delegation-and-review §1: a pinned model string does not pin behavior

[firaen22]

What

Adds one bullet to delegation-and-review §1, directly extending the existing "Model lineups are volatile facts" sentence: a pinned model string does not pin behavior — a measured behavioral property of a hosted endpoint (an edge-guard rate, a failure signature, a latency class) is a measurement of a day, not of the string. Date-stamp such numbers where recorded; when a routing or safety decision depends on one, re-run the probe instead of reusing it; an undated behavioral claim about a hosted endpoint is expired on arrival.

Why

§1 already covers lineup volatility (which models exist) and cross-model-review §1 covers identity discovery — neither covers the measurements themselves drifting behind an unchanged slug. Two independent private incidents in one week, two vendors: a pre-registered one-day re-measurement of a CLI's unstated-edge guard rate (same --model flag, same prompt battery) flipped enough to force retraction of the prior day's published claim; a second vendor's endpoint inverted a reproduced infinite-loop failure into a fully guarded pass over twelve days, model strings unchanged. The failure the rule targets is reusing a stale benchmark number to route edge-risky work.

Covenant status

Both incidents contributor-reported (private benchmark harnesses, verifiable by the contributor, not linkable); cited as shape, numbers not restated. In-body unprobed marker; Provenance notes that scheduled endpoint drift cannot be fixtured — longitudinal re-measurement is the honest test, recorded as debt.

Checks

• Single-file diff (+27 lines), no other files touched
• Unicode sweep of added lines: only house-style §/—/❌
• python3 .github/checks.py: all checks passed

🤖 Generated with Claude Code