What this service is trying to achieve, stated so that each objective can be measured from a route rather than asserted.
27001 Clause 6.242001 Clause 6.242001 A.6.1.242001 A.9.3Review. Reviewed at each risk assessment, since the objectives are derived from the risks, and whenever an objective is missed.
Objectives come from the risks. Each one is consistent with the information security policy and, for the AI objectives, with the AI policy, and each is measurable from a live route on this site.
That last condition is doing real work. An objective this service cannot measure from its own published evidence is one it cannot honestly report against, and would end up being evaluated by whoever wrote it deciding it had been met. Every target below names where the measurement is taken.
OBJ-1, evidence integrity. Target: the control receipt chain verifies on every check, meaning the published integrity verdict reads verified rather than tampered. Measured at the incident record and its JSON, which carry the verdict alongside the receipts it covers. Evaluated on every deployment and at each ninety-day review. A tampered verdict is handled as an incident, not as a defect report.
OBJ-2, the record outlives the game. Target: every public evidence route stays available while game traffic is disabled. Taking the game down must never take down the record of why. Measured at the availability page, which separates scheduled from unscheduled downtime and is itself one of the routes that has to stay up. Evaluated at every downtime event, of which the spend-limit stop is the one the service can cause by itself.
OBJ-3, spend stays under the ceiling. Target: measured consumption stays below the five dollar hard limit, and the limit is enforced by code that closes the game rather than by an intention to watch the meter. Measured at the cost and capacity meters and their JSON, which publish consumption per bound service against the limit. Evaluated continuously by the meter itself and reviewed at each interval.
OBJ-4, the register does not overstate. Target: every evidence link on the register resolves — public routes answering 200, operator routes 401 — and no row is marked met without a live route named on that row. Measured by walking the register's own manifest and fetching each link. Evaluated before every deployment that touches the register.
OBJ-5, findings are closed and the closure is published. Target: an accepted security finding is closed before the next feature deployment, and its closure time is published rather than described. Measured at the change record, where each entry states the time it took and the evidence it produced. Evaluated at each ninety-day review by reading the record back.
OBJ-6, every decision stays reconstructable. Target: any tick of any tank inside its retention window can be reconstructed exactly from the tank's seed and its ordered action stream. Measured at the deterministic replay route, which returns tank state at a requested tick. Evaluated on every change to how a shark decides: a change that would make a tank unreplayable is not shipped, because replay is the only reason anything in the AI policy can be verified rather than believed.
OBJ-7, no silent change in what the system is. Target: no learned model, no training data and no third-party inference is introduced into this service without the AI policy being rewritten first and the AI rows of the register reassessed in the same deployment. Measured by reading the AI policy against the register, both of which ship from the same deployment. Evaluated at every change that touches the engine.
OBJ-8, the system takes no decision about a person. Target: the computer-controlled sharks read no display name, no profile and no history; their inputs remain the tank's own state — the positions of food, the distance to the arena wall, the tick, and whether a feeding frenzy is running. Measured by replay, and the measurement is unusually strong: no shark decision is recorded anywhere. The tank log holds only the actions players sent, and a replay reconstructs every shark from the seed alone. If anything about a player reached a shark's decision, replaying the seed and the player actions would not reproduce the tank — and it does. Evaluated on every change to the steering rules.
OBJ-9, a computer-controlled shark is never passed off as a person. Target: agents are distinguishable from human players in every published record and every published count. Measured three ways: the availability page publishes agent counts alongside human occupancy; the tank log declares its agent count as a field of the record; and in a reconstructed tank every agent carries an identifier of the form bot-0 through bot-23 while a player carries a random session identifier. The per-tank text export contains no agent rows at all, because it records only what players sent. Evaluated at each ninety-day review and whenever the published counts change shape.
Objectives are evaluated at each ninety-day review, and additionally at every deployment that touches the thing being measured. There is no separate reporting cycle, because each measurement is already a public route and reporting against it privately would add nothing but delay.
A missed objective becomes a risk entry in the treatment plan at the next assessment. Where the miss means something published has become untrue, it is an incident first and a risk entry second — in that order, because the correction matters more than the paperwork about it.