Suggestions in the OrderYou Actually Take Them

Every ranking in Plexon has always come from a fixed rule that never finds out whether it was right. Adaptive routing gives that rule a way to be corrected, and it corrects it only in proportion to the evidence. With nothing observed yet, what Plexon learned counts for nothing and the order is the rule's order, ties included.

• No observations means a learned weight of zero, so the ranking is the old one
• Equal scores keep the order the catalog already gave them
• Two observations move a ranking by about a third at most
• Four observations before a learned score counts as much as the original rule
• A test asserts the cold order 500 draws in a row, exactly

Nothing here invents a new thing to measure. Every signal was already produced and thrown away: a suggestion you took or waved off, the verdict the verifier agent filed, the status a loop agent reported, whether a workflow run finished. What you typed next is not one of them.

• You accepted the suggestion, or you dismissed it
• The verifier filed pass, fail, or unverifiable, which is worth half credit
• A workflow run reached completed, or it errored
• A cancelled run counts for nothing. You stopped it, which says nothing about the choice
• Nothing reads "actually, do X instead" as a complaint about the last decision

A verifier verdict lands after the turn that caused the decision, sometimes several turns after. That is the normal case, not the exception, so each decision is filed with the id the later signal will carry and waits for it. If nothing ever comes, nothing is assumed.

• A verdict that arrives three turns later still reaches the decision that earned it
• The same signal twice resolves once
• A decision with no outcome after 7 days is dropped, and no reward is invented for it
• An echoed template like "VERDICT: pass | fail" parses as nothing at all

A preference from last spring should not outrank what you did this week. Evidence decays once a day, keeping about half of itself over a month, and an option that fades all the way back to where it started is removed rather than left sitting in the report saying nothing.

• Daily decay retains roughly half the evidence over a month
• An option decayed back to its starting point disappears from the report
• A long gap between sessions is caught up in one pass, not replayed forever
• A rule that changes between releases cannot overwrite what was learned about it

Settings, Advanced, "What Plexon learned" prints the whole state: one block per context, one line per candidate, each with its score and the number of runs behind it. The run count sits next to the score on purpose. A 100% built on two runs and a 62% built on ninety look identical when you only read percentages, and only one of them is worth anything.

• One block per context bucket, such as capability|code
• Every score carries the count of runs it came from
• Reset one bucket, or reset everything, from the same card
• After a reset, behaviour is day-one behaviour again
• The file path is printed, because deleting it by hand is a legitimate way out

A sub-task that names no agent used to run generic. That was already a choice, just one nobody was watching, so picking the specialist that finished the job before is on by default. Bidding keeps its own switch, because it is the only part that spends anything.

• Learned agent defaults are on, with a switch in Settings that turns them off
• With nothing learned the pick is the plain description match, so day one is unchanged
• An @agent mention or a manual pick is an instruction, and nothing re-ranks it
• An agent whose declared modes exclude the current mode is never selectable
• Bidding costs one short model call per candidate, at most four, and only for a close call

What Plexon learned about your choices is a profile of one person, which is why it is treated like a credential rather than a cache. It lives in a single JSON file at ~/.plexon/routing.json, written with owner-only permissions. No part of it is attached to a request: not a context, not an option, not a score.

• One local file, written atomically, readable only by your account
• Nothing from it is ever sent with a prompt
• A corrupt file costs the learning, not the feature. Plexon starts cold and keeps working
• Turn it off and every ranking is the fixed rule again, everywhere