Tradefloor 0.5.0: sources, evidence and cited reports

Tradefloor 0.5 adds a source-to-report workflow to the local research desk.

• Immutable text, HTML and text-based PDF library, original/extracted hashes and BM25 passage search.
• Explicit SEC discovery and public HTTPS document capture, with date and redistribution-rights review.
• Exact passages frozen into an evidence branch before model context preview and Start.
• Cited reports, Markdown/HTML downloads and browser Print / Save as PDF.
• Interactive research map, recorded replay and evidence withdrawal branches from 0.4.

The public demo uses synthetic fixtures and needs no signup or API keys. Real sources and
model connections run locally. No demonstrated trading edge, broker orders or public multi-user backend.

Verification: 233 Python tests and 23 frontend tests passed locally. The installed wheel's
real stdio MCP workflow exercised 64 tools and the scenario/source/search/freeze/report chain
without provider calls. A real public-document capture succeeded. Browser checks covered
source freezing, saved-project reopen, exact cited reports and downloaded Markdown/HTML,
mobile layout and standalone disabled operations.

Cloud model calls remain skipped by request. Full model-driven Claude Code/Codex sessions,
SEC live capture, externally witnessed forecasting results and factual entailment remain
unverified. Scanned/encrypted PDFs need a verified text transcription; PDF extraction is
time-bounded but is not a complete malicious-document sandbox. Automated browser file
selection was blocked by the ChatGPT extension permission; parser/API tests passed.

Live demo: https://himanshuj16.github.io/tradefloor/

CI and static deployment passed on the reviewed release revision. The demo was verified in a real browser on desktop/mobile.

v0.3.0

0.3.0

• Persistent responsive research workspace: projects, revision history, branch comparisons,
  full archives, evidence inspection and independently frozen evidence revisions.
• Distance-ordered finite-grid counterexample search and endpoint sensitivity priorities.
• Evidence lineage, exact document anchors, source families, contradiction/retraction
  propagation and bounded document changes with exact character spans.
• Immutable information vintages with as-of selection; SEC submissions metadata and ALFRED
  vintage fetchers. Live provider credentials and availability assertions remain explicit.
• Four-model forecast arena, external forecast interchange, paired moving-block loss
  intervals and all-class reliability. No automatic model promotion or skill claim.
• Immutable agent assessments/outcomes, cutoff-safe memory, protocols, trial budgets and
  explicitly unverified witness receipts.
• Factor shock arithmetic, constrained minimum-variance allocation, constant-volatility
  risk-neutral option benchmark and descriptive tape comparison.
• SQLite jobs with idempotency, leases, heartbeat, progress, cancellation and crash retries;
  stronger semantic bundle validation and manual static GitHub Pages deployment workflow.
• 54 CLI/MCP tools. Heavy external research engines remain interchange targets, not bundled
  or verified native integrations. See docs/extensions.md.

Migration: existing run receipts are not rewritten. New evidence graphs use v2; older run
formats remain readable subject to semantic validation. Back up TRADEFLOOR_HOME before
upgrading. The browser server remains loopback-only and is not a multi-user hosted service.

v0.1.1

0.1.1 (2026-09-28)

Added

• Market swarm, /tradefloor:swarm. A seed analyst gathers the events; one persona agent
  per market participant group (India: foreign institutions, domestic funds, retail, prop
  desks, market makers, event funds; the US has its own roster) states its reaction as
  numbers; an agent-based simulation runs 2,000 traders over 1,000 Monte Carlo worlds,
  calibrated to the stock's own volatility, tails and drift, and reports the odds against
  normal behaviour, a fan of outcomes and who moved the price. what-if "<event>" reruns the
  personas with an injected event; afterwards any group can be questioned in character.
  Calls are journaled as source swarm. Inspired by MiroFish; no code shared.
• Two MCP tools, swarm_roster and swarm_simulate, also in the CLI (17 tools in all).
• journal_review breaks scores down by source (quant, full desk, swarm) when there is
  more than one.
• server/scripts/eval_swarm.py and the first calibration writeup,
  benchmarks/results/2026-09-28-swarm-calibration.md.

Changed

• The quant model's 80% range now uses max(EWMA, 1-year realised) volatility instead of
  the EWMA alone. On one-month outcomes of NIFTY 50 and S&P 500 stocks over the past year,
  its 80% ranges held 73% and 70% of outcomes before and 80% and 77% after.
  direction, analyze and trade_plan ranges are wider as a result.

v0.1.0

0.1.0

First public version.

• MCP server with 15 tools: symbol resolution for 41 exchanges and index names, technical
  report, calibrated direction forecast, fundamentals, localized news, sentiment signals,
  macro context, trade plan, watchlist scan, intraday scan (pre-market and live), journal
  record and review, settings. Also available as the tradefloor CLI.
• Two desks as agents: analyze (12 roles for one instrument) and scout (11 roles for a
  market, pre-market or live), with every role writing a file the next one reads.
• Journal grading: hit rate and Brier score against the base rate for swing calls; R
  multiples on 5-minute bars for intraday calls; big-move Brier score and a two-sided
  opening-range trade for pre-market picks.
• Point-in-time as_of and at replays, with no-look-ahead tests.
• Hosts: Claude Code, Codex, Antigravity; generated from one source by tools/build.py.
• Walk-forward evaluation script for the pre-market model.