feat(verify): CoT gate, tax/regulatory anchors, and citation data overhaul

[bensonwong]

Summary

• CoT gate: forces agents to locate the source sentence (f) before picking sourceMatch (k) — eliminates paraphrase failures by making f→k a visible substring check, not a post-hoc rationalization
• Citation data field table: replaced bullet-list field docs with an ordered table that encodes the CoT reasoning sequence (n, r, f, k, p, l) — field order now communicates the required reasoning order
• Hard f→k substring rule: added as a standalone callout in both the data-block section and the sub-agent instructions; if the planned k doesn't appear word-for-word in f, the model must fix f first
• Tax/regulatory anchor examples: added brevity examples and a failure-mode table for dollar amounts, percentages, and named legal tests
• Index/appendix/TOC rule: citations must follow refs to actual evidence pages, not structural pages
• Standards doc housekeeping: moved deep-citation-standards.md reference to the deepcitation repo; updated concepts doc reference

Test plan

• Run /verify on a document with numeric facts (dollar amounts, dates) and confirm citations use terse anchors (≤4 words) that are verbatim substrings of f
• Run /verify on a tax/regulatory document and confirm percentages/thresholds are cited with the exact figure as k
• Confirm sub-agent instructions include the CoT gate and f→k substring check
• Confirm citation data fields appear in n, r, f, k, p, l order in generated output

[claude]

Overall this is a well-reasoned set of improvements. The CoT gate is a genuinely clever fix for paraphrase failures, and the Domain A/B framing + Format 1/Format 2 split adds expressiveness without much cognitive overhead. A few things worth flagging:

---

p field format is a silent breaking change

The old example used "1_0"; the new examples use "page_number_1_index_0". The sub-agent instructions are updated consistently, but any parser or downstream consumer that splits on _ and expects exactly two numeric segments will break silently. Worth confirming the CLI already accepts the new format before this merges — if not, a migration note is needed.

---

scratch/verify-flow.html — two concerns

CDN dependencies with unpinned versions: the file loads mermaid@10 which resolves to the latest 10.x, so a minor Mermaid release can silently change diagram rendering. Pin to an exact version (e.g. mermaid@10.9.3) or inline the script if stability matters. It also imports Google Fonts via CSS, which requires a live network connection and sends request metadata externally — system font fallbacks are fine for an internal reference doc.

Should this be committed at all? scratch/ implies throwaway work. A 2,438-line generated HTML file in version control shows up in every git log, git diff, and blame. If it is a durable artifact, move it to docs/. If ephemeral, add scratch/ to .gitignore.

---

Minor: sourceContext terminology collision in SELF-CHECK step 0

"find the sentence in the evidence that proves the claim and hold it in mind as sourceContext" — sourceContext is the UI component name (rendered from l line IDs), while the field that holds the verbatim sentence is f (source_context). The step does say "write it as f" a few lines later, but the opening sentence creates a brief confusion. Changing it to "hold it in mind as f (source_context)" makes it unambiguous.

---

What is working well

• CoT gate — writing f before k and enforcing a substring check is a principled fix that closes the paraphrase loop at the source.
• Field-order-as-reasoning-order — the n, r, f, k, p, l table cleanly communicates sequencing constraints without a separate prose explanation.
• Tax/regulatory examples — the dollar-amount, percentage, and named-test examples fill a real gap. The failure-mode rows (e.g. "first 5 years of employment" to "5 years") are well-chosen.
• Index/appendix/TOC rule — the bad/good example with "A Club dues 47 F Carrying charge 25..." makes the failure mode immediately recognisable.
• Format 2 syntax — decouples Domain A voice from Domain B verbatim, which the old single-format model could not handle gracefully.

---

Summary: Resolve the p format backward-compat question, decide intentionally about scratch/verify-flow.html (pin CDN if it stays, or drop it), and clarify the sourceContext vs f naming in SELF-CHECK step 0. Everything else looks solid.

[bensonwong]

Thanks for the detailed review. Addressing each point:

p field format — no breaking change here. The page_number_N_index_I format was already the established format in the previous SKILL.md (the diff only converted the bullet-list docs to a table). The CLI has always used this format; the old "1_0" form was never in this file. No migration needed.

scratch/verify-flow.html — agreed. Added scratch/ to .gitignore and untracked the file in the follow-up commit (a3368d1).

sourceContext terminology in step 0 — fixed. Changed to f (source_context) to make the field reference unambiguous. Also updated the substring check reference from sourceContext to f for consistency.

[bensonwong]

@claude review the latest changes