Specification

Draft 2.8 · Status: RFC · 2026-07-10 · Onur Yıldırım · changes: CHANGELOG.md

A DTRExp (read: “DTR expression”) is a compact string expression denoting a (possibly infinite) set of time intervals. It is evaluated for coverage (“is this instant inside the set?”), not enumerated into date objects. Finite windows of it can be enumerated on demand.

T0900:1800 E1:5          business hours
E7#-1 M4                 last Sunday of April, every year
20200106/10D             every 10 days from 2020-01-06

---

1. Model

An expression is one or more components whose denoted interval sets are intersected; the | operator unions whole expressions into a DTRExp.

expression  =  component ∩ component ∩ …
dtrexp      =  expression ∪ expression ∪ …        (via |)

• Components may appear in any order; each designator may appear at most once per expression. Canonical (recommended) writing order is smallest unit first: T… E… D… W… M… Q… Y… <bounds>.
• Whitespace between components is optional but recommended. Whitespace is the ASCII space only, in runs of any length: permitted between components, around |, and at either end of the expression (M3|M5 and M3 | M5 are both valid). Tabs and every other whitespace character are syntax errors.
• Expressions are time-zone agnostic. The time zone is a parameter of evaluation (an IANA time-zone identifier, "Europe/Berlin"), never part of the expression. Default: UTC, under which no DST behavior exists at all; the local-time semantics of §9.3 apply only when the caller asks for a zone that has them.
• All ranges over discrete calendar units (days, months, …) are inclusive on both ends: D1:15 is exactly the first 15 days. Only the continuous time-of-day component T uses half-open ranges: T0900:1200 covers 09:00:00.000 up to but not including 12:00.

2. Designators

Designator	Unit	Domain	Scoped by (nearest of these present, else default)
Y	year	1–9999 (ISO)	absolute
Q	quarter	1–4	Y
M	calendar month	1–12	Y
W	ISO week	1–53	Y (always; week-of-month does not exist)
D	day	1–31 / 1–92 / 1–366	M (default) · Q · Y
E	ISO weekday	1 (Mon) – 7 (Sun)	week; with # ordinal: M (default) · Q · Y
T	time of day	hhmm[ss[.sss]] half-open range	each covered day
H	hour of day	0–23	day
m	minute of hour	0–59	H
s	second of minute	0–59	m
(none)	date literal	YYYYMMDD[Thhmm[ss]]	absolute: bounds (§6) and cadence anchors (§5.2)

Notes:

• A designator is scoped only by the units its table row lists; never by whatever coarser selector happens to be present. D’s scope is the nearest of M/Q/Y in the expression, else M: D40 Q2 = 40th day of Q2; D-1 Y* = last day of each year; plain D1 = day 1 of every month. “Nearest” is by granularity, never by textual position; an expression is an unordered intersection (§1), so Q2 D40 and D40 Q2 are the same expression, and in Q2 D40 M1 the M wins scope, making D40 out of domain: a syntax error.
• W scopes nothing. D-1 W* is the last day of the month (intersected with every week), not the last day of a week; that’s E7.
• M and Q are absolute within the year, never rescoped: M5 Q2 is May (May ∩ Q2), and M-1 Q1 is December ∩ Q1 = ∅, statically unsatisfiable, flagged per §9.1.
• When W is present, Y is interpreted as the ISO week-year for that expression (Dec 29 can be W1 of the next week-year; the pairing must be consistent). Only the Y selector’s own test uses the week-year; M and Q always test calendar fields, so on 2025-12-29 the expression M12 W1 Y2026 is satisfied (calendar December, week 1 of week-year 2026).
• Milliseconds exist only as T literal precision (T093015.250:…); there is no millisecond selector.
• T and H/m/s can coexist (they intersect like everything else), but pick one style: T for clock ranges, H/m/s for unit patterns and strides.

3. Component Values

<designator><value-part> where the value-part is one of:

Form	Example	Meaning
value	D5	day 5
negative value	D-1	1st from the end of the parent’s actual domain (last day: 28/29/30/31 as the month dictates; leap years resolve themselves)
range	M3:7	months 3 through 7, inclusive
open range	D-7:*	7th-from-last through end (last 7 days); * = domain edge (§3.1)
wrap range	M11:2	November through February, each year: start → domain edge, plus domain start → end
list	M1,3,7:9	union of values/ranges
all	M*	entire domain (explicit “every”)
exclusion	M!5,7:9	domain minus the listed set
stride	M1/3, M1/5/2	recurrence (§5.1)
ordinal (E only)	E7#2, E7#-1	nth / nth-from-last occurrence within the scope

Rules:

• Exclusion ! appears only immediately after the designator and negates the whole value list. That is the only position. “Months 1–10 except 5” is written explicitly: M1:4,6:10.
• A component is either an exclusion or carries a stride, never both.
• A range whose start exceeds its end wraps within the parent instance (M11:2, H22:6, D25:5), same model as T’s midnight wrap (§4). Y is the one designator with no parent cycle to wrap around: years are absolute and non-cyclic (wrap is a cyclic domain repeating inside a parent instance — months within a year, hours within a day — and Y sits inside nothing), so Y2030:2020 is an error.
• Wrapping is decided syntactically, on literal non-negative endpoints only. On 0-based domains a literal 0 is an ordinary endpoint and wraps like any other: H22:0 covers hours 22, 23 and 0, the hour analogue of T2200:0100. A range containing a negative or * endpoint never wraps: it resolves per parent instance (§9.1) and covers nothing in any instance where the resolved start exceeds the resolved end. D-1:5 is therefore empty in every month, and when emptiness is statically certain, it draws the §9.1 unsatisfiability warning.
• Negative values are meaningless on Y (years are absolute, not counted back from a parent instance’s edge, §3.1) and are rejected.
• A literal 0 is out of domain on the 1-based designators and a syntax error there (M0, D0, Q0, W0 name no month, day, quarter or week) while on the 0-based H/m/s it is an ordinary value: H0 is midnight’s hour, m0 the first minute. A signed zero is never a value: the - requires a nonzero integer, so H-0 and M-0 are syntax errors everywhere; H0 is the only spelling of hour 0.
• Negative values are domain-checked symmetrically at parse time: -N … -1, where N is the domain’s maximum size, the count of its values (31 for month-scoped D, 24 for H; on 0-based domains this is not the same number as the largest value). D-31 parses — and covers nothing in months without 31 days (§9.1), the mirror of D31 in April — while D-32 is out of domain, a syntax error. At evaluation a negative v counts back from the parent instance’s actual edge, resolving to maxValue + 1 + v where maxValue is the instance’s largest value: D-1 in Feb 2024 is 29 + 1 + (−1) = 29; on H it is 23 + 1 + v, so H-1 is hour 23 and H-24 is hour 0.
• * as a bare list item is rejected (M1,*): a list containing the whole domain is the whole domain; write M*.
• Strides attach to a single value or a single range, not to lists; the stride’s start must be a non-negative value (an end-relative anchor like D-10:*/2 would shift per parent instance; write the positive equivalent). Only the start is so restricted: a negative end merely clips per parent instance while the phase stays anchored, so D2:-1/5 is valid. Wrap ranges take no stride.
• # takes a single ordinal in -5…-1, 1…5. Multiple ordinals (“1st and 3rd Friday”) use union: E5#1 | E5#3. The base value may itself be negative; it resolves against E’s fixed domain first, so E-1#2 ≡ E7#2 (the second Sunday).
• Omitted coarser components default to “every”: M3 alone ≡ M3 of every year. Writing Y* is legal, and redundant unless it rescopes a finer component: D and E# bind to the nearest coarser component present (§2), so D-7:* alone is the last 7 days of each month, while D-7:* Y* is the last 7 days of each year. Deleting a coarse * component is safe only when nothing takes its scope from it.

3.1 Domain, Scope, and *

Two ideas decide what every component value means; they are easy to conflate.

Domain: the set of values a designator can take. Some are fixed (M is 1–12, H is 0–23, Y is 1–9999), some depend on the calendar (D is 1–31 / 1–92 / 1–366).

Scope: which instance of the parent the domain is measured against, per the §2 table. D’s domain is the days of one month by default; D40 Q2 measures days against the quarter, so the domain becomes 1–92.

* is not a wildcard and never means “now”. It denotes the edge of the domain in force, resolved per parent instance; it moves with the calendar, never with the clock. Standalone M* (the whole domain) and M3:* (from 3 to the domain’s edge) are the same * doing the same job:

Expression	Scope	* resolves to	Covers
D25:*	month	last day of that month	Feb: 25–29 (5 days) · Mar: 25–31 (7 days)
D25:* Q2	quarter	last day of Q2	Apr 25 → Jun 30
M3:*	year	December	March through December
H*:3	day	0	hours 0–3
Y2020:*	(none)	9999	2020 and every later year
Y*:2020	(none)	1	2020 and every earlier year

Y’s edges are the format’s own limits: years 1 and 9999, the 4-digit ISO range the 8-digit YYYYMMDD date-literal grammar already commits to (§6). So Y2020:* reads “2020 through the end of the representable timeline”; still the same single meaning of *, resolving to the domain edge like everywhere else.

A range may put * at both ends: M*:* is accepted, each * independently resolving to its edge, and means exactly M*, which is the canonical spelling. (Bounds are the exception: an undesignated *:* is a syntax error, §6.)

Negative anchors have constant length; positive ones do not. A negative start counts back from the edge, so its distance to the edge never varies: D-7:* is always the last 7 days (Feb 23–29 · Mar 25–31), while D25:* is 4–7 days depending on the month. D-7 alone is a single day: the 7th from the end; the :* is what makes it a range. This is also why strides forbid end-relative starts (§3, Rules): D-10:*/2 would shift phase in every month of a different length.

4. Time of Day — T

Thhmm[ss[.sss]]:hhmm[ss[.sss]], half-open [start, end). Lists allowed: T0900:1200,1300:1800 (business hours minus lunch). A single value T12 means the implied unit interval [12:00, 13:00); T1230 means [12:30, 12:31); the unit is the literal’s precision, so T093015 means [09:30:15, 09:30:16) and T093015.250 one millisecond. In a range, an endpoint resolves to its literal’s span-start instant whatever its precision: T0900:12 ends at 12:00, not 13:00; only §6 bounds give a coarse endpoint its full span. A range with equal resolved endpoints (T0900:0900, and by the rule above T12:1200 too) is a syntax error: half-open, it would cover nothing; typo-shaped input fails loudly.

T takes only values, ranges and lists: no *, no exclusion !, no stride, no ordinal (T0900:*, T!0900:1200 and T0900:1800/2 are all syntax errors). A day-complement is written explicitly: not T!0900:1200 but T0000:0900,1200:2400.

Midnight wrap: T2200:0600 is legal and wraps within each covered day: ≡ T0000:0600,2200:2400. It does not spill into the next day.

Gotcha: because the wrap stays within the day, T2200:0600 E5 means Friday 00:00–06:00 and Friday 22:00–24:00, not “Friday night into Saturday morning.” For the latter, write the intent: T2200:2400 E5 | T0000:0600 E6.

Hour 24: midnight-as-an-end is written 2400, valid only as the exact 4-digit token in range-end position. T0000:2400 is the whole day; “22:00 to midnight” is T2200:2400. No other spelling exists: the 2-digit T22:24 and T2300:24, the seconds-precision T0000:240000, and 2400 as a range start are all syntax errors. Hour 24 is not a value, only an edge: one token, one position.

5. Recurrence

Two constructs, split by one question: does the pattern look identical in every parent cycle?

5.1 Stride — Calendar-locked Recurrence

<start>[:<end>]/<interval>[/<duration>] on a normal selector. From start, every interval-th unit, covering duration units (default 1), stopping at end (default: domain edge).

Expression	Covered
H0/4	hours 0, 4, 8, 12, 16, 20; every day
M1/3	Jan, Apr, Jul, Oct; every year
M1/5/2	Jan–Feb, Jun–Jul, Nov–Dec (2-month block every 5)
M1:6/2	Jan, Mar, May (stride stops at June)
Y2020:2040/3	2020, 2023, … 2038
Y2020:*/3	2020, 2023, 2026, … (open-ended, anchored)

Validity:

• 2 ≤ interval ≤ parent-domain size (its maximum, for variable domains: 31/92/366 for D by scope, 53 for W). Larger patterns (“every 14 months”) must use a cadence; a stride cannot drift. Y has no parent cycle to fit inside and therefore no interval ceiling: Y2020:*/100 is valid.
• 1 ≤ duration < interval (equal would be continuous coverage; write a plain range instead).
• The range start is the anchor and must be explicit: Y*/3 is a syntax error (every 3rd year from when?). This kills the epoch problem at the grammar level.

The first number after / is always the interval; the optional second is the duration. (Same order as cadences.) A stride is pure sugar: H0/4 ≡ H0,4,8,12,16,20.

5.2 Cadence — Anchor-based Recurrence

<date>/<n><unit>[/<n><unit>], a component that starts with a date literal: from this anchor, repeat every period, covering duration each time (default: 1 of the period’s unit; …/14M covers 1 month per recurrence, …/10D covers 1 day). Period/duration units: Y M W D H m.

Expression	Covered
20180301/14M	2018-03, 2019-05, 2020-07, … (month drifts; no calendar field can express this)
20200106/10D	Jan 6, 16, 26, Feb 5, … (the thing cron cannot say)
20200106/10D/3D	Jan 6–8, 16–18, 26–28, … (3 on, 7 off)
20200106/2W	every other week from a Monday (biweekly)

Validity:

• The anchor is part of the literal; an anchorless cadence cannot be written.
• The anchor must be a real calendar date: 20230229/1Y is rejected outright.
• Month/year periods use constrain overflow arithmetic (§9.2): 20240131/3M/1D covers Jan 31, Apr 30 (constrained), Jul 31, Oct 31.
• duration < period, compared conservatively with fixed unit lengths: max length (duration side) Y = 366 d · M = 31 d · W = 7 d · D = 1 d; min length (period side) Y = 365 d · M = 28 d · W = 7 d · D = 1 d; H/m exact. The check is duration × max(durationUnit) < period × min(periodUnit); same-unit comparisons are exact. The conservatism is deliberate: 20240101/2M/58D is rejected (58 ≥ 2 × 28) even though no real pair of consecutive months is shorter than 59 days. Consequence: a period of exactly 1 unit (/1M, /1D) requires an explicit smaller-unit duration (/1M/1D); the default would equal the period, and continuous coverage is a bound’s job, not a cadence’s.
• Month/year duration units require a month/year period (/3Y/2M ✓ · /10D/1M ✗): a calendar-month of coverage needs a calendar anchor to constrain against (§9.2), and a day-anchored recurrence supplies none.
• At most one cadence per expression; union more via |.

Cadences intersect with other components like anything else:

T0900:1200 20200106/10D     09:00–12:00 on every 10th day
E1 20180301/14M             Mondays inside each 14-month recurrence window

Maps 1:1 to ISO 8601 repeating intervals (R/2018-03-01/P14M) and to RRULE DTSTART + INTERVAL.

6. Absolute Bounds — Date Literals

An undesignated date-literal range clips the whole expression to an absolute window. Date literals are always 8+ digits: YYYYMMDD[Thhmm[ss]] (no bare YYYY/YYYYMM; that’s what Y/M selectors are for).

Form	Meaning
20150101:*	on and after 2015-01-01
*:20291231	through end of 2029-12-31, inclusive
20180301:20190425	from start of Mar 1 2018 through end of Apr 25 2019
20180120	that entire day
20180120T1800	that instant as a lower/upper edge inside a range

A bare date literal denotes its whole span (a day; a minute if Thhmm given; a second if Thhmmss given); a range covers from the start of its first literal through the end of its last. Note the end is span-inclusive: *:20180120T1800 runs through the end of the 18:00 minute; for a strict “before 18:00,” write *:20180120T1759. A range runs backwards (a syntax error, not an empty coverage) when the first literal’s span begins at or after the end of the last literal’s span: 20180120T1801:20180120T1800 is rejected, because 18:01:00 is exactly the end of the 18:00 minute-span, leaving no window (the empty case fails loudly rather than silently covering nothing). The check is decidable at parse time on the naive spans, zone-free: DST can stretch or shrink the realized window but never invert a naive ordering. Equal literals are valid and cover that literal’s whole span (20180120T1800:20180120T1800 is the entire 18:00 minute) and mixed precision is valid in either order, the coarser literal contributing its full span: 20180120T180000:20180120T1800 covers [18:00:00, 18:01:00). At most one bounds component per expression. At least one endpoint must be a date literal: *:* is a syntax error; an unbounded window is spelled by omitting bounds.

Bounds resolve in the evaluation zone, like every other component: 20180120 is that calendar day as observed in the evaluation time zone, so the window’s absolute position shifts with the zone. “Absolute” means the window does not recur, not that it is pinned to UTC; the expression still carries no zone (§1).

7. Union — |

expr | expr unions two complete expressions. Use it for anything a single expression’s “one designator once, one ordinal, one cadence” rules disallow:

E5#1 M* | E5#3 M*            1st and 3rd Friday of every month
T0900:1800 E1:5 | T1000:1400 E6    weekday hours, plus short Saturdays

8. Grammar (EBNF)

dtrexp      = expression , { "|" , expression } ;
expression  = component , { { " " } , component } ;
component   = selector | cadence | bounds ;

selector    = designator , ( "*" | ordinal | valuelist | exclusion | strided ) ;
designator  = "Y" | "Q" | "M" | "W" | "D" | "E" | "T" | "H" | "m" | "s" ;
valuelist   = item , { "," , item } ;
item        = value | range ;
value       = [ "-" ] , integer | timeval ;            (* timeval only under T *)
range       = ( value | "*" ) , ":" , ( value | "*" ) ;
exclusion   = "!" , valuelist ;
strided     = ( value | range ) , "/" , integer , [ "/" , integer ] ;
ordinal     = value , "#" , [ "-" ] , integer ;         (* E only; single, never in a list *)

cadence     = date , "/" , integer , unit , [ "/" , integer , unit ] ;
unit        = "Y" | "M" | "W" | "D" | "H" | "m" ;

bounds      = date | ( date , ":" , ( date | "*" ) ) | ( "*" , ":" , date ) ;
date        = 8digit , [ "T" , 4digit , [ 2digit ] ] ;

timeval     = 2digit | 4digit , [ 2digit , [ "." , 3digit ] ] ;

Tokens match greedily (longest match): 20180120T1800 is always one date-with-time literal, never a bounds 20180120 followed by a selector T1800; M3Y2018 parses as two components with the whitespace elided. The date literal’s T-glue is unconditional: a T immediately following 8 digits belongs to the literal, so a malformed time-part is a syntax error: 20180120T09 does not re-tokenize as bounds 20180120 + selector T09. Greed never backtracks on semantic failure. The time part of a date literal is an ordinary wall-clock value: hh ≤ 23 and mm/ss ≤ 59; hour 24 is a T-range token (§4) and never appears in a date literal (20180120T2400/1D is a syntax error).

(Static validity rules — domains, stride limits, one-designator-once, W⇒ISO-week-year, etc. — are constraints on top of this grammar, per §§2–6. They apply to the completed expression, not token-by-token: components are unordered, so D40’s domain check must wait until it is known whether an M/Q/Y scope is present; Q2 D40 is valid, and a strict single-pass check would wrongly reject it.)

9. Evaluation Semantics

Every component denotes a set of half-open instant intervals; the expression denotes their intersection; | unions expressions. Formally, for an instant t evaluated in time zone z (an IANA identifier; default UTC, §1):

1. Compute the calendar fields of t in z once: year, ISO week-year+week, quarter, month, day-of-month/quarter/year, weekday, weekday-ordinal-in-scope, time-of-day, hour, minute, second.
2. Each selector tests its field against its normalized span set (inclusive surface ranges become half-open integer spans at parse time; negative values resolve against the actual parent instance: D-1 in Feb 2024 is 29). Strides test (value − start) mod interval < duration within the range.
3. A cadence covers t iff t lies inside one of its occurrence windows. Occurrence k (k ≥ 0) starts at startₖ = constrain(anchor + k·period) (§9.2) and ends at constrain(startₖ + duration); the end is measured from the constrained start, not from the ideal anchor + k·period + duration (for 20240131/3M/1M, occurrence 1 runs Apr 30 → May 30, not May 31). For D/W/M/Y periods this arithmetic is naive local-calendar and the window is a local wall-clock interval: membership compares t’s local fields against the naive window; the anchor is never resolved to an instant (§9.3). For H/m periods the anchor resolves to a single instant (§9.3) and windows are absolute elapsed time. For the exact units (D/W/H/m) this reduces to unitsBetween(anchor, t) mod period < duration; for month/year periods the window start must be computed by constrained anchor arithmetic; the reduction is wrong there (20240131/1M/1D covers Apr 30, which no mod on monthsBetween yields). k is still locatable in O(1) via floor(unitsBetween(anchor, t) / period) ± 1.
4. Bounds test t against the absolute window.
5. covers(t) ⇔ every component passes; a |-union covers ⇔ any branch covers.

Cost: covers is O(#components) integer comparisons after one field extraction, no date object iteration. This is the property that makes DTRExp suitable for hot paths (e.g. per-request permission checks).

Derived operations: intersect(a, b): the covered intervals clipped to a finite window (always a finite list); next(after): first covered interval after an instant (candidate-stepping search, coarsest selector first).

9.1 The Existence Rule

Selectors match against the fields of real instants, so a calendar position that doesn’t exist in a given parent instance covers nothing there, no error, no substitution:

• D29 M2 covers Feb 29 in leap years and nothing in other years. (2100 is not a leap year: the century rule; test vectors must include it.)
• D31 covers nothing in 30-day months; D366 and W53 cover nothing in years that lack them; E7#5 covers nothing in months without a fifth Sunday.
• Ranges resolve per parent instance: D25:-1 is days 25–28 in Feb 2025 but 25–31 in January. An instance where the resolved start exceeds the resolved end covers nothing there.
• validate() SHOULD flag statically unsatisfiable expressions as warnings; they are not syntax errors. D30 M2 can never match in any year; neither can M-1 Q1 (M-1 is December, §2, and December ∩ Q1 is empty). The required minimum (pinned by the warning vectors) is: per-selector satisfiability against the set of domain sizes implied by co-present selectors (catches D30 M2, W53, and its day-of-year twin D366, under concrete years that lack them, and statically-empty non-wrap ranges like D-1:5 and M-2:2; a stride is sugar over a range and participates the same way; D30:*/2 M2 warns), full-domain exclusions (M!1:12), and M∩Q disjointness (M-1 Q1). Concrete-year satisfiability (W53 Y…, D366 Y…) is required only over closed year spans no wider than 1,000 years; an open (*) or wider span MAY stay quiet. When W is present, Y is the week-year (§2) while day-of-year stays calendar, so D366×week-year satisfiability is cross-selector territory and MUST stay quiet; D366 W1 Y2025 really covers Dec 31 2024 (day 366 of calendar 2024, inside week-year 2025). Deeper cross-selector analysis (day×weekday interplay and the like) is quality-of-implementation, not conformance. The converse is also conformance: the quiet vectors pin expressions that MUST parse with no warning; a linter that cries wolf is as non-conforming as one that stays silent. Warnings from every | branch surface on the whole expression; a dead branch is exactly the typo the warning exists for.
• To pin a calendar date, use D+M, never day-of-year: D60 Y* is Mar 1 in common years but Feb 29 in leap years.

9.2 Cadence Overflow — Constrain, Never Skip

When a month- or year-period cadence occurrence lands on a nonexistent date, it is constrained (clamped) to the last valid day:

• 20240131/1M/1D → Jan 31, Feb 29, Mar 31, Apr 30, … (2025: Feb 28)
• 20240229/1Y/1D → Feb 29 in leap years, Feb 28 otherwise

This matches Temporal’s constrain arithmetic. Interop note: plain RFC 5545 RRULE skips such occurrences; the equivalent behavior requires RFC 7529 SKIP=BACKWARD, which toRRule() must emit for affected cadences.

9.3 DST and Local Time

Evaluation in a non-UTC zone follows directly from instant-based coverage:

• Nonexistent local times (spring-forward gap): no instant ever carries those fields, so T0230:0300 covers nothing on that day, the existence rule again.
• Repeated local times (fall-back): both instants carry matching fields; both are covered. Local-time coverage can therefore total 23 or 25 hours on transition days; this is correct for the “local business hours” use case. The transition instant itself carries post-transition fields: Berlin’s 2024 fall-back instant 2024-10-27T01:00:00Z is local 02:00:00 CET, the first second of the second pass, not the last of the first.
• Cadences with D/W/M/Y periods use naive local-calendar arithmetic in the evaluation zone: occurrence windows are local wall-clock intervals, and membership is decided on t’s local fields; the anchor is a naive local date-time, never resolved to an instant. The two rules above then apply unchanged: a repeated local time inside the window is covered both times (a covered “day” on a fall-back date is 25 absolute hours), and gap times are covered by nothing (23 hours). The window duration follows the period’s arithmetic whatever its own unit: an H/m duration inside a calendar-period window is naive wall-clock time: 20241026T2300/2D/12H evaluated in Berlin ends at local 11:00 even across the 25-hour fall-back night (12 wall hours, 13 absolute). H/m periods are absolute elapsed time measured from a resolved anchor instant, and by the same rule, every duration inside such a grid is a fixed absolute length too (D = 24 h, W = 168 h): an absolute grid has no calendar anchor to constrain against, so 20200101/48H/1D covers exactly 24 elapsed hours per occurrence, DST or not.
• Exactly one construct requires a date literal to become a single instant: the anchor of an H/m-period cadence, whose absolute grid needs one phase point. There, a repeated local wall-clock time resolves to the earlier occurrence, and one that falls in a gap resolves forward past it: Temporal’s compatible disambiguation. The rule must not depend on the zone or the sign of its offset. Concretely, the gap is crossed forward, by its own length: in Europe/Berlin the 2024 spring-forward skips local 02:00–03:00 on Mar 31 (the transition is at 2024-03-31T01:00:00Z), so an H/m-period cadence anchored at 20240331T0230 (a wall-clock time that no instant carries) resolves to local 03:30 CEST = 2024-03-31T01:30:00Z: the literal’s fields read with the pre-transition offset and shifted past the gap by the one-hour gap width. It does not collapse to the transition instant itself (local 03:00 CEST, 2024-03-31T01:00:00Z), a plausible but wrong choice that would pin every gap anchor to a single point regardless of where in the gap it fell. Everything else, calendar-period cadence windows (previous rule) and bounds spans (§6), is a local wall-clock interval and never resolves: a bounds minute-span whose local minute repeats on a fall-back day covers both passes.

9.4 Leap Seconds

Not representable, following POSIX/Temporal: s runs 0–59 and T…60 is invalid. 23:59:60 never occurs in the evaluated timeline.

10. Examples

Definition	DTRExp
Business hours (Mon–Fri 9–18)	T0900:1800 E1:5
…with lunch break excluded	T0900:1200,1300:1800 E1:5
Last Sunday of April, every year	E7#-1 M4
Mothers’ Day (2nd Sunday of May)	E7#2 M5
Christmas Day	D25 M12
Last 7 days of every year	D-7:* Y*
First 15 days of each month	D1:15
Every month except July	M!7
3rd and 7th month of each year	M3,7
First 10 days of Q2	D1:10 Q2
Every Wednesday of March 2018	E3 M3 Y2018
Weekends from 2018-03-01 to 2019-04-25	E6:7 20180301:20190425
18:00–20:00 on 2018-01-20	T1800:2000 20180120
Every Monday 10–11 until 2018-01-20 18:00	T1000:1100 E1 *:20180120T1800
Every Friday the 13th	D13 E5
Every 4th hour, first 20 minutes (Leonardo da Vinci’s sleep, per the legend)	m0:19 H0/4
Jan/Apr/Jul/Oct, every year	M1/3
Every 3rd year, 2020 through 2040	Y2020:2040/3
Every 14 months from March 2018	20180301/14M
Every 10 days from 2020-01-06, 3 days long	20200106/10D/3D
Biweekly from a given Monday	20200106/2W
1st and 3rd Friday of every month	E5#1 | E5#3
First 10 ISO weeks of each year, 2015–2029	W1:10 Y2015:2029

11. Open Questions (for draft 3)

1. W + M co-occurrence: currently legal by intersection (W10 M3 = the part of ISO week 10 that lies in March). Confirm or forbid.
2. Ordinal # on units other than E: e.g. D15#… makes no sense, but W#-1 M* (“last full week of each month”) is a recurring wish. Deferred; would reintroduce week-of-month semantics.
3. Cadence duration in a different unit than the period (/10D/36H): currently allowed with a conservative max≤min length check; may restrict to same-unit if the check proves confusing.
4. describe() locale model: deferred to the library spec.

12. Conformance

An implementation is conforming iff it accepts/rejects and evaluates the shared test vectors (vectors.json, shipped with this draft): { expression, tz, instant → expected } coverage groups, plus rejection cases, warning cases, and quiet cases: expressions that must parse with no warning. The vectors, not the prose, are the contract. A recommended library surface (what to name parse, covers and friends) lives in API.md; it is informative, never a conformance requirement.

The suite includes the calendar traps of [§9.1](#91-the-existence-rule)–[§9.3](#93-dst-and-local-time): D29 M2 against 2023, 2024, 2000, and 2100; D-1 across leap February; W53; E7#5; 20240131/3M/1D and 20240229/1Y/1D constrain cases; midnight-wrap + weekday intersection; the hour-24 spellings (the valid T…:2400 and every rejected variant); and DST-transition instants in Europe/Berlin (gap and overlap), including cadence occurrence windows across both transitions, the H-period anchor in the fall-back overlap and in the spring-forward gap, and an H-duration window across the 25-hour fall-back night.