Verification — annotation-mapping (Stream 1b)

Date: 2026-07-11 · Runner: Claude (agent, following SKILL.md) Inputs: tree ../tree-gen/gold/remote-work.tree.json + doc gold/remote-work.source.md (Bloom 2015 Ctrip) Output: verification/2026-07-11-run.annotations.json · Gold: gold/remote-work.annotations.json

Diff vs gold

The 3 gold annotations from ../../claim-tree-annotation.md (expanded to 5 (passage, node) pairs in the gold fixture):

Gold annotation	Node	Relation	Result
an-1 "13% improvement in performance"	`q-root`	supports	match (same quote, node, relation)
an-2 calls-per-minute in quieter environment	`tasks-solo`	supports	match (slightly longer quote span; same node + relation)
an-3 gains narrowed on return to office	`time`	complicates	match
an-4 call-centre setting	`whom`	limits	match
an-5 call-centre setting	`tasks`	limits	match (same passage, second node — per-pair emission works)

Extra beyond gold: an-6 (satisfaction/attrition → q-root, supports) — borderline (satisfaction is not productivity); flagged as weak in its note. Acceptable per gold-fixture criterion ("extra annotations are fine"), but a candidate for tightening the precision guidance.

Mechanical checks (python): every nodeId exists in the tree, every relation in {supports, complicates, limits}, every quote is a verbatim substring of the source doc.

Verdict

Pass by the PLAN.md criterion — each of the 3 gold fragments lands on the correct node with the gold relation (5/5 (passage, node) pairs matched).

Still to do per SKILL.md: try one new doc against the same tree to check generalisation.