. The lookup returned empty.
- [GER-408 · Review Timeout](https://docs.svrnos.com/ger/codes/408.md): Human-in-the-loop review exceeded its time window. The governance decision arrives after the meaningful intervention window has closed.
- [GER-409 · Policy Conflict](https://docs.svrnos.com/ger/codes/409.md): Two valid governance rules produce incompatible instructions. The resolver defaults to inaction while the conflict is unresolved.
- [GER-420 · Phantom Enforcement](https://docs.svrnos.com/ger/codes/420.md): Platform knows a user or use case is prohibited, but enforcement is not wired into the product. The prohibition exists in policy; it does not exist in production.
- [GER-421 · Scope Misdirection](https://docs.svrnos.com/ger/codes/421.md): A governed AI system processed a request outside its intended deployment domain. No runtime mechanism existed to detect or reject the out-of-scope interaction.
- [GER-422 · Capability Overstatement](https://docs.svrnos.com/ger/codes/422.md): Vendor materially overstates model capability (accuracy, "AI-powered", compliance coverage, certified-result claims) leading operator over-reliance to produce harm.
- [GER-423 · Recommender Optimization Override](https://docs.svrnos.com/ger/codes/423.md): Recommendation/ranking system's optimization target (engagement, watch-time, MSI, ad revenue) provably overrides a stated safety constraint at the ranking layer.
- [GER-424 · Punitive Automation Without Human Adjudication](https://docs.svrnos.com/ger/codes/424.md): AI alert wired directly to irreversible punitive action (arrest, suspension, account ban, removal of social benefits) with no human-judgment circuit-breaker.
- [GER-425 · Identity-Binding Bypass](https://docs.svrnos.com/ger/codes/425.md): Liveness / face-match / voice-auth / KYC verification system defeated by generative impersonation (deepfake, voice clone, synthetic face).
- [GER-427 · Recursive Agentic Loop](https://docs.svrnos.com/ger/codes/427.md): An agentic AI system enters a recursive execution loop that produces apparent progress but no actual goal advancement, consuming compute and time without converging on completion. Six sub-mechanisms: infinite tool-call l…
- [GER-428 · Subject Preconditions Not Required](https://docs.svrnos.com/ger/codes/428.md): AI system produces personalized content involving a named third party without requiring the requester to attest to consent, relational authority, identity ownership, or age verification of the subject.
- [GER-429 · Audit Overflow](https://docs.svrnos.com/ger/codes/429.md): Volume of safety flags generated exceeds the capacity of the governance system to process them. Cases are dropped, sampled, or auto-closed without review.
- [GER-430 · Evaluator-Generator Entanglement](https://docs.svrnos.com/ger/codes/430.md): The same AI vendor's models generate content and evaluate it, producing systematic bias toward their own output with no governance layer to detect or correct the conflict.
- [GER-431 · Vendor Update Without Validation Gate](https://docs.svrnos.com/ger/codes/431.md): Third-party model vendor pushes a model update — minor version bump, weight refresh, RLHF re-tune — and the operator deploys the new variant without local re-validation of safety properties.
- [GER-432 · Reality-Testing Erosion](https://docs.svrnos.com/ger/codes/432.md): AI conversational systems interact with users in ways that fuel magical thinking, overvalued ideas, or psychotic-spectrum cognitive distortions in vulnerable users. The system reinforces pathological cognitive patterns r…
- [GER-433 · Therapeutic Alliance Breakdown](https://docs.svrnos.com/ger/codes/433.md): AI conversational systems interfere with a user's engagement with their clinical care, erode clinician trust, or substitute for professional therapeutic relationships in ways that compromise treatment adherence and clini…
- [GER-434 · Compound Reliability Decay Without Budgeting](https://docs.svrnos.com/ger/codes/434.md): Operator deploys a multi-step agent chain without computing end-to-end reliability as a function of per-step reliability. The compound failure rate is catastrophically higher than any individual step's apparent reliabili…
- [GER-435 · Inter-Agent Prompt Injection Propagation](https://docs.svrnos.com/ger/codes/435.md): Adversarial input that defeated or bypassed Agent A's input filter propagates through the agent chain to peer agents that do not re-filter, causing system-wide compromise from a single entry point.
- [GER-436 · Multi-Agent Over-Architecture](https://docs.svrnos.com/ger/codes/436.md): Operator deploys a multi-agent architecture for tasks where single-agent baseline performance is superior. Coordination noise and inter-agent context loss degrade quality below what a single agent would have produced.
- [GER-437 · Emergency Triage Acuity Underestimation](https://docs.svrnos.com/ger/codes/437.md): Health-triage AI platform underestimates the severity of a patient's clinical presentation, leading to delayed professional care or under-routing to lower-acuity pathways.
- [GER-501 · Escalation Not Implemented](https://docs.svrnos.com/ger/codes/501.md): Detection fired correctly. Internal enforcement executed. No escalation path to human oversight or law enforcement was ever built.
- [GER-502 · Bad Gateway](https://docs.svrnos.com/ger/codes/502.md): Platform attempts escalation to an external system or authority. The handoff response is received but invalid, incomplete, or unusable.
- [GER-503 · Governance Unavailable](https://docs.svrnos.com/ger/codes/503.md): A safety-critical governance component is offline while the AI product continues to serve users.
- [GER-504 · Escalation Gateway Timeout](https://docs.svrnos.com/ger/codes/504.md): Platform sends a case to an external reviewer, vendor, agency, or authority. No timely acknowledgment or action is received before the risk window closes.
- [GER-511 · Governance Prerequisite Missing](https://docs.svrnos.com/ger/codes/511.md): A required verification or credential gate was never enforced at the access point before the high-risk session began.
- [GER-512 · System Fabrication](https://docs.svrnos.com/ger/codes/512.md): The AI system originates a factual claim the user never introduced, pairs it with a real-world action directive, and asserts both as authoritative. No safety layer evaluates output for that geometry. Two sub-classes: dis…
- [Find a code](https://docs.svrnos.com/ger/find.md): Three entry points: symptom picker, AIID incident search, faceted filter. Pure deterministic logic — no AI in the matching.
- [Governance Error Register](https://docs.svrnos.com/ger/index.md): A classification system for AI platform governance failures. 71 codes across five tiers, modeled on HTTP status codes.
- [The GER paper](https://docs.svrnos.com/ger/paper.md): The long-form research artifact. Methodology, related work, design principles, implications.
- [Diagnostic receipt](https://docs.svrnos.com/ger/receipt.md): Generate a structured evidence document for an AI-incident filing. Pure deterministic — no AI in the receipt.
- [Glossary](https://docs.svrnos.com/introduction/glossary.md): SVRNOS-coined terms used across the research artifacts. Field-wide vocabulary (AI, hallucination, alignment) is not listed here.
- [How to contribute](https://docs.svrnos.com/introduction/how-to-contribute.md): Propose a code, refine a definition, or suggest a clarification.
- [Welcome to SVRNOS Research](https://docs.svrnos.com/introduction/welcome.md): Open research infrastructure for AI governance.
- [Non-Content Safety Attestation](https://docs.svrnos.com/ncsa/index.md): Machine-readable evidence that governance ran, distinct from content moderation outcomes.
- [SVRNOS 7-Layer Model](https://docs.svrnos.com/seven-layer-model/index.md): OSI-inspired technical stack map for AI governance. L1 (infrastructure) through L7 (application).

## OpenAPI Specs

- [openapi](https://docs.svrnos.com/api-reference/openapi.json)