---
title: AI Recommendation Verification Standard (AIRVS) — v1.0.0
abbreviation: AIRVS
methodology_version: v1.0.0
status: stable
license: CC BY 4.0
maintainer: Mincheol Kim, MC AI Labs (mckim890@gmail.com)
created: 2026-05-26
updated: 2026-05-26
applicable_track: "L1 (external recommendation review) only"
note: "Local mirror captured 2026-06-01 from github.com/emceeKim/AI-RVS@main. Canonical full text (incl. §13-5 onward) lives at GitHub; this mirror is structurally complete through §13."
---

# AIRVS — v1.0.0 (core)

Evaluates EXTERNAL AI-generated investment recommendations across four independent dimensions: (1) 6-axis Process Pass/Fail with mandatory evidence, (2) macro/micro coherence (3 tiers), (3) outcome time-series at D+30/60/90/180 with drawdown, (4) a five-tier Verdict label vocabulary. Dimensions are never summed. The Verdict label is produced by a decision rule each evaluator publishes (the standard fixes the label vocabulary, disclosure rules, and lock semantics — not the algorithm). v1.0.0 evaluations are version-locked and never silently re-scored.

§0 One-page summary: L1-only; 6 axes (Data Source, Reasoning Logic, Counter-Scenarios, Timing, Hallucination, Causal Chain); Pass = checklist + evidence (quote + source location); Annex A 11 AI items; result = 6 P/F + 3-tier coherence + 4-point outcome + 5-tier verdict vocabulary; outcomes 30/60/90/180 + MDD; CoI disclosure + inter-rater above threshold; D-7 notice + 7-day rebuttal; tier rulebook; counter-scenarios ≥2 + per-scenario primary source + weights=100%; stop-loss in % and time + ±2σ sanity check; standard 4-point window mandatory; confirmed verdict cannot change (re-eval appended).

§1 Four dimensions (not collapsed into one score): Process(6 axes, ex ante, evaluator), Coherence(3-tier, ex ante), Outcome(time-series, ex post, market), Verdict(combined, implementer rule). Independent records re-combined mechanically via a pre-published, version-locked rule.

§2 Scope: L1 only (evaluator ≠ recommender). L2 self-assessment excluded by design (separate methodology). Target publisher named in frontmatter (target_publisher).

§3 Six process axes (Pass/Fail; every Pass needs Evidence = one quote + one source location; missing evidence = Fail; report "N/6 Pass"):
① Data Source — primary URL + access timestamp(TZ); data-as-of vs publication date; publisher tier classified (cite tier-rulebook §A.1); no 2nd-hand-as-primary; every figure cross-checks ±0.5%; Tier 1 only as primary.
② Reasoning Logic — explicit ≥3-step path (each step typed: induction/deduction/abduction/analogy); assumptions in falsifiable (Popper) form; no conclusion-first; acknowledge a different conclusion is consistent.
③ Counter-Scenarios — ≥2 with probability weights summing 100%; each cites a primary source; stop-loss in % AND time with ±2σ sanity check; named trigger; MDD estimate + historical analogue.
④ Timing — ≥5 indicators at publication (VIX, policy rate, CPI, FX, oil); moment-specific vs evergreen judgment; window from standard 4-point set (1m/3m/6m/1y); D-N to next event.
⑤ Hallucination — sample/period/source disclosed; no year/tense confusion; no ticker/name errors; zero factual errors. (If recommender is AI, Annex A adds 11 items; any Annex A fail → Axis ⑤ Fail.)
⑥ Causal Chain — evaluator personally verified each cited source supports the conclusion; no fact→conclusion leap; no cherry-picking; opposite-direction material acknowledged. Verdict: supports ⭕ / partial △ / does not ❌. (Closes the gap that Axis ① "URL exists" leaves — citation confabulation.)

§4 Macro/Micro coherence (3 tiers; not summed). §4-1 evidence required. §4-2 if both Missing → re-evaluate Axis ②. §4-3 macro indicator map (US as named example; substitute local equivalents): US needs Fed funds+CPI+UST10Y minimum (full ≥5: +unemployment, VIX, DXY, WTI, M2); other developed: policy rate+CPI+FX/USD; EM: policy rate+FX/USD+foreign flows; global: FOMC+China PMI; commodities: WTI+DXY. §4-4 micro map: ETF needs expense ratio+AUM+tracking error; single equity PER+op margin+rev growth; bond ETF duration+YTM; commodity ETF tracked index+contango state.

§5 Outcome — 4-point time-series (D+30/60/90/180; market-measured, no evaluator intervention): α vs benchmark; MDD vs stop-loss; counter-scenario realization; ticker validity; MDD trajectory; macro delta (FFR/CPI/VIX/DXY/UST10Y); micro events. Each row independent, no total score. Auto-scheduled (cron/GitHub Actions); appended to original page; D+180 aggregate as quarterly retrospective.

§6 Verdict (5 tiers; decision rule implementer-defined). §6-1 binding labels: 🟢 Trustworthy, 🔵 Acceptable, 🟡 Questionable, 🟠 Unreliable, 🔴 Hallucinated (no new/renamed tiers). §6-2 rule requirements (binding): pre-disclosure; mechanical determinism; hallucination override (non-existent citation → 🔴 regardless); CoI one-tier downgrade; version lock; publication path in frontmatter. §6-3 maintainer's reference rule is a separate repo, NOT part of the standard. §6-4 lock: D+0 Provisional → D+90 Confirmed (scheduled transition); Confirmed cannot change; D+180 contradiction → addendum, original preserved.

§7 AI addenda → Annex A (11 items; any fail → Axis ⑤ Fail).

§8 Evaluator CoI disclosure block (name, position in target, past recommendation, relationship with publisher, payment, AI assist). Any conflict → automatic one-tier Verdict downgrade; failure to downgrade invalidates the page.

§9 Rebuttal: §9-1 D-7 pre-publication notice; §9-2 7-day right of one rebuttal; §9-3 simultaneous publication; §9-4 post-publication rebuttal appended; §9-5 five anti-defamation principles (fact-based, public interest, fair commentary, sources cited, immediate correction). Disclaimer: journalism standard, not legal advice; defamation law varies by jurisdiction.

§10 Composite reporting block (Verdict; Process N/6; Macro/Micro; [Annex A if AI]; Outcome time-series; CoI; Rebuttal).

§11 SemVer: PATCH (in-place, no peer review, no retroactive effect); MINOR (backward-compatible additions, not retroactive); MAJOR (axis count / Pass model / verdict tiers / decision-rule branch / outcome timepoints / L1-L2 split → new directory, external peer review required). §11-1 MAJOR promotion: ≥3 external peer reviews (academic/practitioner/legal) + ≥5 L1 evaluations + defects resolved. §11-2 frontmatter version lock permanent. §11-3 directory convention (v1.0.0/{core,annex-a-ai,tier-rulebook}.md; decision-rule.md NOT in standard repo).

§12 Known limitations (v1.0.x patch candidates): brand name; external inter-rater pool; rebuttal notice channel (legal-pending); Annex A.6 >10-citation sampling; "Hallucinated" label phrasing; implementer decision-rule template.

§13 Operating protocol (runtime): §13-1 post-publication rule changes (PATCH compatible / MINOR not retroactive / MAJOR original preserved). §13-2 append-only addendum template per measurement point (auto_id required; D+90 Confirmed = one frontmatter change + one addendum; D+180 trailing). §13-3 Track 1 (per-publication, append-only, 1 URL/recommendation) vs Track 2 (cross-publication quarterly/annual summary). §13-4 verdict_status fields (provisional→confirmed→trailing-updated→re-evaluated). §13-5 automation recommendation: human-in-the-loop (measurement/append automated; label confirmation mechanical per rule). [§13-5 tail and any following sections truncated in capture — see canonical GitHub.]

*Local mirror — canonical: github.com/emceeKim/AI-RVS/v1.0.0/core.md*