57 · Financial coherence audit (2026-05-18)

Senior PM / finance + UX review of all main public pages. Each finding lists the page, the offending claim, the missing context, the verdict, and the copy fix actually shipped in this PR.

Pages reviewed:

• / (landing components in src/components/landing/*)
• /performance
• /backtest
• /methodologie
• /how-it-works
• /recommendations
• /pricing
• /docs
• Blog hero block

A · Findings (10)

A1 · LandingHero "Win rate T+90" badge has no sample or period

• Where: src/components/landing/LandingHero.tsx:299-312
• Claim: XX% of signals with score ≥ 65 are profitable at 90 days
• Missing: n, time period, comparison (vs random ~50% on noisy 90d returns).
• Verdict: Headline-only number. Reader can't tell if it's a curated 196-row subset or the full 23,788-row universe. The default snapshot fallback in LandingTrackRecord is 53.3% (whole-universe OOS), but LandingHero reads whichever number the caller passes — the live wiring uses the curated subset.
• Fix: append "· n=196 filtered, 2021-2026, vs ~50% random" to the badge.

A2 · LandingTrackRecord chart title vs metrics mismatch

• Where: src/components/landing/LandingTrackRecord.tsx:110-145
• Claim: chart is labelled "Win rate · T+90 · 24-month rolling" but the fallback values are documented in code as "last 6 months BUY cohort" (audit 51) and the sparkline points are literally faked (line 34: // Sparkline data: fake historical win rate over 24 months).
• Verdict: Misleading. Either the data is rolling-24m or it is the last 6-month cohort, it cannot be both. The fake sparkline is the most legally risky single line in the codebase: a marketing chart that invents historical values.
• Fix:
  1. Caption changed to "Win rate · T+90 · 6-month BUY cohort".
  2. Sparkline marked aria-label="Illustrative shape — see /performance for audited history".
  3. Code comment upgraded so future devs replace it with real data.

A3 · /recommendations BUY universe stat is fine but jargon-heavy

• Where: src/app/recommendations/page.tsx:53-54
• Claim: BUY universe: mean +0.78%, win rate 46.8% (n=23,788 cross-market)
• Missing: nothing factually, but "cross-market" / "BUY universe" / "composite score v5.1" are pure jargon — a retail user has no idea whether 46.8% is good or bad.
• Verdict: Honest but unreadable. Same number on /performance is contextualised ("Sigma alpha comes from selection, not from buying the universe").
• Fix: Glossary tooltips wired on win rate, composite score, cross-market.

A4 · "Sharpe 1.32" / "DSR 0.311" claims need consistent guard rails

• Where: src/components/landing/LandingFeatures.tsx:191-192
• Claim: Portfolio Sharpe DSR 0.311 (CI95 [-0.08, +2.79]) · cross-sectional 1.87
• Verdict: Excellent disclosure (CI straddles zero is shown), but the same paragraph mixes 3 different Sharpe definitions in one breath. A reader who doesn't know what DSR is will see "1.87" and stop reading.
• Fix: Glossary tooltips on Sharpe, DSR, cross-sectional. The 1.87 figure already carries an asterisk in LandingSigma, propagate the same pattern here.

A5 · "Sharpe 1.0203" style false precision

• Where: searched. None found in shipping pages — STRATEGY_PROOF rounds to 2 decimals everywhere via sharpeDisclosure().
• Verdict: Clean. Keep using .toFixed(2) for Sharpe everywhere.

A6 · Survivor-/look-ahead-bias disclosure

• Where: src/lib/winning-strategy.ts:877-884 disclosure.survivorshipBias and disclosure.pitBiasRisk are present and detailed.
• Verdict: Disclosure exists but only on /performance. Landing, blog hero and /recommendations never link to it. A user scrolling the landing sees Sharpe figures with zero bias warning.
• Fix: Footer rail "Past-performance & bias notice" link added to landing via LandingFeatures Strategy card. Already linked from blog index, kept.

A7 · Mixing whole universe with curated Sigma subset

• Where: across pages. STRATEGY_PROOF distinguishes universeSize=23,788 from filteredSubsetSize=196, and most components use the right one. Risk: a future contributor swapping them silently because the field names are similar.
• Verdict: Architecture is correct, the risk is regression.
• Fix: Glossary terms universe and filtered-subset both seeded, used in tooltips so future copy authors can grab the canonical definitions rather than improvising.

A8 · "554k+ filings" headline number drift

• Where: /how-it-works, landing hero, blog meta.
• Claim: "around 554k filings" hard-coded in several copy strings; live DB count via decls.total is closer to 591k (current count on prod, per audit 48). The hero uses the live count, the "what we do" cards still say 554k.
• Verdict: Minor staleness. The live total is preferred.
• Fix: /how-it-works cards refactored to read from the live total via the existing getDeclarationStats() helper used by the hero.

A9 · "/fonctionnement" is a routing zombie

• Where: src/app/api/docs/route.ts:53 still links to /fonctionnement/, which 301s to /how-it-works. Folder src/app/fonctionnement/ has only _components/ and no page.tsx.
• Verdict: Inconsequential, but the 301 hop is wasteful and the dead folder is confusing. Out of scope for this PR — flagged in audit 56 backlog.

A10 · /pricing Pro tier mentions "Walk-forward backtest + Bailey-LdP DSR"

• Where: src/app/pricing/page.tsx:82
• Claim: pure jargon as feature bullet.
• Verdict: A buyer who doesn't know DSR cannot evaluate the feature.
• Fix: Glossary tooltips on walk-forward and DSR so hovering a pricing bullet explains it.

B · Risks not fixed in this PR

• The "fake 24-month sparkline" data (A2 §2) is now caveated but not replaced with real rolling data. Replacing it requires a backfill job over RecoSnapshot history.
• /recommendations page table headers ("Score breakdown", "Wilson CI") are still untooltipped on small viewports — popovers don't fit on a phone. Mobile copy needs a second pass.
• The "since 2015" framing on landing implies 11 years of OOS testing. Reality is in-sample 2015-2025 + walk-forward 2025-01 → 2026-05. The STRATEGY_PROOF disclosure says so, but the landing copy ("real performance since 2015") does not. Recommended fix: split the headline into "11 years backtested, 17 months walk-forward".

C · Glossary terms shipped

25 terms seeded — see prisma/seed-glossary.ts and the /admin/glossary table. Each carries a slug, EN + FR label, ~80-word EN + FR definition, optional formula, optional example, optional source list.

Terms: sharpe-ratio, deflated-sharpe-ratio, win-rate, max-drawdown, annualized-return, out-of-sample, walk-forward, cross-sectional, insider-buy-signal, composite-score, cluster-bonus, holding-period, filing-date, transaction-date, mar-article-19, sec-form-4, pdmr, beta-neutral, t-plus-90, alpha, regime, universe, filtered-subset, wilson-ci, sortino-ratio.

D · BO additions

• /admin/glossary — list + inline edit per term. Server actions; no client state library.
• /admin/landing-copy — JSON-backed Setting rows (landing.heroHeadline, landing.kpiLabels, landing.ctaCopy). Pure form, no preview UI in v1.