What's new

localStorage quota exceeded errors are now caught gracefully. If storage quota is exceeded, the in-memory state still works for the session — it just won't persist across page reloads. Previously this threw an unhandled error.

restoreState now normalizes loaded deal snapshots the same way as localStorage hydration, backfilling any new schema fields that didn't exist when the deal was saved. Older saved deals can no longer load with undefined fields.

Clause agents no longer frame all contracts as SaaS — contract side context (buyer vs supplier) is now properly enforced throughout analysis.

Integrity engine false positives fixed: single-quote definition parsing and cross-document reference deferrals now work correctly.

Jurisdiction auto-detection no longer false-matches "New South Wales" to Wales.

Termination Rights redline playbook now correctly activates.

Market-standard reference coverage expanded to Auto-Renewal, Fee Increases, Governing Law, and Employment clauses (employer side).

Agent resilience: each of the eight clause agents now retries automatically on a Haiku fallback if the primary model times out or errors. A clear service-unavailable message is shown if all agents fail — no more silent empty results.

Scotland jurisdiction is now fully supported end-to-end. Previously, contracts analysed under Scots law were silently processed without jurisdiction context — this is fixed.

Jurisdiction is now shown as a read-only pill on every review page so you can always see which legal context your analysis used.

API cost reductions: vision extraction for scanned PDFs now uses Haiku instead of Sonnet (same quality, 20× cheaper); redline thinking budget halved from 4,000 to 1,500 tokens; chunk threshold raised from 100k to 120k characters.

Light mode: a full light palette is now available, toggled from the header. Uses crisp white cards on a slate-50 background with cool-grey text. Risk labels are automatically boosted to full-saturation colours for legibility.

Upload flow step labels fixed. The four steps (contract side → upload → commercial context → acknowledgment) now display consistent 1-of-4 through 4-of-4 numbering.

Summary page: risk score now has a label above the number. The "Prepare for negotiation" email section moved to after the clause flags, where it belongs logically.

Removed a dead "Open in PDF" button on the summary page that logged to the console instead of doing anything.

Landing page now includes a "What you get" feature grid explaining the six main outputs (risk verdict, clause checks, negotiation ladder, redlines, cross-clause risks, negotiation email).

Accepted redlines are now exported as a real Word tracked-changes DOCX using docx.js DeletedText and InsertedText revision nodes — ready to send to the other side.

Redline accept and dismiss controls wired on all five clause review pages (Liability Cap, Indemnities, IP, Data Protection, Termination). Previously only available on the summary page.

"Clause not detected" guidance added — when a clause type is absent, the review page now explains what the absence means and whether it is a risk.

Light/dark mode toggle added to the main navigation, with localStorage persistence across sessions and a flash-prevention script to avoid the wrong theme flickering on load.

Beta banner added across all pages with an updated legal disclaimer.

Logout was blocked by CSRF middleware on some sessions — fixed with a resilient session-parse fallback.

Email notifications sent to the team on feedback form submissions.

Redline suggestion feature: click "Suggest redline" on any flagged clause to generate alternative contract language. IP Ownership and Indemnities use Sonnet with extended thinking for higher-quality output; other clauses use Haiku.

Word-level diff viewer: side-by-side original vs proposed clause text with word-level highlights using LCS diffing.

Negotiation email generator: one click to produce a ready-to-send email covering all flagged issues in priority order.

Cross-clause risk engine: deterministic detection of interactions between clauses — for example, an indemnity carve-out that bypasses the liability cap creates a combined exposure not visible when reviewing each clause in isolation.

Risk verdict (0–100 score): a weighted composite score across all flagged clauses with a plain-English verdict — ready to sign, sign with conditions, or not ready.

Pre-classification step added: a single cheap Haiku call detects which clause types are present before running specialist agents. Absent clause types are skipped, saving significant cost on contracts that only cover a subset of the eight types.

Parallel agent architecture: eight specialist clause agents now run concurrently instead of sequentially, reducing analysis time from ~2 minutes to under 30 seconds.

Prompt caching: the contract text is cached across all eight parallel agent calls. Agents 2–8 pay 10% of normal input token pricing.

Extended thinking enabled for IP Ownership and Indemnities — the two clause types that require multi-step legal chain reasoning.

Anti-hallucination check: every extracted clause text is verified against the original document before being flagged. Unverified findings are marked and logged.

Initial launch. Contract upload (PDF, DOCX, DOC), text extraction, and AI analysis across eight clause types: Liability Cap, Indemnities, IP Ownership, Data Protection, Termination Rights, Auto-Renewal, Fee Increases, and Governing Law.

Jurisdiction support for England & Wales, India, Germany, and France, with jurisdiction-specific legal thresholds and risk calibration.

Three-position negotiation ladder for every flagged clause, with ready-to-use scripts.

Commercial context extraction: ACV, contract term, liability cap, insurance cover, and data type detected automatically from the contract text.