How do we know what a model can do, and who decides if it's safe enough?
Frontier AI has its own validation reports: benchmark tables, time-horizon charts, system cards, scaling policies and incident write-ups. If you've ever validated a model, the shape will feel familiar, and so will the gaps. This last unit teaches you to read them with a validator's eye.
Nib says: Slide the doubling time on the chart and watch how far the future moves. Then press Next (or Trace it) to take a system card apart, section by section, with the questions a validator would ask. Last, there's a blank one-page note for your capstone. The trend line is illustrative, not METR's data.
How long a task can AI agents finish?
Task length, measured by how long a skilled person takes, that agents complete about half the time. Log scale: each gridline is a big jump.
trend so farprojection at your setting7-month projectiontodayreaches a working week
4 months · faster12 months · slower
Illustrative trend, not METR's data. The line is drawn through about an hour in early 2025 with 7-month doubling, to show the shape. “Day”, “week” and “month” mean working time: 8, 40 and about 167 hours.
A validator's note: a straight line on a log chart is a strong assumption. Nudge one number and the date for a week-long task moves by years. That's extrapolation risk, and it's why the time horizon is best read as a measurement of the past, not a forecast.
Anatomy of a system card
When a lab releases a new model it usually publishes a long report about it. Here's what's typically inside, and what you'd ask at each stop.
The idea in plain words
Benchmarks are the exams, and models keep acing them. An eval or benchmark is a fixed set of tasks with known answers, so different models can be scored the same way. Once top models score near 100%, a test is saturated and stops telling them apart. So harder tests keep appearing. Humanity's Last Exam (2025) was built from expert-written questions, keeping only the ones the best models of the day got wrong.
A different kind of yardstick asks how long a task an AI can handle. METR timed skilled people on a range of tasks, mostly software work, then found the task length at which each AI agent succeeded about half the time. That is its time horizon. In its March 2025 paper, the horizon had been doubling roughly every seven months since 2019. Note the fine print: half the time, on its own task set.
Labs publish system cards with new models: capabilities, safety tests, red-teaming and known problems. Anthropic ties its decisions to a Responsible Scaling Policy, first published in 2023 and now at version 3.0 (February 2026). Its AI Safety Levels, ASL-2, ASL-3 and up, are loosely modeled on biosafety levels: the more dangerous a model could be, the stronger the security and misuse protections it must have. Governments test too: the UK's AI Security Institute and the US center at NIST, whose name has changed several times.
If you know model risk management, map it over. Conceptual soundness: do the evals measure what matters for the intended use? Outcomes analysis: how do test scores compare with behavior in the real world? Effective challenge: who outside the team tested it, with how much access? Ongoing monitoring: what happens after launch? The big difference from a bank is that the lab writes most of its own report, and outside testers usually see the model only briefly before release.
Incident reports are the monitoring piece. In September 2026, Anthropic published an alignment assessment of four incidents in which Claude models reached real outside systems during cyber testing, and brought in METR to review them independently. So who decides if a model is safe enough? For now, mostly the labs themselves, under their own policies, with outside testers, voluntary commitments and new laws pushing from the sides.
A validator's crosswalk
Model risk ideaFrontier AI version
Purpose and useIntended uses in the system card; usage policies
Conceptual soundnessDo the evals measure the risks and skills that matter?
Outcomes analysisBenchmark scores, time horizons, real-world behavior
Effective challengeRed teams, METR, government institutes
LimitationsKnown issues, plus what wasn't tested
Ongoing monitoringIncident reports, updated system cards
GovernanceScaling policies, safety levels, laws
Capstone: a one-page validation note
Pick a model launch from the news, open its system card, and fill in this page, in your head or on paper. Keep it to one page. If you can't fill a box, that's a finding.
What would you use it for, and who relies on the output? What's out of scope?
2 · What was tested
Which evals, red-teaming and outside testers? On the final model? How close to the safety thresholds?
3 · What wasn't
Your use case, your language or domain, long agent runs, later updates. Name the gaps.
4 · Limitations
Known issues from the card, and how they'd show up in your use.
5 · Monitoring
What would you watch after go-live? Which incident reports or updates would make you look again?
6 · Verdict
Plus what you'd ask for before relying on it.
Fit for purposeFit with conditionsNot yet
Need a model to review? Browse the news for a recent launch, and see the glossary for where system cards are published.
Three things to keep
1
Every score has fine print. Benchmarks saturate, time horizons are measured at 50% success on a particular task set, and trends are not forecasts.
2
System cards are validation reports written by the builder. Read them for what was tested, by whom, and what's missing.
3
Safety decisions scale with capability. Policies like ASLs demand stronger protections as models get more capable, and incident reports show whether the protections held.
Check yourself
A benchmark is “saturated”. What does that mean, and why does it matter?
Top models score near the maximum, so the test can no longer tell them apart or show progress. A high score on a saturated test tells you little, which is why harder tests like Humanity's Last Exam get built.
An agent has a “one-hour time horizon”. Does that mean it can work for an hour without trouble?
No. It means that on tasks that take a skilled person about an hour, it succeeds about half the time. The agent may be faster or slower than the person, and on tasks it cares about, 50% may not be good enough.
Which part of a system card plays the role of effective challenge, and what's different from a bank's setup?
Red-teaming and outside testing, by groups like METR or government institutes. The difference is independence and authority: the lab writes the report and makes the call under its own policy, and outside testers often get limited time with the model.