F102 Tasks an agent files while working include ones that describe nothing wrong, and filing needed a bar

Field	Value
Status	active
Topic	Planning, recording, and what the person decides
Grade	R3/S3
Cases	roadkeep, Shio, winwright, freewilly
Contrary	none
Harness facts	none

Claim

In four projects, tasks filed during work were later retired as observations, duplicates or false premises, and in the one session that was measured only three of ten filings described something wrong.

Mechanism

An instruction to record what a task revealed is read as licence to file anything noticed, and a filing costs the agent one command while its cost lands on whoever picks the task up. A filing made from a hurried reading (a truncated listing, an encoding error of the agent's own, open lines not read) inherits that reading's error. A backlog drained by a loop that stops when nothing remains cannot stop if each task files more than it ships.

Observations

roadkeep. Measured over one session: 5 tasks shipped and 10 filed, so that "a loop told to stop when nothing remains could never stop" [roadkeep@91754240:.claude/skills/roadkeep-dev/SKILL.md#L42-L43],

and "Three of those ten qualified" [roadkeep@91754240:.claude/skills/roadkeep-dev/SKILL.md#L47]; the other seven were observations that belonged in the commit body. The rule written in response is "A task that revealed nothing wrong files nothing." [roadkeep@35fc90c2:.claude/skills/roadkeep-dev/SKILL.md#L41]

Shio. SH519 was retired because the defect was the agent's own: "the mojibake was mine, not the CLI's" [shio@c17ed995a],

with the conclusion that "A filed defect that does not exist costs whoever picks it up more than the bug would have" [shio@c17ed995a]. SH1096 was filed on a claim that the repository had no React SDK, made because "my directory listing truncated before it" [shio@c952ebc33], and was retired in that commit. SH815 was abandoned because "SH815's premise was the same mistake" [shio@05c1660b9].

winwright. WW186 was retired as a duplicate: "Filed without reading Block E's open lines" [winwright@861b82e:docs/CHANGELOG.md#L198].

WW120 and WW169 were retired when a measurement taken before building showed the premise false: "Measured before deciding and the premise did not survive" [winwright@861b82e:docs/CHANGELOG.md#L358], and "Measured before building it, and the premise did not survive" [winwright@861b82e:docs/CHANGELOG.md#L485].

freewilly. DD268 was retired as "Filed against the wrong writer" [freewilly@c1c2eaf:docs/CHANGELOG.md#L98],

and DD195 because it was "Filed with a section heading naming DD193, an unrelated task" [freewilly@c1c2eaf:docs/CHANGELOG.md#L53].

polyweave. No retirement at the pin (git grep -c 🗑 6d1c136 -- docs finds none). A

deferral applies the bar explicitly: "Nothing is filed, because nothing is wrong." [polyweave@5a3a26c] This is neither support nor contrary evidence.

Independence

Four origins. Each wrong filing is a failure in the project where it was filed (rule 1). The bar on filing is one remedy, invented in roadkeep; the retire verb that records the retirements is a shared tool and contributes nothing to the count.

Counter-evidence

Outside roadkeep's one session no source gives a rate, and the retired shares at the pins are small: 24 retired against 1,170 shipped in roadkeep, 4 against 276 in freewilly, 7 against 497 in winwright, none in polyweave (counts from the case studies). A retirement is a lower bound, since a wrong filing can also be shipped or left open, but these figures do not show that most filings are wrong. Not every retirement is a wrong filing: freewilly's DD117 was a decision against, not a mistake [freewilly@c1c2eaf:docs/CHANGELOG.md#L201].

Grade

R3: independent failures in four projects, not all five. S3: roadkeep's source counts the session, ten filed and three that qualified. The other observations are recorded retirements (S2). Not S4: the bar is a written rule, and no check refuses a filing that describes nothing wrong.