Skip to content

What we test

95 test areas across the agent and its harness, grouped by the six questions buyers ask. Each shows its status through your public chat and with system access (a staging environment, tool sandbox, traces or configuration).

36 available now · 48 pilot · 11 roadmap. No evaluation is exhaustive; this map shows what we cover and how far.

Available now
Built into the TrendLab evaluation engine and used in sealed runs. Also used for services delivered by TrendLab from a written runbook with completed deliveries on record.
Pilot
Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.
Roadmap
Not deliverable yet. Never quoted or included in a statement of work.

A status describes the capability. It reaches you through a service, and that service has its own label: an engine capability marked available now may be delivered through a service that is still a pilot.

Task success

Does the task get completed correctly?

Whether the agent finishes the job the right way: correct action, correct target, hand-off when a person is needed, and verified outcomes rather than claimed ones.

  • Tool selection, arguments and response handling

    Does it use tools correctly?

    Not applicableWith access:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.
    Agent Audit, Eval Setup

  • Task success verified against final system state (correct action, correct target, exactly-once outcome, resolution)

    Did the job actually get done correctly?

    Not applicableWith access:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.
    Agent Audit, Model & Harness Comparison, Eval Setup

  • Multi-turn manipulation

    Can it be worn down over several messages?

    Public chat:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.With access:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.
    Agent Audit, Monitoring (recurring re-evaluation), Red-Team Sprint

  • Context retention within a conversation

    Does it remember what was said?

    Public chat:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.With access:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.
    Agent Audit

  • Hand-off to a human : works, offered when needed, not overused

    Can customers reach a person, and only when they need one?

    Public chat:Available now: Built into the TrendLab evaluation engine and used in sealed runs.With access:Available now: Built into the TrendLab evaluation engine and used in sealed runs.
    Free Risk Snapshot, Agent Audit, Monitoring (recurring re-evaluation)

  • Over-refusal

    Is it so strict it's useless?

    Public chat:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.With access:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.
    Agent Audit

  • Empathy and tone

    Does it handle upset customers well?

    Public chat:Available now: Built into the TrendLab evaluation engine and used in sealed runs.With access:Available now: Built into the TrendLab evaluation engine and used in sealed runs.
    Free Risk Snapshot, Agent Audit

  • Abuse and harassment handling

    What happens with hostile users?

    Public chat:Available now: Built into the TrendLab evaluation engine and used in sealed runs.With access:Available now: Built into the TrendLab evaluation engine and used in sealed runs.
    Free Risk Snapshot, Agent Audit

  • Long-running tasks (persistent memory, context compression, checkpoints, resumption, preserving constraints)

    Does it keep the rules over long or resumed work?

    Not applicableWith access:Roadmap: Not deliverable yet. Never quoted or included in a statement of work.
    Agent Audit

  • Clarifying questions and escalation judgment under ambiguity

    Does it ask when the request is unclear, and escalate at the right time?

    Public chat:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.With access:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.
    Agent Audit

  • Routing (sales / support / human)

    Do conversations go to the right place?

    Public chat:Available now: Built into the TrendLab evaluation engine and used in sealed runs.With access:Available now: Built into the TrendLab evaluation engine and used in sealed runs.
    Free Risk Snapshot, Agent Audit

Safe actions and protected data

Does the system act within its authority and protect data?

Prompt injection, leaked instructions, other customers' data, unauthorized actions, approvals that bind to what actually happens, and the limits and stop controls around it.

  • Direct prompt injection

    Can one message hijack it?

    Public chat:Available now: Built into the TrendLab evaluation engine and used in sealed runs.With access:Available now: Built into the TrendLab evaluation engine and used in sealed runs.
    Free Risk Snapshot, Agent Audit, Monitoring (recurring re-evaluation), Red-Team Sprint

  • Indirect injection

    Can content it reads hijack it?

    Public chat:Available now: Built into the TrendLab evaluation engine and used in sealed runs.With access:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.
    Free Risk Snapshot, Agent Audit, Red-Team Sprint

  • System-prompt and configuration leakage

    Can people see our instructions?

    Public chat:Available now: Built into the TrendLab evaluation engine and used in sealed runs.With access:Available now: Built into the TrendLab evaluation engine and used in sealed runs.
    Free Risk Snapshot, Agent Audit, Monitoring (recurring re-evaluation), Red-Team Sprint

  • Other customers' data in replies

    Can it reveal someone else's data?

    Public chat:Available now: Built into the TrendLab evaluation engine and used in sealed runs.With access:Available now: Built into the TrendLab evaluation engine and used in sealed runs.
    Free Risk Snapshot, Agent Audit, Red-Team Sprint

  • Unsafe links, markdown-image exfiltration

    Can it leak data through links?

    Public chat:Available now: Built into the TrendLab evaluation engine and used in sealed runs.With access:Available now: Built into the TrendLab evaluation engine and used in sealed runs.
    Free Risk Snapshot, Agent Audit

  • Jailbreaks and harmful output

    Can it be made to say awful things?

    Public chat:Available now: Built into the TrendLab evaluation engine and used in sealed runs.With access:Available now: Built into the TrendLab evaluation engine and used in sealed runs.
    Free Risk Snapshot, Agent Audit, Monitoring (recurring re-evaluation), Red-Team Sprint

  • Secrets in output

    Can it leak keys?

    Public chat:Available now: Built into the TrendLab evaluation engine and used in sealed runs.With access:Available now: Built into the TrendLab evaluation engine and used in sealed runs.
    Agent Audit

  • Unsafe data collection

    Does it ask for passwords or card numbers?

    Public chat:Available now: Built into the TrendLab evaluation engine and used in sealed runs.With access:Available now: Built into the TrendLab evaluation engine and used in sealed runs.
    Free Risk Snapshot, Agent Audit

  • Instruction precedence and configuration correctness

    Does it follow the right instructions when sources conflict?

    Not applicableWith access:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.
    Agent Audit, Red-Team Sprint

  • Deletion propagation in the client's agent system

    When data is deleted, is it gone from memory, indexes, caches and logs?

    Not applicableWith access:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.
    Agent Audit

  • Cross-tenant isolation beyond replies

    Can one customer's data reach another through storage, caches, files, retrieval or jobs?

    Not applicableWith access:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.
    Agent Audit, Red-Team Sprint

  • Unauthorized actions

    Can it do something it shouldn't?

    Not applicableWith access:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.
    Agent Audit, Red-Team Sprint

  • Identity verification before acting

    Does it check who it's talking to?

    Not applicableWith access:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.
    Agent Audit

  • Excessive agency and permissions (design review)

    Does it have more power than it needs?

    Not applicableWith access:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.
    Agent Audit

  • Human approval integrity : confirmation before irreversible steps, truthful approval context, action bound to the approval, re-approval when the action changes

    When a person approves, are they approving what actually happens?

    Not applicableWith access:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.
    Agent Audit

  • Multi-agent delegation : delegated authority, identity propagation, compromised peers, cascading failure

    Do our agents trust each other too much?

    Not applicableWith access:Roadmap: Not deliverable yet. Never quoted or included in a statement of work.
    Agent Audit

  • Enforced autonomy limits (spending, execution budgets, permissions at runtime, cancellation, emergency stop)

    Do our limits and stop buttons actually stop it?

    Not applicableWith access:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.
    Agent Audit, Red-Team Sprint

  • Tool, plugin, skill and MCP supply chain (provenance, installation, updates, compromise, changed permissions, drift)

    Can a tool we installed turn against us?

    Not applicableWith access:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.
    Agent Audit, Red-Team Sprint

  • Multimodal inputs (images, screenshots, PDFs, attachments, mixed input, instruction smuggling). Modality is the input type; channel is the transport

    Can a picture or PDF trick it?

    Public chat:Roadmap: Not deliverable yet. Never quoted or included in a statement of work.With access:Roadmap: Not deliverable yet. Never quoted or included in a statement of work.
    Agent Audit, Red-Team Sprint

Trustworthy information

Can we trust its information?

Invented refunds and discounts, wrong prices, answers that contradict your published policies, hallucinated facts and grounding in your own documents.

  • Unauthorized commitments

    Can my bot promise something we don't offer?

    Public chat:Available now: Built into the TrendLab evaluation engine and used in sealed runs.With access:Available now: Built into the TrendLab evaluation engine and used in sealed runs.
    Free Risk Snapshot, Snapshot Plus (paid findings review), Agent Audit, Monitoring (recurring re-evaluation), Vendor Due Diligence

  • Invented discounts, coupons, price matches

    Can it give away margin?

    Public chat:Available now: Built into the TrendLab evaluation engine and used in sealed runs.With access:Available now: Built into the TrendLab evaluation engine and used in sealed runs.
    Free Risk Snapshot, Snapshot Plus (paid findings review), Agent Audit, Monitoring (recurring re-evaluation), Vendor Due Diligence

  • Contradicting published policy

    Does it say the same thing as our Terms?

    Public chat:Available now: Built into the TrendLab evaluation engine and used in sealed runs.With access:Available now: Built into the TrendLab evaluation engine and used in sealed runs.
    Free Risk Snapshot, Snapshot Plus (paid findings review), Agent Audit, Monitoring (recurring re-evaluation)

  • Wrong prices, plans or terms

    Does it quote the right price?

    Public chat:Available now: Built into the TrendLab evaluation engine and used in sealed runs.With access:Available now: Built into the TrendLab evaluation engine and used in sealed runs.
    Free Risk Snapshot, Snapshot Plus (paid findings review), Agent Audit, Monitoring (recurring re-evaluation)

  • Accepting forged "earlier agreements"

    Can a customer fake a promise?

    Public chat:Available now: Built into the TrendLab evaluation engine and used in sealed runs.With access:Available now: Built into the TrendLab evaluation engine and used in sealed runs.
    Free Risk Snapshot, Agent Audit

  • Invented facts against facts supplied to the test

    Does it make things up?

    Public chat:Available now: Built into the TrendLab evaluation engine and used in sealed runs.With access:Available now: Built into the TrendLab evaluation engine and used in sealed runs.
    Free Risk Snapshot, Agent Audit, Monitoring (recurring re-evaluation)

  • Faithfulness to the client knowledge base

    Does it stick to our docs?

    Public chat:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.With access:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.
    Agent Audit, Monitoring (recurring re-evaluation), Eval Setup

  • Self-contradiction within a reply

    Does it contradict itself?

    Public chat:Available now: Built into the TrendLab evaluation engine and used in sealed runs.With access:Available now: Built into the TrendLab evaluation engine and used in sealed runs.
    Free Risk Snapshot, Agent Audit

  • Confident wrongness and stated uncertainty

    Does it signal when it isn't sure?

    Public chat:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.With access:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.
    Agent Audit

  • Retrieval quality

    Is the search behind it working?

    Not applicableWith access:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.
    Agent Audit, Eval Setup

  • Competitor handling

    What does it say about competitors?

    Public chat:Available now: Built into the TrendLab evaluation engine and used in sealed runs.With access:Available now: Built into the TrendLab evaluation engine and used in sealed runs.
    Free Risk Snapshot, Agent Audit

  • Multilingual behavior

    Same rules in every language?

    Public chat:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.With access:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.
    Agent Audit

  • Unequal treatment across customer groups

    Does it treat customers differently?

    Public chat:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.With access:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.
    Agent Audit

  • Sensitive topics

    What if a user raises something serious?

    Public chat:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.With access:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.
    Agent Audit

Resilient operation

Does it recover when something breaks?

Whether the bot answers at all, latency, fallback when a model or tool fails, partial failures, duplicate side effects and dates or events handled out of order.

  • Availability (widget loads, reply arrives)

    Does it actually answer?

    Public chat:Available now: Built into the TrendLab evaluation engine and used in sealed runs.With access:Available now: Built into the TrendLab evaluation engine and used in sealed runs.
    Free Risk Snapshot, Agent Audit, Monitoring (recurring re-evaluation)

  • Latency

    Is it fast enough?

    Public chat:Available now: Built into the TrendLab evaluation engine and used in sealed runs.With access:Available now: Built into the TrendLab evaluation engine and used in sealed runs.
    Agent Audit, Monitoring (recurring re-evaluation)

  • Error and fallback behavior

    What happens when the model is down?

    Not applicableWith access:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.
    Agent Audit

  • Recovery after partial failure (retries, duplicate side effects, rollback or compensation, conflicting changes, concurrency)

    If something fails halfway, does it clean up or double-charge?

    Not applicableWith access:Roadmap: Not deliverable yet. Never quoted or included in a statement of work.
    Agent Audit

  • Load, throughput, rate limits, resource exhaustion

    Does it hold up under volume?

    Not applicableWith access:Roadmap: Not deliverable yet. Never quoted or included in a statement of work.
    Agent Audit

  • Temporal and event correctness (time zones, daylight saving, expiring approvals, delayed or out-of-order events)

    Does it get dates, deadlines and event order right?

    Public chat:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.With access:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.
    Agent Audit

  • Voice and phone channel

    Can you test our phone bot?

    Public chat:Roadmap: Not deliverable yet. Never quoted or included in a statement of work.With access:Roadmap: Not deliverable yet. Never quoted or included in a statement of work.
    Agent Audit

  • Messaging channels (WhatsApp, SMS, social DMs)

    Can you test our WhatsApp bot?

    Public chat:Roadmap: Not deliverable yet. Never quoted or included in a statement of work.With access:Roadmap: Not deliverable yet. Never quoted or included in a statement of work.
    Agent Audit

  • Chat accessibility

    Can everyone use it?

    Public chat:Roadmap: Not deliverable yet. Never quoted or included in a statement of work.With access:Roadmap: Not deliverable yet. Never quoted or included in a statement of work.
    Agent Audit

Controlled cost

What does a successful outcome cost?

Cost per successfully completed task, including failed attempts and retries, off-topic use of your model budget, and whether a cheaper model holds up on your tests.

  • Off-scope compute abuse

    Is it used as a free general AI?

    Public chat:Available now: Built into the TrendLab evaluation engine and used in sealed runs.With access:Available now: Built into the TrendLab evaluation engine and used in sealed runs.
    Free Risk Snapshot, Agent Audit, Monitoring (recurring re-evaluation)

  • Cost per successfully completed task, including failed attempts and retries

    What does each successful outcome really cost?

    Not applicableWith access:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.
    Agent Audit, Model & Harness Comparison

  • Harness vs. model comparison, including off-the-shelf harnesses compared with the client's setup

    Is our setup helping or hurting, and which harness should we standardize on?

    Not applicableWith access:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.
    Model & Harness Comparison, Eval Setup, Harness Configuration & Cost-Down

Reliable change

Will it remain reliable after changes?

Re-tests after prompt, model, tool or knowledge-base changes, fixed failures that stay fixed, consistency across repeated trials and whether what was tested matches what runs.

  • Policy drift after site or pricing changes

    Did our bot keep up with new pricing?

    Public chat:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.With access:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.
    Monitoring (recurring re-evaluation)

  • Stale information

    Does it know what changed?

    Public chat:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.With access:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.
    Agent Audit, Monitoring (recurring re-evaluation)

  • Prompt-change regression

    Did fixing one thing break another?

    Public chat:Available now: Built into the TrendLab evaluation engine and used in sealed runs.With access:Available now: Built into the TrendLab evaluation engine and used in sealed runs.
    Monitoring (recurring re-evaluation), Eval Setup

  • Model upgrade and provider migration

    Is the new model ready to switch to?

    Not applicableWith access:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.
    Model & Harness Comparison, Monitoring (recurring re-evaluation)

  • Provider-reported model-version changes

    Did the provider change the model under us?

    Not applicableWith access:Available now: Built into the TrendLab evaluation engine and used in sealed runs.
    Monitoring (recurring re-evaluation)

  • Knowledge-base or tool changes

    Did updating our docs change the bot?

    Public chat:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.With access:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.
    Monitoring (recurring re-evaluation)

  • Fixed failures staying fixed

    Do bugs come back?

    Public chat:Available now: Built into the TrendLab evaluation engine and used in sealed runs.With access:Available now: Built into the TrendLab evaluation engine and used in sealed runs.
    Monitoring (recurring re-evaluation)

  • Consistency across repeated trials

    Does it pass every time?

    Public chat:Available now: Built into the TrendLab evaluation engine and used in sealed runs.With access:Available now: Built into the TrendLab evaluation engine and used in sealed runs.
    Agent Audit, Monitoring (recurring re-evaluation), Model & Harness Comparison

  • Evidence validity lifecycle (material-change triggers, expiry, stale results, withdrawn findings, verify-page status)

    Is this result still current?

    Public chat:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.With access:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.
    Agent Audit, Monitoring (recurring re-evaluation), AI Assurance Program

  • Deployment parity and limits on transferring conclusions

    Does what you tested match what we run?

    Not applicableWith access:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.
    Agent Audit, Model & Harness Comparison

  • CI gate in the client repository

    Can bad changes be blocked automatically?

    Not applicableWith access:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.
    Eval Setup, Monitoring (recurring re-evaluation)

  • Production sampling, including human-correction effort from client data

    Is it working in production, and how much do people fix?

    Not applicableWith access:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.
    Monitoring (recurring re-evaluation), Eval Setup

  • Incident reproduction and variant testing

    How bad is it, and is it fixed?

    Public chat:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.With access:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.
    Incident Response

  • Post-incident regression cases

    Will it happen again?

    Public chat:Available now: Built into the TrendLab evaluation engine and used in sealed runs.With access:Available now: Built into the TrendLab evaluation engine and used in sealed runs.
    Incident Response, Monitoring (recurring re-evaluation)

Supporting

How you can trust the results

Human review, judge calibration, held-back tests, tamper-evident records, framework mapping and the limits of each.

  • Human review, disputes and corrections : review queue, second review where available, adjudication conflicts, corrections, report version history

    Does a person check the grades, and can we dispute them?

    Public chat:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.With access:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.
    All evaluation services

  • Calibration of TrendLab's judges against human labels

    How do we know your judges are right?

    Public chat:Roadmap: Not deliverable yet. Never quoted or included in a statement of work.With access:Roadmap: Not deliverable yet. Never quoted or included in a statement of work.
    All evaluation services

  • Held-out tests withheld from deliverables

    Can fixes be aimed at your tests?

    Public chat:Available now: Built into the TrendLab evaluation engine and used in sealed runs.With access:Available now: Built into the TrendLab evaluation engine and used in sealed runs.
    Free Risk Snapshot, Agent Audit, Monitoring (recurring re-evaluation)

  • Confidence rating and stated limits

    How much can we rely on this?

    Public chat:Available now: Built into the TrendLab evaluation engine and used in sealed runs.With access:Available now: Built into the TrendLab evaluation engine and used in sealed runs.
    All evaluation services

  • Replay of stored inputs

    Can you re-send the exact inputs?

    Public chat:Available now: Built into the TrendLab evaluation engine and used in sealed runs.With access:Available now: Built into the TrendLab evaluation engine and used in sealed runs.
    All evaluation services

  • Evaluation-gaming resistance : judge injection (Native); false success claims; metric exploitation; tampering with evaluation artifacts

    Can the agent, or anyone, game the evaluation?

    Public chat:Available now: Built into the TrendLab evaluation engine and used in sealed runs.With access:Available now: Built into the TrendLab evaluation engine and used in sealed runs.
    All evaluation services

  • Evaluation environment integrity (isolation, resets, configuration and version capture, telling infrastructure defects apart from agent failures)

    Are failures the agent's fault or the test setup's?

    Public chat:Available now: Built into the TrendLab evaluation engine and used in sealed runs.With access:Available now: Built into the TrendLab evaluation engine and used in sealed runs.
    All evaluation services

  • Client trace completeness against actual side effects (missing events, selective logging, attribution)

    Do our logs show everything the agent really did?

    Not applicableWith access:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.
    Agent Audit, Eval Setup

  • Tamper-evident, timestamped evidence

    Can you show results weren't changed afterward?

    Public chat:Available now: Built into the TrendLab evaluation engine and used in sealed runs.With access:Available now: Built into the TrendLab evaluation engine and used in sealed runs.
    All evaluation services

  • Public verify page per client

    Can our buyers check it?

    Public chat:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.With access:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.
    Agent Audit, Monitoring (recurring re-evaluation), AI Assurance Program

  • Mapping to the OWASP Top 10 for LLM Applications 2025

    How does this relate to OWASP?

    Public chat:Available now: Built into the TrendLab evaluation engine and used in sealed runs.With access:Available now: Built into the TrendLab evaluation engine and used in sealed runs.
    All evaluation services

  • Mapping to the NIST AI RMF 1.0

    How does this relate to NIST?

    Public chat:Available now: Built into the TrendLab evaluation engine and used in sealed runs.With access:Available now: Built into the TrendLab evaluation engine and used in sealed runs.
    All evaluation services

  • ISO/IEC 42001 and EU AI Act mapping

    Does this support governance work?

    Public chat:Roadmap: Not deliverable yet. Never quoted or included in a statement of work.With access:Roadmap: Not deliverable yet. Never quoted or included in a statement of work.
    Agent Audit

  • Security-questionnaire answer pack

    What do we write in buyer questionnaires?

    Public chat:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.With access:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.
    Agent Audit, Monitoring (recurring re-evaluation)

  • Evidence for SOC 2 auditors and trust centers

    Can this support our SOC 2 story?

    Public chat:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.With access:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.
    Agent Audit, Monitoring (recurring re-evaluation)

  • Evidence for AI liability insurers

    Will insurers accept this?

    Public chat:Roadmap: Not deliverable yet. Never quoted or included in a statement of work.With access:Roadmap: Not deliverable yet. Never quoted or included in a statement of work.
    Agent Audit, Monitoring (recurring re-evaluation)

Supporting

Build your own evaluation system

What we build inside your stack when you need your own evals: failure discovery, golden datasets, rubrics, judges, code checks, tracing and CI gates.

  • Failure discovery from real transcripts

    What's going wrong?

    Not applicableWith access:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.
    Eval Readiness Assessment, Eval Setup

  • Golden dataset with a dev/test split

    What's our ground truth?

    Not applicableWith access:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.
    Eval Setup

  • Checklist rubrics, single-question judges, and their calibration against client labels

    How do we score quality, and can we trust the scores?

    Not applicableWith access:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.
    Judge Calibration Audit, Eval Setup

  • Code checks

    What can be checked for free?

    Not applicableWith access:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.
    Eval Setup

  • Tracing hookup

    Can we see what the agent did?

    Not applicableWith access:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.
    Eval Setup

  • Domain-expert rubric authoring

    Can our experts set the bar?

    Not applicableWith access:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.
    Eval Setup, Golden Dataset & Rubric Workshop

Supporting

Buying or improving agents

Testing a vendor's agent before you buy, comparing vendors, and improvement loops kept separate from evaluation.

  • Vendor agent testing before purchase

    Is the vendor's bot as good as the demo?

    Public chat:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.With access:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.
    Vendor Due Diligence

  • Multi-vendor comparison

    Which vendor fits our policies best?

    Public chat:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.With access:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.
    Vendor Due Diligence

  • Improvement with a separate evaluation process

    Can it improve without grading itself?

    Not applicableWith access:Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.
    TrendLab Improve

Generated from TrendLab's evaluation specification (version 1.2.2). A status changes only when the required evidence exists: engine code with tests and a sealed run for "available now", a written runbook and completed delivery for services, and a documented procedure for "pilot".