[web] Research Hub /research: trials ledger, evidence matrix, gauntlet, slippage universe

[eaitbrahim]

Why this is category-of-one

No broker (Alpaca), SaaS platform, or open-source competitor (Jesse, Freqtrade) shows historical trials and negative results — their business models depend on trading looking easy. Keel's research rigor is already public and hash-chained; this issue puts it in the navigation. Parallel work stream: read-only, no dependency on #698–#707, shares no code with the cancel action — start after Sprint 1 establishes payload/nav patterns.

Context (all verified)

• Trials ledger: keel/research/ledger.py — append-only, SHA-256 hash-chained (canonical_json + prev_hash, genesis zero-hash), at docs/experiments/trials-ledger.jsonl (93 rows, git-tracked). verify_chain reports per-row breaks. Per trial: id, timestamp, session, rule, params, provenance (a_priori|fitted), kind (7 values), decision (selected|rejected|diagnostic_only), per-trade/per-bar P&L series, summary.
• Trial budget is derived, not consumed: tuning.declared_cells / explored_vs_declared enforce a sweep-vs-declaration refusal; trial_counts returns (M, N_decisions). There is no decrementing counter — the UI reports explored vs declared, never "budget remaining".
• Gauntlet machinery, all present: DSR / E[max SR] / MinBTL (research/deflate.py), CSCV PBO (research/cscv.py + matrix.py, refuses series-missing columns, warns N < 10), Monte Carlo (research/montecarlo.py — reshuffle + moving-block bootstrap with seeded reports), DCA benchmark at real fees/slippage (sim/benchmark.py).
• Slippage universe: slippage_for_quote_volume — 5bp floor, $500M anchor, 183.8bp TON cap (strategy/backtest.py); per-product table renderer with fallback/capped flags (sim/report.py _render_slippage_rows); PR feat(rules): the promotion gate prices fills per product, and the restatement that fires it (#335) #686 landed per-product pricing with the 240-trial restatement (5 rules x 24 assets x 2 regimes; median PF 0.309 -> 0.219; 0 of 120 cells clear 1.0; corpus null 0 of 138 at honest cost).
• (rule, asset, regime) buckets exist (portfolio_sim.per_bucket_pnl) but report losing buckets only — no standing matrix artifact.

Scope — four views

1. Trials Ledger — the hash-chained table: timestamp, trial hash, kind, decision, params; M/N counts; explored-vs-declared per rule; chain-verification badge; Export Provenance Ledger (JSONL). The anti-p-hacking proof in plain sight.
2. Evidence Matrix — rule x asset x regime grid assembled from ledger rows (each trial already records rule/params/series/summary — a thin assembly layer, not a new store). Cells report profit factor, fee drag, trade count with pass/fail verdicts in the real refusal grammar (e.g. n_trades 58 < min_trades 100). Negative cells are highlighted with why they failed, never hidden. PR feat(rules): the promotion gate prices fills per product, and the restatement that fires it (#335) #686's null is the headline exhibit.
3. Promotion Gauntlet scorecard — DSR (trial-count and skewness corrected), PBO, Monte Carlo histogram as pre-computed SVG polylines (the server-quantised pattern the equity curve already uses — zero client-side math), DCA benchmark comparison.
4. Slippage Universe — per-product measured bp vs the 5bp floor (PR feat(rules): the promotion gate prices fills per product, and the restatement that fires it (#335) #686's measured figures; BTC ~5.5bp at 1.1x floor, TON at the 183.8bp cap, ~10x median), fallback and capped flags carried honestly.

API: GET /api/research/trials, /api/research/matrix, /api/research/gauntlet (+ slippage data on the same surface). All read-only.

Constitutional guardrails (hard requirements)

• The Strathern rail in the UI: no sorting or ranking by profit factor anywhere in /research. Views group by rule family; columns report pass/fail gates. (Performance view's realized track-record sorting is untouched — that is reporting, not selection.)
• Zero client-side math: every quantile, DSR, PBO percentage arrives as a presentation-ready string or SVG polyline from Python.
• Feed fields never rank and never place orders — this view is display-only research state.

Acceptance

• Ledger view verifies against verify_chain on a tampered fixture (badge flips to broken).
• Matrix renders a real sweep (e.g. the feat(rules): the promotion gate prices fills per product, and the restatement that fires it (#335) #686 driver's output) with verdicts in the refusal grammar.
• A grep-level test pins "no sort control on profit-factor columns in research views".
• No POST surface; payload contract (money as Fields) respected throughout.