1.Purpose of this disclosureLink to section 1

LiteSurface uses large language models from third-party providers to research, draft, and evaluate product concepts. This page explains where models are used, what they are not allowed to do, how their outputs are checked, and the limits you should keep in mind. It complements our Terms of Service and Privacy Policy.

2.Where AI is usedLink to section 2

AI tasks in the pipeline
Stage	What the model does	What code does
Research planning	Proposes search queries for a research focus	Runs searches through a fixed tool, triages results for relevance, and fetches allowed pages
Claim extraction	Extracts atomic claims from source text, each tied to a quoted passage	Stores claims with their source and content hash; validates structure
Concept generation	Drafts product concepts from capabilities, evidence, and your strategy profile	Enforces schemas, deduplicates, and records lineage and versions
Evaluation	Two evaluators from different providers score each criterion with a rationale and cited claims	Combines the scores into the final result (see Section 4)
Evidence-gap research	Turns evaluators’ stated evidence gaps into targeted questions	Decides when to run the loop, then re-evaluates and appends a new evaluation
Synthesis and artifacts	Summarizes results and drafts documents such as validation plans and build handoffs	Assembles sections, validates structure, and records the prompt version used
Retrieval	Creates text embeddings so relevant claims can be found	Ranks claims by relevance and decision value

Every model output is validated against a schema before it is saved; an output that cannot be repaired is rejected rather than stored. Each call records the provider, model, prompt version, and run it belongs to, so any result can be traced.

3.What models do not doLink to section 3

They do not decide. People shortlist, validate, build, hold, or reject concepts. Each decision is recorded with the AI recommendation it followed or overrode.
They do not compute the final score. Deterministic code does, using the project’s scorecard.
They do not browse freely. Search and fetching happen through typed tools that workflow code authorizes; each task has a fixed allowlist, and claim extraction has no external tools at all.
They do not act outside the Service. Models cannot send email, make purchases, run code on your systems, or call arbitrary web addresses.
They do not see secrets or other workspaces. API keys are never placed in model context, and context is assembled only from the project being worked on.
They do not see fields you exclude. Fields classified as sensitive and not for models are removed before prompts are assembled.

4.Dual evaluation and deterministic scoringLink to section 4

Each concept is scored by two independent evaluators, by default one from OpenAI and one from Anthropic. Using different providers reduces the chance that one model’s blind spots decide the outcome, and their disagreement is shown rather than hidden.

Code then turns the evaluations into a result:

takes the median of the evaluators’ scores for each criterion;
applies the scorecard’s published weights to produce a weighted total;
applies gates, such as build horizon, capability availability, project exclusions, and legal or safety review, which override the numeric band when they fail;
reports confidence, evaluator disagreement, and evidence coverage separately, so none of them inflates the score.

The weights, gates, and bands are documented in our scoring methodology. Evaluations are never rewritten; re-evaluation appends a new record.

5.Citations and evidence coverageLink to section 5

Evaluators must cite the claims that support their material points, and each claim links back to the passage it came from. The Service checks that cited claims exist in the context the evaluator was given.

Evidence coverage measures how many material points are supported by cited evidence rather than by the model’s general knowledge. When coverage is low, or evaluators rate the evidence as weak, the concept is marked as needing evidence: its recommendation is capped and its confidence lowered, and you can gather targeted evidence and re-evaluate.

6.Limitations you should knowLink to section 6

Models can be wrong. They may state false information confidently (sometimes called hallucination), misread a source, or draw an unsupported conclusion.
Sources can be wrong. A correctly cited claim is only as reliable as the page it came from, and web content can be outdated, biased, or promotional.
Coverage is not truth. High evidence coverage means claims are cited, not that the market will behave as they suggest.
Results vary. The same inputs can produce different outputs across runs, models, and model versions.
Knowledge has limits. Models have training cut-offs and may lack context about your market, customers, or constraints.
Bias is possible. Models can reflect biases in their training data, which may shape which ideas are proposed or favored.

Use scores as decision aids

A score is a structured, inspectable opinion, not a prediction or a guarantee of commercial, legal, or technical success. Review the evidence and rationale behind any recommendation before acting on it.

7.Human decision-makingLink to section 7

The Service is designed to support people, not replace them. Recommendations are proposals that a person accepts or overrides, with the rationale recorded. Concepts in regulated or high-risk areas can be flagged for legal or safety review, and an AI score is never legal, safety, or regulatory approval. The Service is not intended for decisions about individuals, as our Acceptable Use Policy explains.

8.Prompt injection and retrieved contentLink to section 8

Web pages and documents can contain text designed to manipulate an AI model. We treat all retrieved content as untrusted data, never as instructions:

a fixed trust order places system policies and our reviewed prompts above your requests, your requests above project data, and retrieved content below all of them;
retrieved text is passed to models inside clearly delimited data sections, with an explicit statement that instructions inside it must be ignored;
tool use is authorized by workflow code, not by model text, so a page cannot trigger a new search, fetch, or action;
the fetcher refuses private and internal network addresses and revalidates redirects; and
we test these defenses with a prompt-injection evaluation suite.

These measures reduce the risk but cannot eliminate it. If an output appears to follow instructions from a source, please report it.

9.Your data and model trainingLink to section 9

LiteSurface does not use your content or outputs to train or fine-tune AI models. The Service sends providers only what each task needs, only to the providers your workspace allows, and after removing fields marked not for models. We record metadata about each call, including a hash of the input, but do not write raw prompts or responses to our logs by default.

Providers process data under their own terms. Both OpenAI and Anthropic state that, by default, they do not use data sent through their commercial APIs to train their models; their current terms and retention practices govern. When you use your own keys, your agreement with the provider applies. See Subprocessors for details.

10.Fixture modeLink to section 10

In fixture mode, the Service returns prepared sample outputs instead of calling AI or search providers, so you can explore the workflow without model costs. Fixture outputs are illustrative only and do not analyze your market; the Service indicates when it is running in fixture mode.

11.Model and provider changesLink to section 11

Application code refers to task roles rather than specific models, and we choose the models behind each role, with a fallback provider for resilience. We may change models as providers release new versions or retire old ones, and we note significant changes in the changelog. Your workspace’s allowed-provider setting always limits which providers are used.

12.Reporting a problemLink to section 12

Inaccurate, harmful, or biased output: support@litesurface.com
Suspected prompt injection or other security issue: security@litesurface.com
Content that infringes your rights: legal@litesurface.com

Include the project, the output, and what you expected if you can. We review every report, and we update this disclosure as the Service and the law evolve, including rules such as the EU AI Act.

Questions about this document

Write to legal@litesurface.com for legal questions or privacy@litesurface.com for privacy requests. Postal notices go to LiteSurface, Inc., [Registered address to be confirmed].