Bound sized in the wrong basis

What goes wrong

A limit is expressed as a share of something else - a percentage, a fraction of a budget, a per-part allowance - while the quantity it must cover is an absolute amount that does not scale with it. The number looks safe because it is large in its own basis, and nobody restates it in the units of the thing it has to absorb.

Example

A reserve set to ten percent of a budget is meant to absorb one extra step, but on the smallest budget ten percent is less than a single step costs.

Not to be confused with

Not Quantity written in wrong unit: both sides here carry the same unit, and the fault is the reference the bound is measured against rather than a missing conversion.

Discovered by

@sks24git GitHub

Field notes · 9

@sks24git2026-11-06beat 3 review rounds

One capped list held words from two sources, and each fix moved the other source to the back. Whichever came first could fill the cap, so a big input pushed the other words out of the list.

Give each source its own quota. Reordering a shared cap only changes who loses.
Test the cap with a large input from each side. One small fixture hides the squeeze.

@sks24git2026-11-06beat 3 review rounds

A batch reader appends each chunk by spreading every row into one push call. Once a chunk holds about a hundred thousand rows, the call overflows the stack and the whole read fails.

Never spread a query result into a call. Argument count has a hard engine limit; append in a loop.
Test batch code at its real chunk size. Small fixtures never reach the limit.

@sks24git2026-11-04

A time budget was split into equal shares in build order, so the first build, often the one that matters most, got the smallest slice. Time a later build never needed could not flow back to it.

Reserve a floor for each later step, and let the current one use the rest. Fixed shares starve the early work.
Test the budget with a slow first step. Equal fast fakes hide the split.

@sks24git2026-10-30beat 1 review round

A warning list was moved to the top of a report so a reader who sees only its first part would not miss it. The list had no cap, so one large failure could fill that first part and push every real result out of sight.

Cap anything you put before the content in a cut-off view. Moving it up to survive the cut makes it what gets cut.

@sks24git2026-10-27beat 1 review round

A feed poll slowed to once a minute after an error, and the comment promised one refused call per minute. The poll fans out one request per category per tick, so a refused viewer logs nine denials a minute.

Count server calls, not timer ticks. A fan-out inside one poll multiplies every rate you promised.
Test the real call site. A one-category test never sees the multiplier.

@sks24git2026-10-16

A depth limit walks a parsed body level by level and counts plain values as one more level. A body at the stated limit with a value inside gets refused, and a boundary test built from empty containers stays green.

Test a limit with a full payload at the edge. Empty edge cases skip the leaf path where an off-by-one hides.

@sks24git2026-10-09beat 1 review round

A fix made old pages keep their head, not their tail, but kept the small budget sized for tails. The key line sat just past the cut, and the test, built on a short page, stayed green.

Size a test fixture from the real input. A fixture shorter than the budget cannot see a bad cut.
Re-check a budget when its direction changes. Head and tail need different sizes.

@sks24git2026-09-15

A constant was defended in a comment as measured rather than guessed, and the figure it cited measured setup hooks. Hooks are charged to another budget, so it reserved time out of a total nothing spends.

Check the citation measures your claim's subject. A real number from the wrong thing reads as rigour.
Pin a derived constant below its measurement. Asserting it exists lets it drop to one.

@sks24git2026-09-09beat 4 review rounds

A safety margin was a percent of a per-model budget, but the overshoot it must absorb is an absolute token count. On the smallest budget the accepted maximum covered one tool output, and the rescue request itself failed.

Size a reserve in the unit of the overshoot. A percent of a small budget covers less than one step.
Pin the boundary, not just the default. The maximum the validator accepts will be run.