Pressure benchmarking

When to use

Use when the person asks to raise load, tighten resource budgets, prove a hot path stays near-linear, or drive optimizations against measured ceilings rather than intuition.

Skip for pure documentation or contract edits with no executable path, and say so. Route one-change dependency base/head deltas to dependency-audit upgrade resource compare. Route continuous ship-path absolute size and build-timing gates to delivery-budgets. Route utilization review claims to engineering-audit resource and budget mode.

Quick reference

calibrate meter -> baseline -> stress -> tighten -> (optimize -> prove) x N -> accept|rollback

Rules that fail the episode if skipped:

• calibrate before baseline (stage-distinct self-test; uncalibrated is not a baseline)
• name one primary series before each optimize cycle; change only that series until prove
• samples-per-load >= 3 for gate, propose, and write; n=1 is observe only
• episode iterations N >= 10 before accept (pressure cooker); one green prove is not enough
• do not claim stage RSS without an isolated probe

Read references/pressure-loop.md before changing ceilings or claiming cheaper. Read references/known-practices.md for layered gates, noise floor, and stop criteria.

Binding (consumer)

Record next to development process docs when the tree has local measure commands:

• measure command that prints stable samples
• self-test command with distinct known p95 per named stage
• ceilings file path
• load axis definition
• samples-per-load default (>= 3)
• min episode iterations (>= 10)
• propose-basis per series (median or max, times headroom; default headroom 1.15)
• Memory probe class (process_lifetime or isolated_stage)
• gate / propose / write command flags

Absent a binding, the agent still runs calibrate and the minimum series with explicit one-off commands and records them in the live-work note.

Related

• engineering-audit resource and budget mode - utilization claims; prioritized candidates for the loop
• engineering-audit performance mode - wait-time claims
• dependency-audit upgrade resource compare - base/head deltas for lockfile or runtime-pin changes
• delivery-budgets - continuous ship-path absolute size and build-timing gates