Every page
HTML sitemap
Every public page on inoku.com, grouped by type and topic, with a one-line description of each. For people; the XML sitemaps are for crawlers.
The XML sitemap index for crawlers is at /sitemap.xml.
01
Product and company
The pages that say what Inoku builds, for whom, and on what terms.
- HomeInoku is an independent research lab building ways to measure and reduce false success claims by AI agents. Proof, not promises.
- An independent record of what your AI agents didFor teams running agents on real systems: a record of agent actions kept outside the agent's reach, for review and audit. Research stage.
- Design partnersApply to the design-partner programme: early evaluations of your agent's completion claims, free during the programme, for teams running real agents.
- Brand and press kitInoku's logo files, core colours, typefaces and usage rules, with downloads for press and partners. The dotted-zero wordmark and the frame symbol.
- Contact InokuContact Inoku about design partnerships, private evaluations, the enterprise record, press, research collaboration or security reports.
- Custom task environments for AI agentsCustom task suites for training and evaluating agents, checked by an independent ground truth, including tasks that can't be completed.
- SitemapEvery public page on inoku.com, grouped by type and topic, with a one-line description of each. For people; the XML sitemaps are for crawlers.
- Inoku VerifiedAn independent evaluation of how often an AI agent's completion claims are true, on a private test suite. In development; design partners first.
- BetaApply to test the first beta of Inoku's evaluation tools on your own AI agent and tasks. Free, invites in waves, for teams running agents on real systems.
- PricingFree for research, education, open source and organisations under 10 people and $1M revenue. Paid above that, on request during the design-partner phase.
- Private evaluations for AI agents and modelsTest how truthfully your AI model reports its own work, on private tasks it has never seen, including tasks that can't be done.
- Security and responsible disclosureHow to report a security problem in anything Inoku runs or publishes, what to expect from us, and where our security.txt lives.
02
Solutions and comparisons
Who Inoku works with: AI labs and model builders, teams building AI agents, and regulated teams deploying agents on real systems.
- SolutionsWho Inoku works with: AI labs and model builders, teams building AI agents, and regulated teams deploying agents on real systems.
- For AI labs and model buildersContamination, self-grading credibility and RL data that teaches nothing. Independent evaluations and environments for labs building agentic models.
- For teams building AI agentsCustomers lose trust after a false "done". Independent evidence of how often your agent's reports are true, for coding, operations and data agents.
- For regulated teams deploying AI agentsAudit trails the agent can't touch, EU AI Act timelines and what an independent record of agent actions means for review and audit.
- Inoku vs LLM-as-judgeAn LLM judge reads the agent's own transcript. Where that works, where it fails (AUROC 0.65 at best) and when you need an independent record.
03
Research notes
Inoku's blog: evidence-first notes on false success in AI agents, agent verification, reinforcement learning for agents, sandboxing and compliance.
- Research notesInoku's blog: evidence-first notes on false success in AI agents, agent verification, reinforcement learning for agents, sandboxing and compliance.
- Agent benchmarksWhat an AI agent benchmark measures, how contamination and exploitable tests inflate scores, and what a leaderboard must state before you trust it.
- Agent complianceWhat the EU AI Act's logging duties mean for AI agents: Article 12, the AI Omnibus dates, certification, and what an auditor can trust.
- Agent sandboxingContainers, microVMs and gVisor for AI agents, why identical images don't give identical runs, and a practical Incus setup for an agent lab.
- Agent verificationOutcome checks, process evidence, LLM judges and independent records: a guide to verifying AI agent actions, with public sources only.
- False successWhat false success in AI agents is, how common it is, why judges miss it, and how to measure an agent's false-completion rate. A guide with sources.
- RL for agentsRLVR, GRPO, the zero-advantage problem, process rewards and reward-hacking defences: a guide to reinforcement learning for LLM agents.
04
Glossary
Plain definitions of the terms used in AI agent evaluation: false success, upward deception, reward hacking, RLVR, LLM-as-judge and held-out tests.
- GlossaryPlain definitions of the terms used in AI agent evaluation: false success, upward deception, reward hacking, RLVR, LLM-as-judge and held-out tests.
- Agent traceAn agent trace is the recorded sequence of an AI agent's steps, tool calls and outputs during a task.
- Benchmark contaminationBenchmark contamination is when test tasks or their answers appear in a model's training data.
- False successFalse success is when an AI agent reports a task complete although it was not.
- GRPOGroup Relative Policy Optimization (GRPO) is a reinforcement learning algorithm that scores each sample against the mean of a group of samples for the same prompt.
- Held-out evaluationA held-out evaluation tests a model on tasks kept out of its training and development data.
- Honest failureHonest failure, or abstention, is when an AI agent says it could not complete a task, and that is true.
- Impossible taskAn impossible task is a test task that cannot be completed as stated, used to check whether an agent admits it.
- LLM-as-judgeLLM-as-judge is the practice of using a language model to grade another model's output.
- OverclaimingOverclaiming is when an AI agent states more progress or capability than it has evidence for.
- Process reward modelA process reward model scores the intermediate steps of a solution rather than only its final answer.
- Reward hackingReward hacking is when a model gains reward through behaviour the reward was not meant to encourage.
- RLVRReinforcement learning with verifiable rewards (RLVR) trains a model with rewards computed by a checker rather than by human preference.
- Sandbox escapeA sandbox escape is when code running inside an isolated environment gains access outside it.
- Tamper-evident logA tamper-evident log is a record designed so that any later change to it can be detected.
- Upward deceptionUpward deception is when an AI agent hides or misrepresents its failures to the person or system it reports to.
- Zero-advantage problemThe zero-advantage problem is when every sample in a group gets the same reward, so the training update carries no signal.
05
About and legal
Who is behind Inoku, and the terms the site runs under.
- About InokuInoku is an independent research lab working on one question: how do we make AI agents tell the truth about their own work? Founded in 2026.
- Ivan PopovIvan Popov, founder of Inoku: 20+ years in web and systems engineering and founder of the IvanLabs advisory practice. Started Inoku in 2026.
- Press releasesAnnouncements from Inoku, newest first: programmes opening, publications and milestones. Each release states what is built and what is not.
- Case studiesCase studies from work with design partners and beta testers. None yet: the programmes have just opened, and nothing is published without permission.
- LegalInoku's legal pages: the privacy policy for inoku.com, the terms of use for the site and its brand files, and the legal notice naming the operator.
- Privacy policyWhat inoku.com stores about visitors (server logs and one cookie for your cookie choice), what happens to messages you send us, and your GDPR rights.
- Terms of useThe terms under which inoku.com is published: what you may do with the content and the brand files, what we don't promise, and which law applies.
- Legal noticeWho operates inoku.com: the operator's legal name and form, address, contact email and registration details, as required for an EU-based website.