We won't give you a score. We'll give you the records.

Every task an agent does is logged as a run: whether it really passed, what it cost, and which slice of memory it was given. Three documents come out of that. Their format below is the real one. The numbers are samples from Fernhill, a made-up team. Yours come from your own tasks.

1. One run

Sample numbers

The smallest unit. One task, one agent, one attempt. Everything else on this page is built from these.

{
  "task_id": "t1-thursday-order-arrival",
  "mode": "pack",
  "verified": true,
  "checks": [
    {
      "type": "command",
      "level": "L2v",
      "pass": true
    }
  ],
  "harness_measured": {
    "wall_ms": 48210,
    "input_tokens": 18422,
    "output_tokens": 1310,
    "total_cost_usd": 0.18
  },
  "context_manifest": [
    "pk_01M775JMX8HXFVCQ3Q7JX2MNBS"
  ],
  "started_at": "2026-11-10T09:14:02Z"
}

verified: Decided by tests the agent never sees. They are copied in only after it finishes. What the agent says about its own work is kept as a claim and never counted.
harness_measured: Time, tokens and cost, metered from outside the agent.
context_manifest: The exact slice it was given. Any result traces back to what the agent knew at the time.
mode: "pack" is a run with memory. "baseline" is the same agent, cold.

2. With memory, and without

Sample numbers

The same agent does the same tasks twice. Each task is one where something an earlier session learned decides whether it works. These four are samples. Yours are your own tasks.

Task	What an earlier session learned	Cold agent	With memory
Tell a customer when a Thursday evening order arrives	The courier skips weekends for live plants	Failed · $0.21	Passed · $0.18
Refund a plant returned after 20 days	A decision in force: 14 days	Failed · $0.34	Passed · $0.22
Quote 80 olive trees for a garden centre	Wholesale discounts stop at 10%	Passed · $0.41	Passed · $0.29
Fix the delivery date shown at checkout	The checkout tests need the payment mock	Passed · $0.88	Passed · $0.52
Really passed	Checked by tests the agent never saw	2 of 4	4 of 4
Cost of a pass	Everything spent, divided by what really passed	$0.92	$0.30

How the verdict reads. One of three words: better, worse or unresolved. When there are too few tasks to call it, the report says unresolved instead of rounding up.

Why cost of a pass. A cheap run that fails isn't cheap. Dividing the whole bill by the passes puts a failed attempt where it belongs: in the price of the ones that worked.

Run card: model: One model, the same in both arms
Run card: task set: 4 tasks, frozen before the runs
Run card: checked by: Held-back tests, on every number
Run card: tools allowed: The same list in both arms

3. Your week

Sample numbers

What your agents did with memory, counted from the audit log. It arrives once a week.

Fernhill, week 3	Count	What it means
Sessions	212	Each one started with a slice
Memories written	61	Kept by agents in their own memory
Lessons proposed	9	6 approved, 1 rejected, 2 still waiting for a person
Contradictions	4	3 settled. Agents act on neither side of the one still open.
Refusals	7	Times an agent asked for something outside its slice, each with the reason
Secrets blanked	3	Replaced before anything was stored
Held for you	1	Text that read like orders to an agent. It stays out until you release it.

Three things that never appear in a report

One number that blends everything. You see passes, cost and the level each was checked at, side by side.
An agent's word that it succeeded. Its claim is recorded and labelled as a claim.
Cost the agent reported about itself. If the meter couldn't measure it, the report says it is missing.

The real numbers come from your own tasks

Request access Back to what you get