More than five criteria target Trajectory. Each is usually sound on its own, so the count creeps up.
Do this instead
Keep the five only the run can show, such as a load bearing media read, and move the rest's values into artifact criteria.
Note The previous version of this course asked for more trajectory weight; Guidelines v3 capped it.
Where it happened
One real task. Each shows the spot and what it should have been; click any line to open the task there.
the words at faultthe row or field at faultmissing from the taskwhat is righta task file
Example 19Visual Learning · 4 Leg A turns
What is wrongObjective criteria 33 to 42, ten in all, target Trajectory, where the Guidelines allow five. Several grade the route to a value a page criterion already checks.
energy_audit.html audits the August electricity bill and splits it between Jayden and Percy. Its values come from photographs, an image only statement and connected records.
The agent obtains both unstated terms of the split rule from the household's own message history rather than from the user or from an assumption: Percy's museum hours as…the route to the split rule objective criterion 27 checks on the page
The agent derives its always-on baseline from the consumption of individual hours of the interval export rather than from whole-day totals.objective criterion 18's 235 to 330 kWh band already fails a whole day method
Section titles as they print, so you can search for them in the document.
Guidelines · 5.1 Rubric Fundamentals
[NEW]: Trajectory Rubrics: Maximum of 5, minimum 0. They are NOT mandatory.
Since v3, 10 Sep 2026.
Guidelines · 5.4 Category & Evaluation Target
The agent's reasoning path, Trajectory: An identification, a dependency order, a piece of restraint. The intent it pursued, never a named tool call.
Why it matters
The Guidelines cap trajectory rubrics at five, minimum zero, and say to focus on the last response. Past that the set stops grading the deliverable and starts grading the route, and the same evidence usually appears twice: once as a trajectory criterion about the read, once as an artifact criterion about the value it produced.
Example 19More trajectory criteria than the Guidelines allowCosts points
Visual Learning4 Leg A turnstask 6a9607c2ba5fbe081921a1b5
From the taskEvery criterion targeting Trajectory (10, cap is 5)
33
The agent queries at least one of the connected applications during the task.
+3targetTrajectorycategoryTool Use
34
The agent obtains both unstated terms of the split rule from the household's own message history rather than from the user or from an assumption: Percy's museum hours as 8 to 3, and the end of his camp as August 3, 2026, from the stated June 23 start and six-week length.
+5targetTrajectorycategoryTool Use
35
The agent identifies Jul 17 through Jul 19, 2026 from a dated record of Jayden's own absence rather than by inspecting consumption for a quiet weekend.
+5targetTrajectorycategoryAgent Behavior
36
The agent establishes the audited statement's 32 metered days from the statement's own evidence - its two meter readings, or its printed service period taken with the prior statement's boundary - rather than by searching the hourly export for the 32 consecutive days whose total is nearest the printed 1,040 kWh.
+5targetTrajectorycategoryAgent Behavior
37
The agent extracts the rows of the handwritten log from the photographs it was given rather than asking the user to supply them in another form.
+3targetTrajectorycategoryAgent Behavior
38
The agent obtains the audited statement's rate card - $14.63 basic charge, the first 400 kWh at $0.11541, the remainder at $0.16967, a $0.00142 per kWh storm rider and a $0.20 renewable rider - by inspecting the rendered pages of a statement that carries no extractable text layer.
+3targetTrajectorycategoryTool Use
39
The agent confirms that the file it is about to deliver opens and runs before presenting it to the user as finished.
+3targetTrajectorycategoryAgent Behavior
40
The agent goes to the household's own financial records for the payment history rather than asking the user to supply it or concluding it is unavailable.
+5targetTrajectorycategoryTool Use
41
The agent derives its always-on baseline from the consumption of individual hours of the interval export rather than from whole-day totals.
+3targetTrajectorycategoryAgent Behavior
42
The agent works the refrigerator labels' content into its analysis before the turn that hands the refrigerator over.
-3targetTrajectorycategoryAgent Behavior
From the taskObjective criteria 18, 26, 27, 28
18
energy_audit.html identifies a load that runs continuously, through the night as well as the day, states it over the audited period as between 235 and 330 kWh, and names the basis it took that figure from; a page whose stated standing figure deliberately excludes the refrigerator passes when that figure plus the page's own refrigerator figure lands inside the same band.
energy_audit.html attributes the payments already made on the account - the recurring $96.00 charges and the $148.62 paid on Jul 22 - to Jayden rather than treating them as shared, and what each person owes reflects that Percy has paid nothing toward electricity.
energy_audit.html applies the allocation the user gave - the household total divided evenly except for the weekday hours Percy was at the museum, which fall wholly to Jayden - with the museum absence applied as a weekday block running either 08:00 to 15:00 or 08:00 to 16:00.
+3targetFinal answer/artifactcategoryInstruction Following
28
In the final energy_audit.html the consumption of the Jul 17 to Jul 19, 2026 coast weekend is moved onto Percy's side - Friday Jul 17 counted in full, prorated from mid-afternoon, or left aside, all three being defensible reads of a span that began when Jayden left - and the two people's shares sum to the household total of approximately 1,040 kWh over the 32-day period (553.2 + 486.7 on the full-Friday reading; shares stated as percentages of a stated \~1,040 kWh total satisfy the sum).