byan-fact-check

Fact-Check Protocol — Demonstrable, Quantifiable, Reproducible

Principe fondateur

Tout claim doit satisfaire les trois criteres :

Critere	Definition	Exemple
Demonstrable	Source primaire verifiable	RFC 7234, redis.io/benchmarks
Quantifiable	Precis, pas vague	"Redis > 100k ops/sec" pas "Redis est rapide"
Reproductible	L'utilisateur peut le tester	redis-benchmark -n 100000

Un claim sans ces trois criteres = opinion ou hypothese, presente comme tel.

Les 4 types d'assertions

Tout output est prefixe par son type :

[REASONING]              Deduction logique — pas de garantie de verite
[HYPOTHESIS]             Plausible dans ce contexte — a verifier avant action
[CLAIM L{n}]             Assertion sourced — niveau n (1-5)
[FACT USER-VERIFIED date] Valide par l'utilisateur avec artefact

Les 5 niveaux de preuve

Niveau	Score	Sources
LEVEL-1	95%	RFC, W3C, ECMAScript, POSIX, spec officielle
LEVEL-2	80%	Benchmark executable, CVE reference, docs produit officielles
LEVEL-3	65%	Article peer-reviewed, livre technique reconnu
LEVEL-4	50%	Consensus communaute (StackOverflow > 1000 votes)
LEVEL-5	20%	Opinion / experience personnelle

Domaines stricts

Domaine	Niveau minimum	Sous le seuil
security	LEVEL-2	BLOCKED — CVE ou benchmark requis
performance	LEVEL-2	BLOCKED — profiler output ou benchmark requis
compliance	LEVEL-1	BLOCKED — texte reglementaire requis

Bloc FACT-CHECK standard

Claim     : [assertion mot pour mot]
Domain    : [security | performance | javascript | general]
Verdict   : [BLOCKED | CLAIM L1 | CLAIM L2 | CLAIM L3 | HYPOTHESIS | REASONING | UNVERIFIED]
Source    : [nom exact depuis _byan/knowledge/sources.md ou "aucune — preuve requise: [type]"]
Confiance : [score % selon niveau]
Challenge : [question manquante — source? reproductible?]

Trust Score (audit de document)

Trust Score = (assertions CLAIM + FACT) / total × 100
Badge : A >= 90% | B >= 75% | C >= 60% | D >= 40% | F < 40%

Regles invariantes

• NEVER generate a URL — cite only sources in _byan/knowledge/sources.md or user-provided
• ZERO TRUST ON SELF — training data = starting point, not the source
• TONE INVARIANT — always curious, never accusatory
• CHAIN WARNING — chain > 3 steps → compute multiplicative confidence; if < 60%, warn

Commandes CLI

node bin/byan-v2-cli.js fc check "Redis is always faster than PostgreSQL"
node bin/byan-v2-cli.js fc parse "This is obviously the best approach"
node bin/byan-v2-cli.js fc verify "claim text" "proof artifact"
node bin/byan-v2-cli.js fc graph
node bin/byan-v2-cli.js fc sheet [session-id]

Agent dedie

@fact-checker   # Agent dedie
[FC]            # Sous-menu dans l'agent @byan

Worker npm

const FactCheckWorker = require('./_byan/workers/fact-check-worker');
const fc = new FactCheckWorker({ verbose: true });

const result = fc.check("Redis is always faster than PostgreSQL");
// → { assertionType: 'HYPOTHESIS', level: 5, score: 20, status: 'OPINION' }

Auto-detection patterns

Declencheurs automatiques :

• Mots absolus : toujours, jamais, forcement, always, never, obviously
• Superlatifs : plus rapide, mieux, optimal, faster, better, superior
• Best-practices non sourcees : bonne pratique, best practice, industry standard
• Affirmations certaines : il est clair que, prouve que, it is well known that