review(ted): complete PR #65 grading contract

[TabsPhasers]

Summary

This draft pull request is the reviewer-authored remediation package for WebHarbor PR #65. It completes and hardens the TED task grading contract that was reviewed against the remote PR head 065e39a4e1bf33e61902abdfb3e9d35f1e62428e.

The original PR remains unchanged. Its remote head contains 18 tasks and 18 verifiers; this draft adds the missing two high-difficulty comparison tasks and carries the fixes required by the Review Environment Skill. Maintainers can review this as a replacement/companion PR before deciding how to land the remediation.

What changed

• Added TED--18 and TED--19, two filtered, multi-page view-count comparison tasks with dedicated deterministic verifiers and judge rubrics.
• Re-anchored narrow tasks on visible TED searches with distractor results and explicit selection steps.
• Tightened verifier contracts for exact filters, required detail/listing navigation, title/speaker/year constraints, non-empty final answers, and state deltas.
• Kept all ground truth inside the reviewer-authored verifier code. sites/ted/tasks.jsonl has no answer key and contains only the permitted task and grading-contract fields.
• Retained raw Browser Use trajectories, screenshots, and before/after SQLite snapshots in the local review archive. They are intentionally excluded from this public PR because they contain session-sensitive artifacts; public-safe aggregate reports and controls are included below.

Review outcome for PR #65

The correct status for the current remote PR head is REQUEST_CHANGES until equivalent changes are present on that PR. The remote head has the following blocking gaps:

1. It has 18 rather than the complete 20-task contract.
2. Several tasks omit a broad visible search path or use narrow queries, leaving first-result and distractor-quality risks.
3. Task 11's verifier does not bind the requested TED2026 and duration filters.
4. Task 3 accepts a speaker-only token for a talk-identification task.
5. Task 16 does not require visiting a real /talks listing.
6. Task 17 checks November without requiring the TEDNext 2025 event/year.
7. Stateful verifiers do not consistently reject empty final answers.

Validation

• 20/20 clean Luna Browser Use runs passed their deterministic verifiers.
• 20/20 homepage-only, empty-answer no-op controls failed their verifiers (exit code 1).
• 20/20 primary trajectory, screenshot, and database audits passed.
• 20/20 sanitized blind Claude judge packets passed.
• Full local environment smoke passed: control plane healthy, all 17 site ports returned HTTP 200, TED reset was ready, and instance/seed MD5s matched.
• The pinned Hugging Face asset revision 597623a2f32898afa12e3bbeda15520f559aa7c7 was checked directly and returned HTTP 200.

The full-environment smoke used the existing local webharbor:ted-review image. A fresh source rebuild of the exact remote PR head was not established because unrelated local asset locks affected the earlier build; CI should remain the final reproducibility check.

Evidence

• Final by-task matrix
• Verifier logic review
• Primary trajectory audit
• Independent blind Claude results
• Full environment smoke
• No-op verifier matrix
• Raw per-task trajectories, screenshots, and SQLite snapshots remain in the local review archive and are available for maintainer inspection through an approved channel.
• Remote PR metadata
• Hugging Face revision check

Review contract

Each task has one verifier under sites/ted/verify/verify_<N>.py and a corresponding judge_rubric in sites/ted/tasks.jsonl. The verifier is deterministic-first and checks navigation, visible task facts, final output, and SQLite state where applicable. The LLM judge is secondary; no-op failure and state-delta checks are retained as independent controls.

Checklist

• 20 tasks and 20 dedicated verifiers
• No answer key in the agent-facing task rows
• No-op controls fail all verifiers
• Clean positive-control matrix passes all verifiers
• Browser screenshots and operation logs retained in the local review archive
• Public PR excludes session-sensitive raw artifacts
• No remote branch was modified by the original PR review
• Maintainer review and merge decision pending