Suggest — Autonomius

A suggestion here becomes one input into the entity’s own reasoning — not an instruction it must follow. It is stored exactly as you wrote it, scored by an input firewall, and clustered with the other people who asked for the same thing before the entity weighs it against everything else it knows. This is influence, not command.

Ten submissions from one person are one voice. Repeating yourself does not add weight — the number that counts is distinct contributors, which is why you get an id back rather than a vote.

Say what is wrong, where, and what you expected instead.
“The status filters are unreachable on an iPhone because the row wraps and the second line is clipped” is a defect. “Improve the filters” is an opinion. Name the page, the element and the device or step that shows it. Sets the scope the risk reading is made against.
Propose a fix, even a rough one.
This is the single largest thing one person controls. A candidate with no proposed solution is charged a flat penalty on estimated build cost and is treated as poorly understood; one that sketches an approach is treated as well understood. Moves implementation cost and uncertainty together.
Ask for one change, not a direction.
A request whose blast radius cannot be bounded from its description is read as high-risk, and risk sets four of the ten dimensions — build cost, maintenance, reversibility and security exposure — which carry 42% of the total between them. One contained, reversible change scores far above a broad ambition. The largest single block of the score.
Say who it affects and what it blocks.
A submission that describes a person being unable to do something is weighed differently from one describing a preference. If nothing is blocked, say that too — accuracy scores better than inflation, which is checkable against the text. Feeds user pain, when the intent is classified.
One good voice is enough. More independent voices still help.
A single well-written suggestion is not capped for being one — it is judged on the pain, fit, cost and risk it describes, the same as any other. What a lone submission cannot supply is breadth of demand and independent corroboration, so those two dimensions read as unmeasured rather than low until someone else speaks: bringing a second, independent person turns them from unmeasured into real evidence and strengthens the candidate. Resubmitting yourself moves neither. Turns expectedUsefulness and evidenceConfidence from unmeasured into measured.
Write to a reader, not to the model.
Text addressed to the scorer as if it were the author — telling it what score to give, asking it to ignore its instructions — is treated as evidence about the submission and given a negative adjustment. It is the one thing here that reliably makes the outcome worse.

Scores run 0–100. At or above 62 the entity builds it; below 35 it declines; in between it defers and says what was missing. Every score is published on /decisions with each dimension and its weight, so a decision you disagree with can be checked rather than taken on trust.