AI Transparency Notice

1. Purpose and scope

Neutron AI is a scoped memory and bounded consequence platform for developers and organisations building AI agents. It stores customer-directed memory, retrieves relevant context, and can help compare possible actions and record approvals and outcomes. It is not a general autonomous decision-maker and does not guarantee future events.

This Notice describes AI and machine-learning processing in Neutron’s hosted Service. It supplements the Terms of Service, Privacy Notice, and Acceptable Use Policy. A customer’s own application may add models, tools, data, prompts, and decisions outside Neutron’s control.

2. What Neutron does

Neutron combines several techniques with different risk profiles:

• deterministic controls validate identifiers and policies, enforce budgets and permissions, hash or redact data, rank candidates, manage retention, and structure reviewable records;
• embeddings convert authorised memory summaries into numerical representations so semantically related information can be found;
• retrieval and context packing select and bound relevant memory for a repeated workflow;
• optional generative inference is invoked only by an authorised feature or customer instruction and can draft or analyse content; and
• Consequence Thought creates bounded scenarios, decisions, approval records, outcome observations, and reflection lessons using deterministic-first orchestration and, where configured, model assistance.

A response may therefore be fully deterministic, retrieval-based, AI-assisted, or a combination. Product surfaces and API fields should be read together to understand the mode used.

3. Embeddings and semantic retrieval

For hosted semantic search, Neutron may send an authorised memory summary to Cloudflare Workers AI to generate an embedding using the configured embedding model, currently @cf/baai/bge-base-en-v1.5. The numerical vector is stored in Neutron’s scoped semantic index with tenant and scope metadata. Vector search is treated as eventually consistent and is not the authoritative record; scoped Durable Object storage remains the hot source of truth.

An embedding is not intended to reproduce the original text, but it can still reveal relationships and must be protected as Customer Content. Neutron applies the same tenant and scope boundary to semantic retrieval and validates authoritative records before returning context. Deleted memory is tombstoned so it is not intentionally restored through vector indexing, queues, caches, compaction, or archives.

4. Optional hosted and customer-selected model inference

Some features can call a model through Neutron’s hosted Cloudflare Workers AI binding. Supported responses expose execution metadata such as the mode and model identifier so an integrating application can distinguish deterministic or hosted-model behaviour. The exact model may vary by customer configuration, feature, availability, and documented updates.

Customers may also use their own model provider or gateway after retrieving Neutron context. Those downstream prompts, provider credentials, model choices, safety settings, and outputs are controlled by the customer and governed by that provider’s terms. Neutron does not need a customer’s provider API key for ordinary memory retrieval, and customers should keep those keys inside their own secure boundary.

Customer Content is sent to a hosted model only when an authorised workflow invokes that processing. Data sent through Neutron’s hosted inference is handled by the Subprocessors identified on the Subprocessor List.

5. Consequence Thought

Consequence Thought is a bounded decision-review workflow. It can create alternative scenarios, identify assumptions and risks, require approvals, observe outcomes, and write reviewed lessons back to scoped memory. It predicts possible consequences under stated assumptions; it does not prove what will happen or identify a mathematical optimum unless a separate verified solver and proof boundary expressly support that claim.

Simulation depth, nodes, branches, runtime, model calls, scenarios, and cost are bounded. Deep runs at depth 50 or greater use supported asynchronous and cancellable processing. Depth 1000 requires explicit confirmation. High-risk or irreversible actions require a recorded human approval, and external action requires a separately authorised, policy-checked, idempotent adapter and recovery plan.

The system stores structured artifacts—such as assumptions, alternatives, risks, decisions, approvals, observations, and concise rationales—not raw hidden chain-of-thought, private model scratchpads, or system prompts. Sensitive-domain reflection lessons require review before memory write-back.

6. Model training and Service improvement

Neutron does not use Customer Content to train general-purpose models or Neutron models unless the customer gives separate, explicit written opt-in consent identifying the data and purpose. Buying or using the Service is not consent to model training.

We may use Usage Data and de-identified, aggregated statistics that exclude the substance of Customer Content and do not reasonably identify a person or customer to measure reliability, security, cost, latency, and feature performance. Customer-selected model providers may have different training or retention terms, which the customer must assess.

7. Known limitations

AI-assisted and retrieval outputs can:

• be inaccurate, incomplete, outdated, fabricated, overconfident, biased, or internally inconsistent;
• omit relevant memory because of scope, permissions, expiry, deletion, indexing delay, ranking, or context budgets;
• retrieve plausible but irrelevant context or reflect bias and errors in Customer Content;
• vary across model versions, configurations, repeated requests, languages, and regions;
• resemble third-party content or another user’s independently generated Output without implying a transfer of ownership; and
• misestimate likelihood, impact, causation, or the effect of an action.

Confidence labels, citations, scenarios, scores, and explanations are aids to review, not guarantees. Customers must test with representative data, monitor production behaviour and drift, verify important sources, and maintain a safe fallback.

8. Human review, approvals, and external action

A qualified person must meaningfully review material Output before it informs a high-impact, regulated, safety-critical, irreversible, or externally executed action. Review must consider the underlying data, uncertainty, alternatives, affected people, bias, legal duties, permissions, recovery, and whether no action is safer.

Approval must be specific to the proposed action and current evidence. A prior approval, model recommendation, memory entry, or broad agent instruction must not be treated as approval for a materially different action. An authorised operator must be able to reject, pause, cancel, correct, and where feasible reverse an action.

Neutron is not a substitute for medical, legal, financial, employment, safety, or other professional judgement. Customers are responsible for placing qualified professionals in the workflow where required.

9. Disclosure, marking, and provenance

Neutron product surfaces identify AI-oriented memory and consequence features, and supported hosted-inference responses include model-mode metadata. Customers must preserve relevant metadata and clearly tell people when they are interacting directly with an AI system unless that interaction is obvious or an applicable exception applies.

Where law requires synthetic audio, image, video, or text to be marked in a machine-readable format or disclosed to recipients, the customer application that creates or publishes that content must implement and preserve the required marking. Neutron’s memory APIs do not by themselves label every downstream artifact created by a customer’s separate model provider.

Customers should retain source references, timestamps, model or execution mode, policy version, approvals, and relevant transformation history in proportion to risk. They must not remove provenance or safety notices in a deceptive way or claim that Neutron has independently verified an Output when it has not.

10. Customer and integrator responsibilities

Before deploying a Neutron-enabled AI workflow, customers and integrators must:

• identify their role under applicable AI, privacy, consumer, equality, sector, and product-safety law;
• classify the use case and avoid prohibited practices under the AUP;
• provide required instructions, transparency notices, accessibility, explanations, and routes to human contact or contest;
• ensure training, testing, validation, logging, cybersecurity, human oversight, quality management, registration, and post-market monitoring duties are met where applicable;
• use only necessary, lawful, accurate, and appropriately retained personal data;
• evaluate model and integration providers, geographic transfers, intellectual-property rights, and contractual restrictions;
• test representative languages, groups, edge cases, attacks, failure modes, and recovery paths; and
• stop or change a workflow when evidence shows unacceptable risk or non-compliance.

Neutron documentation and technical controls support this work but do not determine a customer’s legal classification or complete its compliance assessment.

11. Questions, feedback, and individual rights

Questions about a specific Output should first be directed to the customer or application that presented it, because that organisation controls the input, purpose, model configuration, and decision. Privacy rights are explained in the Privacy Notice.

To report a harmful Output, apparent policy violation, systemic risk, or transparency concern involving Neutron’s hosted Service, email admin@neutronai.dev with the workspace, request or trace identifier if available, a concise description, and no unnecessary personal data or credentials.

Contact and company information

Questions about this document may be sent to admin@neutronai.dev.