breakpilot-compliance/ai-compliance-sdk/internal/iace/pattern_relevance.go

package iace

import "strings"

// Pattern relevance gating (engine-side backstop).
//
// The pattern engine fires any pattern whose required tags are present. Some
// patterns are gated only on near-universal tags (e.g. "structural_part" — every
// machine has a frame) or on strong-but-broad tags a machine genuinely has
// (gravity_risk, moving_part, high_force). That lets context-specific patterns
// from other environments leak in (tick bites, confined-space oxygen, palletizer
// reach-in). IsPatternRelevant is the text backstop: a pattern made only of
// GENERIC hazard vocabulary (quetschen, stromschlag, absturz, person ...) is a
// universal machine hazard and stays; a pattern carrying a machine-, environment-
// or organism-specific word (palettierer, klettergeraet, zeckenbiss) only applies
// if that word actually appears in this machine's limits.
//
// This is a BACKSTOP, not the primary gate — the authoritative gate is the
// engine's machine-type + required-tag matching (see patternMatches). Keeping the
// relevance logic in this package lets the precision test exercise the exact
// production path.

// genericSafetyTerms are exact words that appear in almost all risk assessments
// and must NOT be treated as machine-specific.
var genericSafetyTerms = map[string]bool{
	"maschine": true, "anlage": true, "bereich": true, "gesamte": true,
	"arbeitsplatz": true, "gefahrbereich": true, "gefahrstelle": true,
	"gefahrenstelle": true, "person": true, "werker": true, "bediener": true,
	"steuerung": true, "schutzeinrichtung": true, "sicherheit": true,
	"betrieb": true, "wartung": true, "instandhaltung": true, "reinigung": true,
	"bewegung": true, "beweglich": true, "feststehend": true, "teil": true,
	"teile": true, "oeffnung": true, "zugang": true, "gefahr": true,
	"verletzung": true, "quetsch": true, "scher": true, "schneid": true,
	"stoss": true, "schlag": true, "einzug": true, "brand": true,
	"motor": true, "antrieb": true, "achse": true, "achsen": true,
	"kabel": true, "leitung": true, "schaltschrank": true, "spannung": true,
	"schutz": true, "gehaeuse": true, "oberflaeche": true, "boden": true,
	"leitfaehig": true, "elektrisch": true, "mechanisch": true,
	"bedienfeld": true, "display": true, "anzeige": true,
	"energie": true, "druck": true, "temperatur": true,
	// Abbreviations and synonyms that should not trigger relevance filter
	"kss": true, "emv": true, "esd": true, "dcs": true, "plr": true, "sil": true,
	"hmi": true, "sps": true, "rcd": true, "loto": true, "psa": true,
	// Common action words
	"bersten": true, "platzen": true, "abspringen": true, "spritzen": true,
	"einatmen": true, "ausrutschen": true, "herabfallen": true,
	"durchschlaegen": true, "wegschleudern": true,
	// Common structural terms that don't indicate a specific machine
	"gesamter": true, "gesamtes": true, "bereichs": true, "stelle": true,
	"innen": true, "aussen": true, "transport": true, "seite": true,
	"front": true, "rueck": true, "ober": true, "unter": true,
	"fuehrung": true, "lager": true, "verschleiss": true, "welle": true,
	"getriebe": true, "kette": true, "riemen": true, "feder": true,
	"spindel": true, "werkzeug": true, "werkstueck": true, "flucht": true,
	// Generic anatomy (too short for safe prefix stems)
	"arm": true, "arme": true, "bein": true, "beine": true, "fuss": true,
	"fuesse": true, "kopf": true, "koepfe": true, "hand": true, "haende": true,
	// Common German function words (prepositions/conjunctions/determiners) that
	// are not machine-specific but survive the >=5-char specific-word cut.
	"zwischen": true, "durch": true, "gegen": true, "neben": true,
	"hinter": true, "waehrend": true, "sowie": true, "dabei": true,
	"dadurch": true, "wodurch": true, "beim": true, "etwa": true,
	"jeder": true, "jede": true, "jedes": true, "dieser": true, "diese": true,
	"dieses": true, "welche": true, "welcher": true, "deren": true,
	"dessen": true, "sodass": true, "damit": true,
	// Location prepositions — never machine-distinctive.
	"ueber": true, "oberhalb": true, "unterhalb": true, "innerhalb": true,
	"ausserhalb": true, "entlang": true, "angrenzend": true, "darunter": true,
	"umliegend": true, "benachbart": true,
}

// genericStems cover inflected generic words by prefix (German adds suffixes:
// person→personen/personal, arbeit→arbeiten/arbeitsraum, quetsch→quetschen).
// Only stems long/distinct enough that a prefix match cannot catch an unrelated
// specific compound are listed. This is the lemma half of the filter.
var genericStems = []string{
	// actors / organisation
	"person", "arbeit", "taetig", "mitarbeit", "bedien", "werker", "nutzer",
	"betrieb", "wartung", "instandhalt", "reinig", "einricht", "transport",
	"qualifik", "unterweis", "schulung",
	// hazard phenomena / kinematics
	"quetsch", "scher", "schneid", "schnitt", "stich", "stoss", "schlag",
	"einzug", "einzieh", "erfass", "wickel", "absturz", "abstuerz", "sturz",
	"stuerz", "kollision", "anprall", "anstoss", "verbrenn", "verbrueh",
	"verletz", "gefaehrd", "klemm",
	// energy / electrical / thermal descriptors
	"stromschl", "spannung", "elektr", "thermi", "energie", "leitung",
	// anatomy (long enough for prefix)
	"finger", "koerper", "gliedmass", "extremit",
	// structure / location / generic qualifiers
	"bereich", "struktur", "gehaeuse", "oberflaech", "beweg", "feststehend",
	"schutz", "sicher", "maschine", "anlage", "betriebs",
	// common generic verbs / adjectives (contact, motion, causation, state)
	"beruehr", "greif", "treff", "fall", "faell", "loes", "oeffn", "schliess",
	"gelang", "erreich", "direkt", "schwer", "offen", "scharf", "teil",
	"moeglich", "fehlend", "unerwart", "ploetzl", "unkontroll", "versehentl",
}

func isGenericTerm(w string) bool {
	if genericSafetyTerms[w] {
		return true
	}
	for _, s := range genericStems {
		if strings.HasPrefix(w, s) {
			return true
		}
	}
	return false
}

// narrativeTokenSet builds the set of words the machine actually describes
// (limits text + component names), normalised and de-duplicated.
func narrativeTokenSet(narrative string, compNames []string) map[string]bool {
	set := make(map[string]bool)
	add := func(text string) {
		for _, t := range strings.Fields(NormalizeDEPublic(text)) {
			t = strings.Trim(t, ".,;:!?()/-\"")
			if len(t) >= 4 {
				set[t] = true
			}
		}
	}
	add(narrative)
	for _, cn := range compNames {
		add(cn)
	}
	return set
}

// specificWordInNarrative reports whether a machine-specific pattern word is
// present in the machine's vocabulary. Matches on token boundaries (full token,
// or either word a prefix of the other for ≥5 chars) so German inflection is
// tolerated ("behaelter" ~ "behaeltern") without substring false positives
// ("arbeiten" inside "bearbeiten").
func specificWordInNarrative(sw string, tokens map[string]bool) bool {
	if tokens[sw] {
		return true
	}
	if len(sw) < 5 {
		return false
	}
	for t := range tokens {
		// Only the inflection direction: a narrative token is the specific word
		// plus a German suffix ("behaelter" → "behaeltern"). The REVERSE
		// direction is dropped — it let a long pattern word anchor on a short
		// common narrative token (pattern "uebertragen" matching "ueber",
		// "zugangsbereich" matching "zugang"), the dominant false-positive class.
		if len(t) >= 5 && strings.HasPrefix(t, sw) {
			return true
		}
	}
	return false
}

// IsPatternRelevant checks whether a pattern applies to the machine in the
// narrative. A pattern with no machine-specific word is generic → relevant. A
// pattern with specific words is relevant only if at least one appears in the
// machine's own vocabulary.
func IsPatternRelevant(mp PatternMatch, narrative string, compNames []string) bool {
	patternText := NormalizeDEPublic(mp.ZoneDE + " " + mp.ScenarioDE + " " + mp.PatternName)

	var specificWords []string
	for _, w := range strings.Fields(patternText) {
		w = strings.Trim(w, ".,;:!?()/-\"")
		if len(w) < 5 || isGenericTerm(w) {
			continue
		}
		specificWords = append(specificWords, w)
	}
	if len(specificWords) == 0 {
		return true
	}

	tokens := narrativeTokenSet(narrative, compNames)
	for _, sw := range specificWords {
		if specificWordInNarrative(sw, tokens) {
			return true
		}
	}
	return false
}