Agent packs, measured by whether the code they wrote builds

Leaderboard

Every listed pack, one row

The economics are per cell and measured under this condition only — pooling the two conditions would report a cost that describes neither. Invalid cells are in the pool: the money was spent whether or not the cell was later invalidated, and a mean over the survivors would report the run as cheaper than it was. A few rows recorded no measurement at all, and those are absent from the mean rather than counted as zero — a figure taken over fewer rows than its arm's pool prints how many.

There is no composite index on this page, and that is deliberate. A single score needs weights across build rate, cost, turns and silence, and this corpus gives no basis for choosing them, so an index would advertise a ranking the measurements cannot support. What they do support is stated plainly: of the arms with a graded frontend rate, ponytail clears the baseline's interval outright.

Trade-offs

Build rate against what it costs you

Where the difference is

Packs separate on frontend work and not on backend work

The two domains are graded by different gates and are never folded into one number. The bar is the Wilson interval, the tick is the point estimate, and the label carries the counts both are computed from.

Every backend interval on this corpus overlaps every other one, so the backend tickets do not tell the packs apart — a finding about the tickets as much as about the packs, and the reason one headline rate spanning both domains would mislead in either direction.

Per ticket

Which tickets each pack is actually good at

One spoke per ticket, in the order the records introduce them. The solid shape is the measured rate; the pale shape behind it is the upper bound of the interval, drawn unclipped because its size is the finding.

Token economics

Where the tokens go

Mean tokens per cell, split by kind and printed as counts beside the bar. The segments keep a fixed order, so the row reads in greyscale.

The corpus

Every cell, one mark per repetition

This is what the rates are computed over, with nothing aggregated away. A row of one shape is a pack that behaves the same everywhere; a column of one shape is a ticket that decides itself regardless of pack. Select a mark for the cell behind it.

• built
• did not build
• wrote no code
• invalid cell

83/312 cells are invalid, and they are not spread evenly — the worst are fe-command 17/40, fe-wizard 17/40, fe-colorpicker 15/33, fe-rating 15/33. Nearly all of them are one instrument failure, a session limit hit mid-run, and they are drawn rather than dropped so a reader can judge how much each column is worth.