HomeBecome a HostScorecard

Scorecard: four metrics, one score, an open formula

The scorecard is not a black box. It is computed from four components over a 30-day window, its weights are exactly as written here, and it drives the placement order directly. A high score means more work at the same price — not a marketing sentence, but an input to the scheduler.

The scorecard model was defined in Phase 0; measurement for community hosts starts with the agent in Phase 1. There is no live host scorecard today.

Host section

Four metrics

The weights are written down; there is no hidden multiplier.

The score is computed out of 5. Each of the four components is normalized to a 0–1 range within itself, multiplied by its weight and summed. If we change the formula or the weights, we write it in the changelog.

0.35

Uptime

Heartbeat coverage over the last 30 days. Power and network outages, unplanned restarts and a crashed agent are all written here. This is the heaviest component, because the customer's most expensive problem is a job left stranded.

0.30

Benchmark consistency

The variance of throughput across regularly scheduled tests. What is measured is not speed but predictability: a card that is sometimes fast and sometimes slow scores worse than one that runs steadily at medium speed.

0.20

Job acceptance rate

How many of the jobs routed to you that you accepted. Repeatedly declined offers break placement and keep the customer waiting. A declared maintenance window is not counted here — when you say up front that you are closed, no job is routed to you.

0.15

Network

The median of measured download and upload bandwidth, normalized against an upper bound. It counts because model and dataset transfer shape the customer's first experience; going beyond the bound earns no extra points.

ComponentWeightWindowHow it is measured
Uptime35%30 days, rollingAgent heartbeat coverage; scheduled maintenance windows excluded
Benchmark consistency30%30 days, rollingThroughput and variance of a regularly scheduled reference job
Job acceptance rate20%30 days, rollingNumber of jobs routed / accepted
Network15%30 days, rollingMedian of measured download + upload, normalized to an upper bound

A new machine does not start with a low score but in an "insufficient data" state: for the first 72 hours it warms up on a limited job flow, and the score settles as measurements accumulate. That warm-up period is not a penalty, it is data collection.

Placement

The score builds the ranking; it does not decide alone.

The scheduler picks the most suitable row, not the cheapest one. Scoring weighs price, scorecard and latency/network fit together; if the customer wants to, they can shift the weighting on their side (cheapest, fastest, Verified only, and so on).

First, elimination: can this job run on this machine?

VRAM, card count, disk, region constraint and operating system class are checked. A long training job rules out machines on the WSL2 profile from the start; that is a question of fit, not of score.

Then, scoring: price × scorecard × network

The remaining candidates are reduced to a single score. A machine with a high scorecard can come out ahead even if it is slightly more expensive — because the cost of a job that fails is larger than the price difference.

Tie-breakers

When scores are close, priority goes to the machine in the same region, with the image already in its cache, and that has taken less work in the last 24 hours. The last item exists to stop supply piling up on a single host.

What a low score concretely means

As the score falls you drop out of long and critical jobs first, then out of the general pool. Below a certain threshold the machine stays listed but is only routed short, interruptible jobs. Delisting is the last step and appears in the panel as a warning beforehand.

Raising the score

Concrete steps, in order.

The score is not something to guess at: the panel shows the 30-day trend of every component and which event lowered the score. The items below are the interventions that make the biggest difference.

Define a maintenance window: enter a schedule instead of shutting the machine down without notice. A planned window does not lower uptime; an unplanned shutdown is written to the heaviest component.
Turn overclocking off, fix the power limit: the consistency component punishes variance. A steady, predictable card scores better than an intermittently fast one.
Bring the temperature down: thermal throttling produces benchmark variance directly. Dust cleaning and airflow are the cheapest improvement your score can get.
Filter out jobs you can't accept: the fastest way to lower your acceptance rate is having jobs routed to you that you cannot take. Set your capacity and job-type filters to match reality.
Switch to a cable: Wi-Fi produces measurement variance; it is the most common cause of a low network component. On most machines Ethernet is a one-off fix.
Free up space in the NVMe cache: when the disk fills, image pulls slow down, cold start stretches out and job acceptance risks timing out.
If you're on Windows, move to Linux: update-driven restarts drag uptime down. Booting Linux from a second disk gives the biggest score jump on the same hardware.

Transparency

Scorecard data is public, and manual intervention is not hidden.

Everyone sees it

Every machine's score and its four components appear in the marketplace listing and in the API. A customer can filter by "above this score only". You cannot hide your score; a machine that works well wants to show it anyway.

The computed score is never deleted

If we intervene by hand in an incident — for example, offsetting an outage caused by us — the computed score is not deleted. Two values sit side by side in the panel: the measured score and the adjusted score. The reason for the intervention and its date go on the record.

How to object

If you believe an outage was not yours, you report the incident from the panel. We review it; if you are right we make the correction and leave it visible as described above. We do not quietly add points.

If the weights change

If the formula and the weights change, it is announced in advance and written in the changelog. If a retroactive recomputation is done, that is written too.

The scorecard model is defined but is not yet collecting data for community hosts: the agent ships in Phase 1. Today the supply in the marketplace comes from integrated providers and is scored by our own benchmark bot. Last updated: 29 August 2026

Next step

The scorecard ceiling lifts in the Verified tier.

A high score means more work in the general pool. Enterprise workloads come through a separate door: the Verified tier, with an on-site audit and an SLA commitment. It opens in Phase 3, and its criteria are already written down.

; scores and measurements are illustrative.