Skip to content

v2.2.0 — Grounded Comparison Briefing

Latest

Choose a tag to compare

@laanito laanito released this 21 Oct 17:53
· 14 commits to main since this release
039c395

Grounded Comparison Briefing

From Cross-market context, apply an anchor, proxies and window, then choose Generate briefing. The terminal prepares fresh server-side evidence and returns observations, limitations and unknowns with citations into that exact snapshot.

What's new

  • Inspect each citation's actual pair window, returned numbers, sources, quote units, freshness and caveats.
  • Explicit progress, cancellation and manual retries; selection changes clear prior work, and failed fresh runs never leave an older answer looking current.
  • Bounded evidence/model/repair work and validated NDJSON delivery. Only final reference-checked prose is shown; interrupted streams are not automatically replayed.
  • JWT-authenticated evidence, synchronous briefing and progress-streaming APIs under /api/market-context.
  • Fix crypto prices in stock detail: chart and snapshot use the crypto feed, unsupported Finnhub zero quotes are rejected, missing prices stay unavailable, and disconnected crypto streams have Yahoo polling fallback.

Scope and limits

The briefing includes shared-UTC native provider-close returns only, not optional news, fundamentals, SEC, macro, technical, FX/adjusted/local-date panels or private research—even when visible elsewhere on the page. Citation-reference and structure checks are not semantic fact-checking, causal explanation or investment advice. Model capability still affects prose quality. Nothing is automatically saved or indexed into Brain/Brain Log.

Upgrade and verification

No database migration or new provider setting is required. Briefings use the existing OpenAI-compatible LLM configuration. Binance's existing USDT-to-USD-symbol mapping remains a separate known limitation.

Host acceptance is complete. Release preparation passed all three CI jobs: build-and-test, clean-install and PostgreSQL contracts. Local validation passed the full backend gate with 74.90% coverage, 447 frontend tests, production build and six Chrome smoke tests.