Measure per route

A platform rarely documents the edges its users rely on: whether a range's end is inclusive on every request shape, whether a flag is stable between two fetch times, when a diagnostic line is printed, which build a restart lands on, what the measuring machine injects. One measurement answers for the route it was taken on, and it reads as a general fact once written down. The next route then fails closed in a gate, or silently in the data.

The rules

1. Record the route with the fact. Write each measured platform property with what it was measured on: the request shape, the build or version, the time of day or fetch timing, the machine and its tools. A fact without its route is a guess about other routes.
2. Measure again on the new route. Before relying on the property on a different request shape, build, timing, device or machine, measure it there. One clean rehearsal on one day is one route.
3. Compare whole ranges. Compare two routes item by item over a whole range (every tick, every row, every page), because an edge that fails one item in a hundred passes any sample.
4. Know when the platform speaks. Before a gate reads a platform's diagnostic output, measure when that output is emitted (a cache reused between runs may suppress it), so a correct run is not failed for a missing line.
5. Name the build under the run. Before any restart of a tool that updates itself, check for a pending update, and record the build each step ran on; a comparison across a build change says so.
6. Subtract the environment. Before a local reading is recorded as the system's, list the origins, processes and services it saw and exclude what the system does not serve (an antivirus's injected script, a gateway holding a peer's old address).
7. A stop is proven by its effect. After stopping or restarting a server through a wrapper, check the port is closed and the child processes are gone, not that the wrapper said it stopped.

Provenance

passive income L-057 (a restart applied a downloaded update under a registered run), L-062 (a fetch boundary measured on one route dropped 8 of 692 boundary ticks on another), L-066 (the tester names generated-tick dates only when it prepares ticks), L-088 (an identity check rested on a flag bit whose stability across two fetch timings was never measured); ICC L-071 (the local gateway kept the Auth container's old address after a reset), L-087 (a Lighthouse reading included the machine's antivirus script); AI domain L-119 (stopping a backgrounded dev server left its child process holding the port). docs/SKILL_PROPOSALS_2026-10-22.md, Tier 1 item 5.