Repository navigation
fix(routing): verify handoffs before model comparisons and retain normal-work feedback - #50
Conversation
◈ PR Lens
Architecture 10 components touched across 3 lanes. Play the interactive walkthrough Inside the changed components — 2 viewsComponent view — Bounded input delivery and study guardrails Chunking reference materials, host-visible verification, budgeted recovery, and study runner guards. Component view — Local feedback summarization and usage correlation Correlating test results, observed host settings, and token meters to detect model substitution locally. Data flow
Follow each request, response and payload The other flows — 1 sequence
View
Tip Run 🪧 More tips
Thanks for using PR Lens! It's built by Coldtea, free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. |
Truncated reference reads can otherwise be graded as model failures and consume a whole study budget. Add bounded input manifests and host-visible chunk checks, require manifests before new live study claims, and stop later claims after invalid delivery. Invalid/unknown live inputs are excluded before grading and prevent comparison export; original outcomes, grades and usage remain intact.
4 of 4; depends on #49, #48 and #47. Targets
routing/dynamic-setupso this diff contains the final step. Review/merge #47 → #48 → #49 → this PR; retarget to main when preceding branches merge. Final package version: 0.17.3.No paid model/Jev calls, study reruns, personal activation or marketplace changes. Supplied visible-read records establish delivery only within their evidence scope; they cannot prove comprehension or authenticate fabricated source labels. Hosts without visibility report unknown.
Full verification:
docs/reviews/2026-09-29-dynamic-routing-verification.md.