Evidence & Methodology

Why methodology is our differentiator

Most AI tools optimise for a fluent answer. For lawyers, accountants, and the professionals KLNR serves, a fluent wrong answer is worse than no answer at all. Our differentiator is not a louder model — it is a stricter method. We call it evidence, not words.

This page describes how KLNR products (Search, Lexor, FRRE.ai, CoLab, AgentOS, and others) are designed to produce answers you can defend: grounded in citable sources, dated, honest about uncertainty, and always under the control of a qualified human. It is written to be candid about both what our method does well and where it can fail.

This is a working draft describing our design intent and operating principles; it is not a warranty of outcomes. Professional users remain responsible for their own work product.

Evidence-first answering

Every substantive claim our systems make is meant to be traceable to a source you can inspect:

• Grounding in cited sources: Answers are constructed from retrieved, identifiable sources — statutes, case law, filings, the customer's own documents — and present those sources as citations rather than asserting facts from the model's memory alone.
• Evidence floor: Where a product is configured with an evidence threshold, a claim that cannot be supported by sufficient retrieved evidence is not asserted as fact. The system surfaces what it found, and what it could not.
• as_of date discipline: Legal and regulatory material changes over time. Our systems track and display the as_of date of the sources relied upon, so you know the temporal validity of an answer. An answer is only as current as the corpus and date it was grounded in, and we make that explicit.
• Inspectable evidence: Wherever feasible, you can open the underlying source behind a statement, rather than being asked to take the system's word for it.
• Separation of retrieval and generation: We separate the act of finding evidence from the act of phrasing an answer, which reduces the risk of fabricated citations and lets the evidence constrain the language — not the other way around.

The abstention policy

The hardest and most important design decision in a professional AI system is knowing when to say nothing.

• Silence over guessing: When the system lacks sufficient evidence to support a claim, it is designed to abstain — to state that it cannot answer, or to answer only the part it can support — rather than fabricate a confident-sounding response.
• Abstention is a feature, not a failure: An honest "I don't have evidence for that" preserves your trust and your professional standing. We treat unjustified confidence as the more serious error.
• Scoped answers: Where the system can support part of a question but not all of it, it answers what it can and clearly flags what it cannot.
• No invented authority: The system does not manufacture case citations, statutory references, or quotations to fill a gap. If it cannot find supporting authority, it says so.
• Routing to a human: Abstention is paired with handoff — the question is returned to the professional, who can investigate, decide, and, where appropriate, supply the missing input.

Human-in-the-loop: AI assists, the professional authors and decides

KLNR's model of work is explicit and non-negotiable: AI is the assistant and second pilot; a qualified professional is the author, the decision-maker, and the signatory.

• Assistant, not replacement: Our products research, draft, organise, and propose. They do not practise law or accountancy, and they do not replace the professional's judgement.
• The professional authors and signs: The qualified human reviews, edits, decides, and takes responsibility for the final work product. In Sign, the human signature — not the AI — is the binding act.
• Consequential actions require human approval: In AgentOS and similar products, high-impact actions are gated behind human confirmation. An agent cannot exceed the authority of the person it serves.
• Professional secrecy preserved: The human-in-the-loop model is also a confidentiality model. Attorney–client privilege and professional secrecy are respected; we do not train models on client or legal content.
• Transparency about AI involvement: We disclose where AI is used so that professionals — and, where relevant, their clients — understand the role it played.

Known and fixed failure modes

Honesty about limitations is part of the method. Like any AI system, KLNR's products have failure modes. We document the ones we know about, monitor for them, and fix the corpus, prompts, or guardrails when they recur:

• Stale corpus: An answer can be outdated if the underlying source material has changed since the as_of date. Mitigation: as_of disclosure and corpus refresh.
• Retrieval gaps: If relevant authority is not in the indexed corpus, the system may abstain or under-answer rather than reach beyond its evidence.
• Misgrounding: A retrieved source may be relevant but mis-applied. Mitigation: human review and citation inspection.
• Prompt injection from untrusted content: Documents and emails may contain adversarial instructions. Mitigation: treating retrieved content as data, not commands (see Trust & Security).
• Over- or under-abstention: The system may occasionally stay silent when it could have answered, or answer when it should have abstained. We tune these thresholds and measure them (below).

We will not claim these failure modes do not exist. We claim that we surface them, measure them, and improve.

Our metrics commitment

A method that cannot be measured cannot be trusted. We commit to evaluating our products against held-out test sets and to publishing meaningful metrics over time, including:

• Accuracy / groundedness: the rate at which claims are correctly supported by their cited sources.
• Abstention quality: how often the system correctly abstains when it lacks evidence, versus over- or under-abstaining.
• Citation integrity: the rate of valid, resolvable citations versus fabricated or broken ones.
• Known failure-mode tracking: a maintained, public-facing list of recognised failure modes and their status.

We will report these honestly — including where we fall short of our targets — because the alternative is the very over-confidence our method exists to prevent. Specific figures and an evaluation methodology will be published here [metrics to be published — target Q[...] 20[...]].

Scope, limitations & no legal/professional advice

KLNR products are professional tools, not a professional. Outputs are drafts, research aids, and proposals intended to be reviewed by a qualified professional. Nothing produced by a KLNR product constitutes legal, tax, accounting, or other professional advice from KLNR, and KLNR does not establish a lawyer–client or comparable relationship with end users by virtue of providing the software. The professional user is responsible for verifying outputs and for the final work product.

Regional supplement — European Union (GDPR, EU AI Act & EU consumer law)

For users and processing in the EU/EEA:

• Human oversight & transparency (EU AI Act): Our human-in-the-loop and disclosure practices are designed to align with the human-oversight and transparency expectations of Regulation (EU) 2024/1689 (EU AI Act). Where any KLNR functionality falls within a regulated risk category, we will provide the corresponding documentation and safeguards (see Trust & Security — EU AI Act readiness).
• No training on personal/client data: Consistent with GDPR principles of purpose limitation and data minimisation, we do not use client or legal content to train foundation models.
• Automated processing: Because a qualified human authors and decides, KLNR products are not designed to make solely automated decisions producing legal or similarly significant effects under GDPR Article 22 on the user's clients.
• Consumer transparency: For consumer-facing use, we provide clear information about the role and limitations of AI, consistent with EU consumer-protection and fairness rules.
• Controller: KLNR Labs P.S.A., [address], Gdańsk, Poland ([KRS], [NIP], [REGON]). Contact [privacy@klnr.ai] / [dpo@klnr.ai].

Regional supplement — United States (CCPA/CPRA & other state laws)

For users and processing subject to US state law:

• No sale or sharing of personal information: We do not sell or share personal information, and we do not use client content to train foundation models.
• Automated decision-making & profiling: Where emerging state rules (e.g., under the CCPA/CPRA, Colorado CPA, and similar) address automated decision-making and profiling, our human-in-the-loop model is designed so that consequential decisions remain with the qualified professional, supporting applicable opt-out and transparency obligations.
• Transparency: We disclose AI involvement and the limitations described above, consistent with US consumer-protection principles against unfair or deceptive practices.
• Contact: US-related requests may be directed to [privacy@klnr.ai].

Changes & contact

We will update this page as our methodology, metrics, and the legal landscape evolve. Effective date: [effective date]. Questions: [kontakt@klnr.ai] · Privacy/DPO: [privacy@klnr.ai] / [dpo@klnr.ai].