The LiangShen V4.1 Flash comparison needed evidence before any default could
move, and the tooling could not produce it: a run could not price itself, could
not tell a workspace that failed to reach a host apart from a model that
guessed, and recorded a Windows shim scripting error as the DSH version.
Runner and report:
- add tools/prices/deepseek-flash.json (CNY per million, published off-peak row)
and price runs and the budget gate from it via --budget-cny; cost is no longer
a conversion factor living in a reader's head;
- separate WEB_* transport failures from the tool errors the model caused, so an
unreachable host never reads as a model mistake;
- record research calls and the inspections preceding the first write, the
guardrail the external we-need experiments measured falling from 30 to 6;
- report a metric an older record does not carry as unmeasured rather than as a
measured zero;
- derive the DSH version from a version token and fall back to the package the
dsh shim resolves to.
Corpus and analysis:
- add three external-verification tasks whose facts are published outside the
workspace, plus two workspace-local authority tasks that keep the same
guardrail measurable without the network; every new check fails on its seed;
- state in analyze-session whether the log carried reasoning text, because a
signed block with an empty string makes an empty counter look measured.
Decision record:
- move the improvement plan to implemented with the measured outcome: all five
arms passed 15/15 and every paired comparison was 0.0pp, so the shipped
persona and presentation 'both' stay the defaults; the candidate persona
showed no advantage and the ptc presentation traded a cleaner tool surface for
63 percent more output tokens;
- archive the evaluation snapshot and update the README pair.
Gates: pnpm typecheck, pnpm test (350 tests in the package, 8 focused additions),
pnpm test:standards, pnpm docs:check, pnpm i18n:check. pnpm test:scripts fails on
this host for a pre-existing unrelated reason: scripts/ pass GNU tar's
--force-local while the system tar is bsdtar 3.8.8.
EDDYCRAZY-CC
767512ecfd
