Changelog

DTRExp has been in the works since 2017. Draft 1, then under the name DTRE, is dated August 24, 2018; the format spent the years between in production notes and a shelved implementation. Draft 2 (2026) is a ground-up revision with one added discipline: five clean-room implementations, each built from the prose and vectors alone, plus the reference; every divergence between them resolved as a spec fix and a test vector. Newest first.

Changes from Draft 2.7 (draft 2.8)

Vocabulary, plus one late vector addition. No grammar or evaluation change. The cadence-far-horizon-{daily,weekly} coverage groups span one full Gregorian cycle (400 years = exactly 146,097 days) ending at the year-9999 domain ceiling: a D/W-period cadence evaluated that far from its anchor caught a real implementation bug (dtrexp-go’s elapsed-time estimate clamped at time.Duration’s ~292.5-year ceiling, silently turning covered instants into false negatives), and the class (fixed-width duration types saturating inside the 1–9999 year domain) is exactly what a vector should pin for every future port.

Was	Now	Why
Date-Time Range & Recursion Expression in the title; “repetition” as the umbrella term in the prose (§5, repetition.md): two words for one concept, both wrong-shaped	Date-Time Range & Recurrence Expression; recurrence is the one umbrella term: §5 retitled, repetition.md → recurrence.md	to a programmer, recursion is self-reference (a function calling itself), a claim this spec never makes, and a false signal to exactly its audience; repetition is correct but generic; it names no domain. Recurrence is the established calendar-domain term (RFC 5545 RRULE = recurrence rule; Google Calendar, Microsoft Graph, EventKit all follow), so the title now states what the spec covers in the vocabulary implementers already search for, and one concept gets one word (the doctrine §3’s -0 decision already follows). Draft-1 syntax quoted in this file (the /duration/interval/repetition tail) stays verbatim; it names the old grammar, not the concept

Changes from Draft 2.6 (draft 2.7)

From the fifth clean-room implementation’s (dtrexp-java) journal: 21 entries, probed across all six codebases (five ports + reference). Ten behaviors pinned; two were owner doctrine calls; one probe exposed a crash no vector had ever reached.

Was	Now	Why
T range endpoints’ precision semantics unstated	endpoints resolve to their literal’s span-start instant (T0900:12 ends at 12:00); equal resolved endpoints reject (T12:1200); bounds (§6) remain the one place a coarse literal contributes its full span	6-way unanimous, previously held by luck; the §4↔§6 asymmetry is now stated instead of discovered
whitespace never defined; the EBNF allowed spaces only between components while §7’s own examples put them around |	ASCII space runs: between components, around |, at either end; tabs and all other whitespace are errors	6-way unanimous
D/W durations inside H/m-period cadences undefined at evaluation	fixed absolute lengths (D = 24 h, W = 168 h); an absolute grid has no calendar anchor to constrain against; vectored across the Berlin spring-forward	5-way agreement, and dtrexp-rs crashed (unreachable!) on the very first evaluation, proof the corner had zero coverage
scope resolution (“nearest of M/Q/Y”) ambiguous between granularity and textual order; single-pass parsers would wrongly reject Q2 D40	”nearest” = finest granularity present, never position; static validity applies to the completed expression (Q2 D40 valid, Q2 D40 M1 rejected)	6-way unanimous; two journals independently flagged the ordering trap
-0 accepted as hour 0 by four of six implementations	a signed zero is never a value: the - requires a nonzero integer; H0 is the only spelling of hour 0	owner decision (one way to say it); matches the reference
E-1#2 rejected by one of six	the ordinal’s base may be negative and resolves against E’s fixed domain first: E-1#2 ≡ E7#2	owner decision (text as written; the grammar derives it, no rule bans it)
stride range ends implicitly restricted alongside starts by a broad reading of §3	only the start must be non-negative; a negative end clips per parent instance (D2:-1/5 valid)	6-way unanimous; the anti-drift rationale only ever applied to the anchor
statically-dead strides unflagged in the reference (D30:*/2 M2)	a stride is sugar over a range and participates in the §9.1 minimum; js gains the check	5–1 inside the written minimum
hour 24 in date literals and star-endpoint ranges in lists, both unstated	20180120T2400/1D rejects (hour 24 is a T-range token only); M11:*,5 and M*:3,7 are valid list items (only the bare * is banned)	6-way unanimous
—	D29 M2 Y2021: rs warns (deeper, correct analysis), five stay quiet. Left unpinned by design: §9.1 already licenses quality-of-implementation warnings beyond the minimum, and a vector could only ban one of two legitimate behaviors	recorded so it is never mistaken for an oversight

Changes from Draft 2.5 (draft 2.6)

An owner decision, closing the Y domain question the rs round surfaced: cold readers had split 3–2 on Y0 and 2–3 on an upper bound, each implementation quietly inventing its own limits where the prose was silent.

Was	Now	Why
Y domain “unbounded”; Y0 and Y12000 implementation-defined (js/py capped at 1–9999, go/swift/rs each drew different lines)	Y takes 4-digit ISO years: 1–9999, everywhere; selector values match the range the 8-digit YYYYMMDD date-literal grammar already commits to; Y0 and Y12000 are syntax errors; * on Y resolves to the domain edge (9999 / 1) like every other designator, so Y2020:* still reads “2020 and every later year”	one rule for values, literals, anchors and bounds; ISO 8601 without expanded representation is exactly the 4-digit years; year 0 is 1 BCE, an invitation to the zero-confusion §3 just closed
wrap and negative-value rejections on Y justified by “no edge / unbounded domain”, a rationale the finite domain broke	rationale re-anchored: years are absolute and non-cyclic: wrap models a cyclic domain inside a parent instance and negatives count back from a parent’s edge; Y sits inside nothing, so neither applies	the rules stand; only their reasons moved

Changes from Draft 2.4 (draft 2.5)

From the fourth clean-room implementation’s (dtrexp-rs) journal: 21 entries, 18 of which the 2.4 prose already decided (the round validated the spec). The three [VECTOR-GAP]s were probed across all five implementations; two are pinned below, one (Y0 and the Y domain bounds, a live 3–2 split among cold readers) is deferred to the owner.

Was	Now	Why
gap anchors named “Temporal’s compatible” but only the overlap→earlier direction was vectored; rs silently resolved gap anchors to the transition instant and passed the full suite	§9.3 example makes the gap direction concrete: 20240331T0230 in Berlin resolves to local 03:30 CEST (01:30Z): the pre-transition offset pushed forward by the gap’s length, never the transition instant; vectored (dst-cadence-anchor-gap-forward); rs fixed	4–1 with normative prose on the majority side; an unvectored normative sentence is exactly how divergence hides
backwards-bounds threshold undefined beyond one clearly-inverted vector	§6: a bounds range is backwards (syntax error) iff the first literal’s span begins at or after the last literal’s span end; equal literals cover the literal’s span, mixed precision is valid in either order; vectored (invalid …T1801:…T1800, coverage equal-literal + mixed-precision)	5-way unanimous behavior, previously held by luck
zero on 1-based domains implied only by the domain tables (the swift journal’s stumble #2, re-raised by rs)	§3 rule: literal 0 is out of domain on 1-based designators (M0, D0, Q0, W0 are syntax errors); ordinary value on 0-based H/m/s; D0 vectored beside the existing M0	two cold readers independently had to infer it; one sentence ends that

Changes from Draft 2.3 (draft 2.4)

Closes the round-3 triage (clean-room dtrexp-swift + the go/swift journals’ prose findings). Two behavior decisions, two unanimous agreements pinned, and a new vector class.

Was	Now	Why
”2400 is valid only as a range end”; silent on spelling: T22:24, T2300:24, T0000:240000 accepted by half the implementations	hour 24 is the exact 4-digit token 2400, in range-end position only; every other spelling is a syntax error; canonical “to midnight” is T2200:2400	both fresh clean-room readers independently derived the narrow reading (2–2 split against the older implementations); one-way-to-say-it doctrine
D366 Y2021 accepted silently by half the implementations	D366 under concrete years that lack it must warn; it falls inside the §9.1 written minimum (domain sizes implied by co-present selectors; the day-of-year twin of W53 Y2021); range forms included (D366:* Y2021 resolves 366:365, backwards in every instance)	the minimum already required it; the vectors didn’t pin it
§9.1 minimum said nothing about how far concrete-year checks must scan	required only over closed year spans ≤ 1,000 years; open/wider spans MAY stay quiet; with W present, Y is the week-year and D366×week-year MUST stay quiet (D366 W1 Y2025 covers Dec 31 2024)	implementers could not tell how much enumeration conformance demands; the week-year corner produces a genuinely covered day 366 in a 365-day selector year
warnings testable in one direction only: nothing pinned what must NOT warn	new quiet vector section: expressions that must parse with zero warnings (D366 Y2020, W53 Y2020, open spans, the week-year corner, true wraps)	a linter that cries wolf is as non-conforming as a silent one; false positives were unfalsifiable
H/m durations inside calendar-period cadence windows undecided in prose (4-way unanimous in code)	pinned: the duration follows the period’s naive wall-clock arithmetic (20241026T2300/2D/12H in Berlin ends at local 11:00 across the 25-hour night: 12 wall hours, 13 absolute)	unanimity is not a contract until vectored; this was the last un-pinned DST corner
M*:* accepted by all four, undocumented	pinned in §3.1 and vectored: M*:* ≡ M* (canonical: M*); bounds *:* stays an error	same: verification without pinning rots (the hour-24 lesson)
§3 “max” meant element count in the parse check and largest value in the resolution formula, three lines apart	parse check uses N = domain size; resolution is maxValue + 1 + v with maxValue the instance’s largest value	the swift journal’s first stumble: the examples disambiguated, the prose didn’t
—	§9.3: the fall-back transition instant itself carries post-transition fields (Berlin 2024-10-27T01:00:00Z = local 02:00:00 CET)	the boundary case was decidable only from a vector, not the prose

Changes from Draft 2.2 (draft 2.3)

From the second clean-room implementation’s (dtrexp-go) one substantive finding: no vector exercised a cadence on a DST transition, and the three implementations had quietly built three different models there. §9.3 now decides it.

Was	Now	Why
§9.3 internally torn: “occurrence windows align to the local clock” (b3) vs. an instant-resolving anchor rule (b4) that only means anything in absolute time	calendar-period (D/W/M/Y) occurrence windows are local wall-clock intervals: membership on t’s local fields; the anchor is a naive local date-time, never resolved to an instant. Fall-back: both passes covered (25-hour “day”); gap: nothing (23)	extends the T-selector’s emergent DST model to cadences; one time model, not two; matches the reference and the clean-room py reading; correct for the business-hours use case
anchor disambiguation (earlier/forward) stated for all cadence anchors	scoped to where a literal must become one instant: H/m-period anchors and bounds literals	under local wall-clock windows the rule was dead prose for calendar periods, and two of three implementations silently violated it where it was live
—	new vectors: cadence across the Berlin spring-forward gap and fall-back overlap (calendar-period), and the H-period anchor in the overlap (earlier occurrence pins the absolute grid)	the corner had zero vectors; dtrexp-py and dtrexp-go both passed the full 2.2 suite while diverging from the spec (later-anchor grids) and from each other (window model)

Changes from Draft 2.1 (draft 2.2)

From the findings of a clean-room second implementation (dtrexp-py, built from prose + vectors alone under an ambiguity-journal protocol). Its journal surfaced 28 underdetermined readings; 4 were live divergences: the reference and the clean-room implementation both passed the full vector suite while disagreeing on all four. Every fix below is vectored.

Was	Now	Why
wrap decided on “literal positive endpoints"	"literal non-negative endpoints”; on 0-based domains a literal 0 wraps: H22:0 = hours 22, 23, 0	”positive” fit only 1-based units (the same editorial slip 2.1 fixed for stride starts); H22:1 wrapped while H22:0 was empty, a discontinuity exactly where “22:00 to midnight” is written
T0900:0900 silently empty (or implementation-defined)	equal T endpoints are a syntax error	half-open, it covers nothing; typo-shaped input fails loudly
20180120T09 tokenization unstated	the date literal’s T-glue is unconditional: a malformed time-part is a syntax error, never re-tokenized as 20180120 + T09	one implementation errored, the other re-tokenized; “greedy” alone did not decide
*:* bounds derivable from the EBNF	at least one endpoint must be a date literal; EBNF tightened	an unbounded window is spelled by omitting bounds; grammar overgenerated
bounds zone unstated (“the absolute window”)	bounds resolve in the evaluation zone; “absolute” means non-recurring, not UTC-pinned	both implementations chose local, by luck; no vector distinguished the readings
cadence window end formula ambiguous for M/Y durations	end = constrain(startₖ + duration), measured from the constrained start	the two readings diverge (Apr 30 + 1M = May 30 vs May 31); all constrain vectors used 1-day durations
Y stride interval ceiling undefined (table had no Y row)	Y has no interval ceiling: Y2020:*/100 valid	the cap exists so strides cannot drift; a year stride cannot drift at any interval
negative-value parse limits unstated	symmetric domain check: -max…-1; 0-based resolution max + 1 + v (H-1 = 23)	D-31 vs D-32 was implementation-defined
"Y* is legal but redundant"	"…redundant unless it rescopes a finer component” (D-7:* ≠ D-7:* Y*)	the flat claim was false: deleting a “redundant” Y* changed meaning
§9.1 warning SHOULD unbounded	required minimum defined (per-selector domain-size satisfiability, full-domain exclusions, M∩Q disjointness); union branches all surface	implementers could not tell how much static analysis conformance demands
—	new vectors: bounds-in-zone (Berlin), month-duration window end, H22:0 wrap, Y stride, D-31/D-32, H-1, equal-T, T-glue, *:*, dead-union-branch warning	six agreements previously held by luck are now pinned; four divergences decided

Changes from Draft 2 (draft 2.1)

Adopted during implementation review. Breaking against draft 2, which nobody consumed: no compatibility owed.

Was	Now	Why
- doubles as range and negation (M3-7, D-7-*, D-7--3)	: is the range operator everywhere: M3:7, D-7:*, T0900:1800, 20200101:20210630; - means only “from the end”	one symbol, one meaning; D-7--3 was legal and unreadable. (Not draft-1’s :, which was exclusive; this one is inclusive)
backwards ranges rejected	wrap ranges for every scoped designator: M11:2 = November through February, each year (Y keeps rejecting: no edge to wrap around)	T2200:0600 already wrapped midnight; periodic cross-boundary spans without out-of-domain arithmetic
§2 note: “D follows its nearest coarser selector present”	explicit scope table; W scopes nothing; M/Q/W stay Y-scoped	W is coarser than D, so the note promised D-1 W* = last day of the week; the table and the evaluator say month: the note lied
* defined in one parenthetical (“* = domain edge”)	new §3.1: domain vs scope; * is the edge of the domain in force, never “now”; negative anchors have constant length, positive ones don’t (D-7:* is always 7 days, D25:* is 4–7)	the model was load-bearing and undocumented
M-1 Q1 valid, silent, empty	statically-unsatisfiable intersections draw a warning; warnings vectors added	negatives resolve against their own scope: M-1 is December, and December ∩ Q1 = ∅, legal, but never what you meant
§8 timeval requires 4 digits	2-digit timeval legal (T12)	§4 already declares T12 valid; the grammar catches up
§3 stride start “must be positive"	"must be non-negative”	H0/4 is legal; “positive” fit only 1-based units
month/year cadence durations checked only by the reference parser	codified: /nM, /nY durations require month/year periods; rejection vector + H/m-period coverage vectors added	the vectors are the contract, not the parser
§9.3 silent on cadence-anchor disambiguation	explicit: an anchor at a repeated local time is the earlier occurrence; in a gap it resolves forward (Temporal compatible)	the reference implementation resolved it by the sign of the zone’s offset: Berlin got later, New York earlier; mutation testing flushed it out

Review Fixes Folded Into 2.1 (independent review)

Was	Now	Why
§3 “start exceeds end wraps” vs §9.1 “resolved start > end covers nothing”, unreconciled	wrap is decided syntactically, on literal positive endpoints only; negative/* endpoints resolve per instance and never wrap; statically-always-empty inverted ranges (M-2:2, D-1:5) now warn	the two sections contradicted each other and no vector pinned the reading
§9 step 3: unitsBetween mod period < duration for all cadences	occurrence windows [constrain(anchor + k·period), +duration); the mod reduction holds only for exact units (D/W/H/m)	the formula contradicted the spec’s own §9.2 constrain vectors (20240131/1M/1D covers Apr 30)
EBNF could not derive M*/Y*; ordinals derivable inside lists	selector = designator , ( "*" | ordinal | valuelist | … ); item = value | range	the grammar undergenerated the spec’s own §10 examples
Y-1 and M1,* accepted, silently useless	both rejected (Y has no edge to count from; a list containing * is *)	typo-shaped inputs should fail loudly
T’s permitted forms unstated	§4: values, ranges, lists only; no * / ! / stride / ordinal; rejection vectors added	grammar generated forms the evaluator rejects
§5.2 duration check (“max ≤ min length”) ambiguous	exact fixed unit lengths specified (M = 28/31 d, Y = 365/366 d); 20240101/2M/58D rejection vectored	prose and implementation disagreed; “min length of 2 months” is 59 d, the check uses 2×28
W⇒week-year effect on M/Q unstated; tokenization unstated; bare-literal-with-seconds span undefined	§2: only Y tests the week-year; §8: greedy longest-match, any-width separators; §6: day/minute/second span	independent implementers could not answer deterministically
W53 Y2021 accepted, silent	statically-decidable W53-in-short-years warns	§9.1 SHOULD-flag, now honored; vectored

Considered and refused: out-of-domain overflow (M14 = February, D40 M3 = April 9). It would break the existence rule (D29 M2 = leap day only), silently change M5 Q2 (May → August), and stop catching typos, all fatal for the access-control use case. Wrap ranges + date-literal bounds cover both cross-boundary cases.

Changes from Draft 1 / DTRE (August 24, 2018)

Stated as of draft 2; ranges have since moved to :, see above.

Was	Now	Why
: (exclusive) and - ranges	- inclusive only; T half-open	one operator, one rule; off-by-ones die
! in 5 positions	! only after the designator	one negation, defined set semantics
F / L / + ordinals	negative values + E…#n	one mechanism; -1SU-style familiarity
/duration/interval/repetition tail	stride /i[/d] + anchored cadence	split by evaluation model (§5); anchor required by grammar
recurrence indices /*3;5	removed	bound the cadence with dates instead
<ᐧ <=ᐧ >ᐧ >=ᐧ = operators	date-literal bounds (§6)	same power, less grammar
Y*3 (anchorless stride)	syntax error	epoch problem solved by construction
week-of-month W	removed; W = ISO week-of-year, Y-scoped	ill-defined; notes-issues already conceded
T always UTC	tz is an evaluation parameter	business hours are local; DST
”a DTRExp should not be parsed"	"cannot generally be enumerated; evaluated by coverage”	it is parsed; it isn’t expanded
4/6-digit date literals (2015, 201703)	8-digit minimum	removes Y2015 vs 2015 ambiguity
millisecond selector S	removed (T literals keep .sss precision)	no real use case

Migrating the Old Examples

Old (draft-1 / DTRE)	Now
M3/14/14/* Y2018:* (every 14 months after 2018-03)	20180301/14M
m0:20 H00/1/4/* (da Vinci’s sleep)	m0:19 H0/4
M3/1/2/* (every 2nd month from March)	M3/2
D3:100/5/2/10:15 (recurrence indices)	gone: bound the cadence with a date range instead

What Died, and Why

• The 3-part tail on selectors (/duration/interval/repetition): duration+interval inside a parent is expandable to a plain range list at parse time: sugar, kept as the stride’s block form (see recurrence.md). Across parents it’s a cadence, moved to the anchored-cadence construct.
• Recurrence indices (/*3;5, “only the 3rd and 5th recurrence”): counting occurrences from an anchor and filtering by index; expressible as a date bound when finite; appeared in zero real examples.