Module: Iriq::Specificity

Defined in:

lib/iriq/specificity.rb

Overview

Per-Recognizer claim strength. Higher specificity wins when multiple Recognizers fire on the same segment; the ensemble picks the max(specificity × confidence).

The bands below capture the current type taxonomy at coarse-grain: they’re explicitly NOT linear “how confident” scores. They encode “how surprising would it be for this Recognizer to fire by accident on a different actual type.” UUID’s shape is so distinctive that a non-UUID producing that string is vanishingly unlikely (SEMANTIC); a 4-digit integer could plausibly be a year, an HTTP status, or an ID, so ‘:integer` claims only TYPED.

Calibration corpus tests in spec/iriq/calibration_spec.rb / Go’s calibration_test.go are the source of truth for whether these values are well-chosen — adjust them and re-run to validate.

Constant Summary

SEMANTIC =

Unambiguous semantic shapes — the regex effectively can’t fire by accident. (UUID, JWT, email with @, URL with ://, color hex.)

1.0

STRUCTURED =

Restrictive structured patterns. Could collide with broader types at edges. (date, file with known ext, ipv4, mime.)

0.8

BOUNDED =

Digit-shaped with an additional bound — range or allowlist — that makes the shape alone meaningful. (timestamp, currency, country, boolean.)

0.7

TYPED =

Lexically broad but typed. (integer, float, version.)

0.5

PATTERN =

Generic pattern-based shape. (slug.)

0.3

FALLBACK =

Generic fallback shapes. (literal, opaque_id.)

0.1