Skip to content

conformance-ladder.md

Latest commit

 

History

History
118 lines (96 loc) · 7.88 KB

conformance-ladder.md

File metadata and controls

118 lines (96 loc) · 7.88 KB
Raw

The conformance ladder (how zic-rs admits IANA slices)

Reference standard (Arthur David Olson / Paul Eggert / Kenneth Murchison, strict): IANA slices are not trophies. They are a deliberately chosen conformance ladder. zic-rs admits a slice only when it forces or validates a distinct reference-zic semantic, proves behaviour under zdump, records provenance, and expands the support frontier honestly.

zic-rs is a producer-side compiler for one of the world's strangest living civic datasets. The tz database is collaborative civil-time infrastructure: it records historical timezone and daylight-saving rules and is updated as governments change civil-time law (the maintenance process is governance, per RFC 6557 / BCP 175 — not local preference). Our job is to compile it, never to correct, simplify, or editorialize it.

Authority stack (unchanged, restated)

IANA tzdb source  →  reference zic / zdump  →  RFC 9636 (TZif validity)  →  zic-rs + evidence

Wikipedia and articles are orientation only. Every subtle behaviour is pinned by experiment against reference zic/zdump, never inferred from prose (see reference-zic-semantics.md).

The five rules (the "make-them-proud" standard)

  1. Every slice has a reason. A slice is admitted because it forces a new semantic class or validates an important implemented one — never because it is famous.
  2. Every slice has a receipt. Provenance, the transform (if any), reference-zic output, zdump parity over a declared horizon, type/transition counts, footer, alias/link status, consumer-bench status (if run), and any caveats. Receipts live in fixtures/iana-slices/SLICES.toml; the live enforcement is the per-slice zdump oracle tests plus support-report.
  3. Behaviour first, structure second. zdump behaviour over a declared horizon is the binding contract. TZif structural parity (slim/fat transition counts) is optional unless explicitly claimed — the Test/Mixed slim/fat result is the pinned doctrine (reference-zic-semantics.md §6).
  4. Do not editorialize civil time. We compile tzdb as reference zic does. We never "fix" or normalize it (tzdb-governance.md).
  5. Links are first-class production identifiers. Production systems use aliases, not only canonical zones, and a current-offset table does not describe a zone's historical data. The alias map and support-report therefore account canonical zones, links, and total identifiers separately — never conflated.

The admission gate

A slice graduates from "pressure probe" to "admitted" only when all of these exist:

  • provenance — tzdb release + source kind, and (for de-abbreviated slices) a byte-identity proof against the verbatim extract under reference zic;
  • reference-zic output to diff against;
  • zdump parity over a declared horizon (the contract);
  • explain output (the evidence trace);
  • manifest output (--alias-map / --manifest provenance);
  • a documented semantic reason (which class it forces/validates).

Pressure-probe first (extract the slice, run zic-rs, record the first fail-closed blocker, classify it as parser vs semantic vs emission-shape), implement only what the real source forces, and call it admitted only once the oracle is green. (This is exactly how Europe/London and America/New_York were admitted — see research/SESSION-CONTEXT.md.)

The ladder

Tier Slice Semantic class it forces / validates Status
0 Europe/London multi-era, LMT, GMT/BST/BDST, non-zero fixed-offset era, recurring EU footer, effective-in-era classification, direct tzdata.zi ingestion ✅ admitted (T4.0)
1 America/New_York U.S. rule history, WWII war time, genuinely mixed-in-era finite+recurring final era ✅ admitted (T4.1)
2 Asia/Hong_Kong real inline-save / wartime behaviour (proves T3.2b on a real zone) ▷ next candidate
3 Asia/Kolkata (+ Asia/Calcutta alias) non-integer offset, historical naming / alias↔canonical ▷
4 Australia/Sydney southern-hemisphere seasonal inversion (mental-model discipline) ▷
5 Pacific/Apia date-line / skipped-day civil-time discontinuity ("final boss") ▷ later
6 Etc/UTC, Etc/GMT±N, selected Link aliases fixed/simple zones, reversed POSIX-sign Etc/GMT names, link/canonical integrity ▷

Slice quantity is not progress. "5 zones covering 8 semantic families" beats "10 zones." Pick the next slice by bucket impact (see support-report), not by fame.

support-report — the frontier map

zic-rs support-report --input <file> compiles every zone in a source file and buckets the outcome, distinguishing canonical zones from links. It reports compile support (a valid TZif is produced) — not behavioural correctness, which only the zic/zdump oracle establishes. Every zone lands in exactly one bucket and the accounting is exact (supported + Σ unsupported == zones parsed), with a catch-all other bucket that keeps the raw diagnostic — nothing is hidden.

This is how the next milestone is chosen: by the largest unsupported bucket, not by vibes. Running it on the installed tzdata.zi (2026b) at the time of writing reports 162 / 341 canonical zones compile-supported (345 / 598 identifiers including links), and named the biggest unlock precisely: 176 zones blocked by a no-rules era with a %z FORMAT (e.g. 0:30 - %z, which zic renders as the numeric offset of the era's standard offset). Implementing that single small fix took compile-clean coverage to 338 / 341 — a result no amount of guessing would have produced. That is the point of the ladder: let the real database choose the work.

Three statuses — compile-clean is not behaviour-verified

The headline number is honest only if its meaning is precise. Three distinct statuses:

  • compile-clean — support-report says zic-rs emits a valid TZif for the zone. (338 / 341.)
  • behaviour-verified — the emitted TZif matches reference zic under zdump over a declared horizon. This is the binding contract, and it is strictly stronger than compile-clean.
  • fail-closed — the zone hits an explicit unsupported diagnostic. None among canonical zones in 2026b (the negative-SAVE and non-POSIX-day-form buckets are both cleared, laws 7 + 10).

A comprehensive oracle sweep — all 341 canonical zones over 1900..2040, each compiled and zdump-diffed against reference zic — reports 341 match / 0 mismatch / 0 fail-closed (341 + 0 = 341 compile-clean), after T5 #1–#5 + law 7 + law 10: every canonical zone behaviour-matches reference zic over 1900..2040 — the canonical-zone behaviour frontier is closed. The match count rose 264 → 328 → 333 → 338 → 339 → 341 (the old 35-zone sample "28 / 6", which caught only 6 of 74, is retired): T5 #1–#5 (standard-LETTER; final-recurring anchor; standard-clock absorption −64; prior-active rule-state seeding −5; clock-reference normalization −5), law 7 (inline negative SAVE −1: Prague), law 10 (non-POSIX ON + v3 footer −2: Gaza/Hebron). Catalogue in reference-zic-semantics.md §10.

So the honest sentence is: "zic-rs compiles all 341 canonical zones from tzdata.zi 2026b and behaviour-matches reference zic under zdump over 1900..2040 for all 341; none fail closed." — over 1900..2040, a declared conformance sweep, not infinity; not full-zic replacement. The next milestone is a support-report --verify mode that folds this comprehensive sweep into the tool. Only zones that are behaviour-verified + receipted are admitted to compatibility.md.