ntkit v2.3.0

ntkit v2.3.0

/research-nt turns a web question into a cited report. It plans with three independent agents, requests plan approval, researches sections, edits, and checks sampled claims.

Plan memory also gains quoted provenance, standing questions, and recall over prior records.

[2.3.0] — 2026-09-30

Added

• /research-nt: a question-to-report workflow with optional vault lookup, an approval stop, bounded web fetches, section research, editing, and sampled citation checks.
• /research-nt L2–L5: three independent planners and a merge judge, one critic/reviser pass, one scoped editor, and sampled claim verification with one repair allowance. L6 comparative research quality remains unmeasured.
• evals/research-bench.py blinds the five preregistered report pairs and computes the documented shipping rule from complete rubric verdicts and claim-support counts.
• quoted provenance. A tag may carry the record's own words, [from: 2026-09-10-summary "parked until usage passes 10k rows"]. plancheck checks that the words are in the record it names and reports a new hard finding, misquote, when they are not. The comparison ignores case, whitespace, typographic quotes, dashes and markdown emphasis; a soc: quote must come from that one entry. Until now plancheck could prove a record existed but not that an item said what it said (MEMORY.md §6); a quoted item's wording is now checked. /replan-nt and /windup-nt quote where the record has the words. Bare tags stay valid.
• standing questions, an optional plan/standing.md. The human writes the ## questions; the reconcile pass rewrites the answers from the records, tagged and quoted; /resume-nt prints them without recomputing. plancheck checks its answers like pending.md's. No file, no change. MEMORY.md §7.
• recall over the records (MEMORY.md §8). /resume-nt answers "when did we decide X / why did we drop Y" by searching plan/ and _archive/ with scholia history <terms> --plan ., which dates each hit; without scholia it falls back to rg.
• tests/plancheck.test.sh: clean, orphan, ghost, misquote, soc-entry scoping, standing.md as a derived file, and --json. Against the previous plancheck it fails 8 checks.
• Research workflow documentation, updated workflow and social images, and research, hardening, chaos-leg, and ATTEST promo videos.

Changed

• G3 reconstructs the full ten-claim sample, binds verdicts to the report and stored evidence, and rejects missing repair reviews. Skipping G3 produces a diagnostic result, never VERIFIED.
• /research-nt status prints the run slug, state, and original question. Offline eval fixtures cover BUDGET resume and infrastructure failure.
• README software comparisons are optional in the packaging gate. The chaos-leg brief describes the action-sequence walk explicitly.

Fixed

• concurrent planners share a locked fetch budget, preventing simultaneous requests from exceeding their cap.

Validation limits

• The L6 comparison stopped before completion. This release makes no claim of measured superiority over the baseline research workflow.
• Native role validation does not establish Claude-host compatibility. Notification integrations remain untested.

Verification

• sh tests/run.sh: 10 passed, 0 failed, 0 skipped.
• Ruff and git diff --check passed.
• Native fixtures exercise the approval stop, BUDGET resume, infrastructure failure, and claim-repair handling.
• The five-prompt L6 comparison remains incomplete. The user explicitly requested release after stopping that comparison.
• No application deployment exists: ntkit distributes Markdown skills and scripts through GitHub.

Full changelog