Glossary
Glossary
One sentence first, then the detail, a simple example, and where to read more. Each term links to the notes that use it.
Sixteen terms today, listed alphabetically. Each entry gives the definition, a plain explanation, a simple hypothetical example, why it matters for teams building or auditing agents, how the research community measures or handles it, related terms, and external reading.
A
Agent trace
An agent trace is the recorded sequence of an AI agent's steps, tool calls and outputs during a task.
Full entryB
Benchmark contamination
Benchmark contamination is when test tasks or their answers appear in a model's training data.
Full entryF
False success
False success is when an AI agent reports a task complete although it was not.
Full entryG
GRPO
Group 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.
Full entryH
Held-out evaluation
A held-out evaluation tests a model on tasks kept out of its training and development data.
Full entryH
Honest failure
Honest failure, or abstention, is when an AI agent says it could not complete a task, and that is true.
Full entryI
Impossible task
An impossible task is a test task that cannot be completed as stated, used to check whether an agent admits it.
Full entryL
LLM-as-judge
LLM-as-judge is the practice of using a language model to grade another model's output.
Full entryO
Overclaiming
Overclaiming is when an AI agent states more progress or capability than it has evidence for.
Full entryP
Process reward model
A process reward model scores the intermediate steps of a solution rather than only its final answer.
Full entryR
Reward hacking
Reward hacking is when a model gains reward through behaviour the reward was not meant to encourage.
Full entryR
RLVR
Reinforcement learning with verifiable rewards (RLVR) trains a model with rewards computed by a checker rather than by human preference.
Full entryS
Sandbox escape
A sandbox escape is when code running inside an isolated environment gains access outside it.
Full entryT
Tamper-evident log
A tamper-evident log is a record designed so that any later change to it can be detected.
Full entryU
Upward deception
Upward deception is when an AI agent hides or misrepresents its failures to the person or system it reports to.
Full entryZ
Zero-advantage problem
The zero-advantage problem is when every sample in a group gets the same reward, so the training update carries no signal.
Full entry
Questions
Frequently asked questions
- What is this glossary for?
- It defines the terms used across inoku.com in one place: a quotable one-sentence definition, a plain explanation, a hypothetical example, and links to the research notes that use the term and to external resources.
- Are the definitions Inoku's own?
- The wording is ours, but each term is grounded in public research listed on the term’s page. Where a term has a standard meaning in the literature, we follow it; where usage varies, the page says so.
- Why do some terms have numbers in them?
- Where a published study puts a figure on the behaviour, such as the share of failures that were false success, the page quotes it and cites the source. Numbers never come from Inoku’s own unpublished work.
- Can I link to or quote a definition?
- Yes. Each term has a stable URL and a one-sentence definition written to be quoted. Attribution to inoku.com is appreciated. The brand page has the logo files if you need them.
- Which terms are missing?
- Terms specific to Inoku’s own method are not published yet; they appear when the method is published. Suggestions for public terms are welcome through the contact page.
- How do the glossary and the guides fit together?
- Each guide, such as false success in AI agents, explains a whole topic. The glossary gives the precise meaning of one term and points back to the guides and notes that use it.