Skip to main content

​
Street-address baseline: approved labelled scope

Status: Option A scope approved by Sid on 2026-10-28 (America/Los_Angeles) for 0.5.0-beta.2; implementation is a draft under verification/review. This is the remaining street-address slice of D2 / #144, independent of the masked-account slice. The proposal is based on #159’s partial-rule/model compatibility foundation. The implementation adds the approved labelled rule while retaining configured model coverage.

​
Approved decision

Use A: explicit labels first, using the bounded number-first English street-line format below. Keep unlabelled addresses with the configured model and user review. This adds useful deterministic coverage without deciding that every number/name/suffix phrase in prose is a private address. B: unlabelled baseline is not activated. The reviewed alternative would apply the same value grammar anywhere in text, without a label. It catches Send to 12 Willow Road but also matches duration/navigation prose such as We need a 10 minute drive. That tradeoff remains outside this implementation. Neither option is universal address validation or complete UK/US coverage. The decision-history corpus fixtures/proposals/street-address.jsonl makes both options explicit. Executable rule acceptance uses fixtures/street-address.jsonl, selecting only A and adding boundary regressions. Neither corpus is a model qualification set. false_positive_tradeoff rows describe B’s cost rather than claiming the phrase is an address. Sid approved A. Any expansion to B or other geographic formats requires separate review.

​
Approved value contract

  • Canonical type: STREET_ADDRESS, distinct from Small’s full ADDRESS.
  • A requires exactly street address:, shipping address: or billing address: (ASCII case insensitive), followed by horizontal spacing or no spacing after the colon. Bare address:, JSON keys, labels without a colon and intervening prose do not supply context. B needs no label.
  • Value begins with 1–6 contiguous ASCII house-number digits, optionally followed by one ASCII letter, then horizontal spacing. This lexical bound is not an allocated-address or house-number validation rule.
  • Street name: 1–5 Unicode letter tokens, permitting apostrophes (' or ’) and hyphens inside a token. Embedded periods, numeric street names, leading unit lines, number ranges and slashed numbers are outside this first format.
  • Suffix vocabulary, ASCII case insensitive: Street/St, Road/Rd, Avenue/Ave, Lane/Ln, Drive/Dr, Court/Ct, Terrace/Ter, Way, Close, Place/Pl, Crescent/Cres. Abbreviations may have one final period. Optional following direction: N, S, E, W, NE, NW, SE, SW.
  • Optional single trailing unit: Apt, Suite, Unit, Flat or Floor plus a 1–8 ASCII alphanumeric identifier, with an optional intervening comma and horizontal spacing. No leading-unit or multiple-unit grammar is proposed.
  • Match the entire supported street line, including a supported trailing unit and direction. Stop before a city/region/postcode/country. A comma, semicolon, line boundary or end of input terminates the line; sentence-final punctuation is outside the span except a period that belongs to a suffix abbreviation. Do not accept a shorter prefix of an unsupported leading-unit, numeric range or continued value. Malformed/additional Apt, Suite, Unit, Flat, Floor continuations and the unsupported English markers Apartment, Ste, Room, Building, Level are rejected after a comma or line ending; the captured span never crosses lines. Empty comma/line components do not hide a following known unit marker. Other comma components terminate the street span. Unknown unit vocabulary cannot be distinguished universally from a city by this lexical rule; retain model/user review for those formats. Fixtures make these boundaries concrete.
  • Horizontal spacing is space, tab, NBSP or narrow NBSP. A label and its street line stay on one line; no multiline capture. Unicode letter/digit/underscore adjacency cannot be used to find a candidate inside another token.
For decision history, B would use the value and termination grammar even where A’s label cannot apply (for example, after a newline). Neither option introduces a global code exemption: a supported literal phrase in a fenced code block still matches. A closing quote is not a supported terminator in this initial format. This is an English-suffix display format that includes some UK/US-style examples, not a country-wide postal parser. Non-English number-last formats, named premises, PO boxes, intersections and international component orders remain model/user review work. A stricter geographic deliverable requires a separate reviewed corpus; do not infer one from the UK-market discussion alone.

​
Existing corpus reconciliation

Reviewed source: PR #115 fixture file at 54dad9d. It has 85 records and 38 STREET_ADDRESS spans. All annotated codepoint ranges select their exact declared text. Its stated street-line convention is not consistent with 12 spans in 11 rows that include city/postcode/country or multiple lines: d4-addr-0007, 0008, 0010, 0027, 0028, 0030, 0034, 0035, and d4-mixed-0002, 0005 (two spans), 0010. For example, d4-addr-0007 includes 10 Downing Street\nLondon\nSW1A 2AA; d4-addr-0030 includes the street plus city/postcode and Canada. Keep their offset validity separate from the boundary-policy mismatch. #115 also includes PO boxes, intersections and leading-unit forms outside the proposed baseline. Do not silently relabel its gold data or reuse its reported native GLiNER2 figures as Small qualification. Its model-quality work remains separate from D2’s narrow rule coverage. Before pilot scoring, approve street-only vs full-address annotations, reconcile these records, and score actual Small ADDRESS output under its own agreed alignment. The rule proposal does not solve the separate ADDRESS to STREET_ADDRESS scoring contract.

​
Compatibility requirements for implementation

The implementation registers a partial street rule without replacing configured native/custom model STREET_ADDRESS coverage, using #159’s internal distinction. Keep model ADDRESS labels unchanged. Default scans may combine a narrow street rule with a model’s wider span; existing Core selection and evidence composition apply, with no blanket source precedence. Do not rename or truncate model full addresses to make a street-only score pass. Runtime acceptance must cover positive/hard-negative formats, selected and default scans, all offset coordinate systems, structured field boundaries, allowlists, provider round trips and model routing. Retain existing precision and recall thresholds, installed SDK smoke and live parity. The implementation includes a user-facing changelog and public examples. Implementation, CI and release review still apply before #144 or D2 can be marked complete. This scope approval does not qualify the model or release beta.2.

​
Implementation evidence

The executable rule corpus contains 117 cases: 64 supported labelled examples and 53 negatives, including the unlabelled-policy alternatives, unsupported formats and boundary regressions. Installed Rust/Python/Node/browser-WASM checks consume the same expected street-only ranges. Contract tests separately retain native/custom model STREET_ADDRESS, equal-span evidence and wider model ADDRESS; synthetic backend spans are not an inference qualification claim. Before separator hardening, on macOS ARM64 with Node 24.18.0, the 112 short fixture texts scanned 1,000 times each after warmup averaged 0.63 µs on #159 and 1.66 µs with the new coverage. This includes returning the new findings and is an uncontrolled local measurement, not a stable performance claim. Release WASM changed from 2,458,647 to 2,466,285 bytes (+7,638 bytes, about 0.31%) with separator hardening. cargo dfv bench remains the separate provisional performance gate.