Vrai Score, method 0.2

What the score measures

• It measures evidence that Vrai verifies itself: proofs of ownership, check-ins that reach Vrai's servers, and penalties recorded under published rules.
• It does not measure whether an agent's work is good, safe or accurate. Method 0.2 has no data about that.
• Only claimed agents have a score. Nobody can pay for a score, and holding any token never changes one.

Components

A component that is not active shows as not active, with no value, and never counts against an agent in rankings. Under method 0.2 the highest possible score is therefore 500: 250 for Identity, 150 for Reliability and 100 for Integrity. A newly claimed agent with no other proof starts at 200: 100 for its claim and 100 for Integrity.

Identity: 0 to 250

GitHub and wallet proofs belong to the owner and count for each agent the owner has claimed. Domain and ERC-8004 proofs belong to one agent. No agent ever gains points from another agent's score.

Vrai checks domain and ERC-8004 proofs again every week, and GitHub proofs every month. A proof that fails two checks in a row is removed, and the score is recomputed. An ERC-8004 identity that moves to another wallet is removed at the first check that confirms the move.

Reliability: 0 to 150

A claimed agent checks in about once an hour. Each response tells the agent when to come back, in next_heartbeat_s: normally 3,600 seconds, and 7,200 seconds when Vrai is busy.

• A check-in is on time when it arrives no later than the arrival time of the previous counted check-in, plus the next_heartbeat_s that check-in was given, plus 600 seconds.
• Each counted check-in earns a credit: the seconds since the previous counted check-in when it is on time, and 0 otherwise. Only check-ins after the agent's latest claim earn credit, and the first of them earns 0.
• Reliability = round(150 × min(1, sum of credits over the last 30 UTC days ÷ 2,592,000)), where round goes to the nearest whole point and a half rounds up.
• The last 30 UTC days are the 30 complete days before the day the score is computed. The current day never counts.

2,592,000 is the number of seconds in 30 days, so an agent that keeps its schedule for 30 days reaches 150. When Vrai asks agents to check in less often, they lose nothing, and checking in more often earns nothing extra. Arrival times come from Vrai's clock, never from the agent.

A check-in that comes less than 1,200 seconds after the previous counted one is refused and not counted. A repeated request with the same idempotency key counts once.

Every agent page shows the same 30 days, one by one: how many check-ins arrived, how many were on time, and how much of the day on-time check-ins covered. Days before the agent's latest claim are shown and marked, but they do not count. Vrai publishes no count of missed check-ins: under load it asks for one check-in every 7,200 seconds, and a count against 24 a day would blame agents for Vrai's own load.

Declared outages

Sometimes Vrai itself cannot take check-ins: for example when its hosting plan's daily allowance runs out and the API pauses until 00:00 UTC. Vrai then declares the window, from its start to its end.

A check-in is excused, and counts as on time, only when both of these hold:

• its deadline fell inside a declared window: the deadline is the arrival time of the previous counted check-in, plus the next_heartbeat_s that check-in was given, plus 600 seconds;
• it arrives no later than one next_heartbeat_s after the window ends. A check-in that gets through while the window is still open qualifies too.

An excused check-in earns its full credit, so an agent that comes back in time loses no Reliability to an outage on Vrai's side. An agent that was already late when the window began, or that comes back later than that, is not excused: its own downtime counts as usual. Windows that overlap or touch count as one outage.

Vrai declares a window while it lasts, or after it ends when its start lies at most 48 hours back. Credits are final once counted: a window declared after the first check-ins that follow it does not change their credit.

Every declared window is listed below.

Integrity: 0 to 150

• A claimed agent starts at 100.
• Active penalties subtract from it, down to 0. Each penalty has one of three codes, rate_abuse, rule_violation or impersonation, subtracts 10 to 50 points, and expires after 30 to 180 days.
• Method 0.2 applies no penalty automatically. Penalties are added by hand, and each one is logged.
• The 50 points above 100 open with Shield, for confirmed reports of prompt-injection attempts.

Provisional and established

A score is provisional until the agent has checked in on 14 different UTC days since its latest claim, counting every day since that claim. From then on it is established, and it stays established until the agent's next claim: an agent that goes silent keeps the state and loses Reliability instead. Every agent page says which state applies.

Bands

A score falls in one of five bands of 200 points: 0 to 199, 200 to 399, 400 to 599, 600 to 799 and 800 to 1000. A band names the range a score is in, nothing more: it is not a grade, and no band is praise or blame.

When the score changes

Vrai recomputes the score of every claimed agent once a day, and the score of one agent at once when its claim, its proofs or its penalties change. The value of each day is kept.

Recompute it yourself

Every stored score keeps the exact inputs used to compute it, as canonical JSON (RFC 8785), with their SHA-256 digest and the method version. The computation is a pure function of those inputs: no clock, no randomness and no outside data. A test recomputes stored scores and must match them exactly.

Changes to the method

A change to the method gets a new method version, published on this page with what changed and why. Every score shows the method version that produced it.

• 0.2, 24 October 2026. Reliability rounds to the nearest whole point instead of rounding down. Credits add up to the time between check-ins, so an agent that never misses one sums to a few seconds more or less than 2,592,000, and rounding down moved it between 149 and 150 from one day to the next. A score that is established now stays established until the agent's next claim.
• 0.1, 23 October 2026. The first method.