internal/pathrule checks the paths and the texts of a format 3 head
(spec 29.5, 29.6) with its own tables, never with the Unicode functions
of the platform, whose version changes with each runtime.
- gen reads the 19 pinned data files (UnicodeData, DerivedCoreProperties,
CaseFolding and emoji-variation-sequences of Unicode 18.0.0, and the
15 WindowsBestFit tables), checks their SHA-256 and writes tables.go:
assigned code points, Default_Ignorable_Code_Point, full canonical
decompositions and combining classes, C and F folding, the bases of
the emoji variation sequences, and the non-ASCII code points each
code page maps to ASCII. The data files stay out of git, in .cache.
- NFD, Fold and the key of R7; CheckPath with R2 to R6c and R10,
CheckTree with R7 and then R9, and CheckComment and CheckAuthor with
the invisible-character rule. Errors name the rule and never echo the
creator's text, so that another implementation can match them.
- The canonical text of the tables has a SHA-256, TablesDigest, which
the tests recompute and a TypeScript implementation will share.
- Checked against golang.org/x/text (Unicode 15.0.0) outside this
module: NFD matches on every code point both know, and folding only
differs on the 86 Cherokee letters that CaseFolding.txt folds to upper
case, as these tables do.
- The spec pins the SHA-256 of the 19 files in 29.5.1.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The author approved it on 30-09-2026, after Unicode 18.0.0 warned that
sequences of invisible variation selectors are used to attack AI
applications. The comment and the declared author allowed the tag
characters and loose variation selectors, which hide text a model
reads, and paths allowed runs of ZWJ, ZWNJ, VS15 and VS16.
- R4b: VS15 and VS16 only right after a character that
emoji-variation-sequences.txt pairs with that selector, and ZWJ and
ZWNJ never first, last or twice in a row in a segment.
- Section 29.6: the comment and the declared author carry no
Default_Ignorable_Code_Point except that whitelist, placed as R4b
says, each line standing for a segment.
- The tables add emoji-variation-sequences.txt; section 76 records the
change as number 12, with mutations and vectors.
- The five new references of section 77 regain the two trailing spaces
that break their line, like the others.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The author chose Unicode 18.0.0, released on 16 September 2026, over
the 17.0.0 of the design: the tables freeze with capsule format 3, so
an older version would refuse for good characters that are already
standard. Section 29.5.1 now says the Unicode version is fixed with
the format, and that moving to another needs a new format (22).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- R6c only rejects a best-fit projection with '/', '\', ':' or U+0000,
and still applies R3, R5, R6 and R6b to it: bestfit1250 maps "¿" to
'?' and bestfit874 maps "§" and "♥" to C0 controls, so the stricter
rule would have refused ordinary Spanish names.
- Step 17: codes other than ERR_INTEGRITY are reported only after
reading PAYLOAD_AGE to EOF; 17.1 leaves the padding to 17.8; 17.3 and
17.6 never fail. The precedence case is the cut right after a
complete chunk, where filippo.io/age and age-encryption differ.
- Verdicts: the first matching row decides, and alg or seal_type are
read only from content that meets its schema. Sections 58 and 70
exempt SECURITY_CBOR, which only changes verdicts.
- Normative verdict texts, and the presentation rule for any text
output, paths included.
- The sink rule of section 70, the test-vector exemption of writer
rule 13 and the mtime rule 16.
- More mutations and vector coverage; a complete list of changed test
data in section 76 (checked: only two of the 125 mutations and none
of the 4380 differential cases leave VERSION 3); corrected cases.
- A closed list of the Unicode and WindowsBestFit tables, and
editorial fixes.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Delivery 1 of capsule format 3, as the author decided on 30-09-2026:
several files with their paths, sizes, SHA-256 and dates, encrypted in
the plaintext of PAYLOAD_AGE, and a security area that later versions
will fill with an author signature and a timestamp seal without
changing the format.
- VERSION 3 and CONTROL_CBOR schema 3; writers write only format 3,
readers open the three formats.
- BODY with a 12-byte frame (AREA_LEN, SECURITY_LEN, HEAD_LEN), the
security area fixed by spec version (512 bytes here), the head and
the files (new sections 29.2 to 29.7).
- The head is always version 1 in format 3; path rules R1 to R10 on
pinned Unicode 17.0.0 and WindowsBestFit tables; text rules; the
verdicts X, F0, F1, S0, S1 and S2 and their presentation.
- Step 17 split into 17.1 to 17.8 with its precedence, and the new
code ERR_HEAD_INVALID.
- Atomic delivery of several files (56), limits (57), writer rules 13
to 18, mutation tests, fixtures and the change log in 76.
- The CDDL gains control-v3, security, author-signature, seal, head and
file.
The reference implementation still implements v0.9. The design, its
reviews and the author's decisions are in the private docs repository,
spec_v0.10/formato3_diseno.md.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The reference moves to the DateKeys Protocol Specification v0.9, approved
by its author on 29 September 2026. Encrypt writes capsule format 2 only;
Open and Inspect read formats 1 and 2, and a format 1 capsule keeps the
verdict v0.8.2 gave it.
Format 2 (spec §22, §29.1, §31, §39):
- VERSION in the PRELUDE is the capsule format, capsule.Format; any other
value is ERR_UNSUPPORTED_VERSION at step 2.
- CONTROL_CBOR has the schema version of its format. Version 2 adds key 6,
payload_length (8 bytes, big-endian, at most L_MAX = 2^53 - 2^46), and
key 7, padding (1 bloque256, 2 reforzado); it is 103 bytes without
extensions, whatever L.
- The payload is the content padded with zeros to P = rule(L). Step 17
checks the length and the zeros, and Open writes only the first L bytes.
- INNER_ACCESS_AGE holds exactly 16 X25519 stanzas: 1 to 16 credentials,
and a dummy in each slot left, in a uniformly random order.
Writer rules (spec §62.1): EncryptOptions.Length is required and the
source must deliver exactly that many bytes; recipients that are not
canonical or of low order are rejected (agewrap.CheckX25519Recipient);
self-checks of the header, the control, INNER_ACCESS_AGE and PAYLOAD_AGE.
The CLI measures its input, takes -padding and reports the format.
Test data: seven format 2 fixtures, padding vectors checked against
math/big, format 2 CBOR vectors, and the mutation corpus in both formats
with the 22 cases of the third list of spec §64, built without randomness
by sealing the fixtures again with their known keys and nonces. The
format 1 fixtures are kept byte for byte and never regenerated; the
differential corpus keeps its 1825 cases and adds a block per format 2
fixture. The spec copy loses its "to be implemented" markers, and the
READMEs, CHANGELOG, traceability and testdata/README.md follow v0.9.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The final review of the draft found 22 problems, and none had been applied
yet. All of them are applied now, together with four places that repeated
them (§62.1 rules 1 and 4, §76 changes 4 and 10).
- §29.1, §76 change 4: 32-bit arithmetic first gives a wrong P at
L = 2 113 929 217 with signed operators and at L = 4 227 858 433 with
>>> 0, not at 2^32 + 1. The vector table gains a row for each, checked
against a BigInt Padme. Above 2^32, Padme exceeds bloque256 except at
L_MAX. The Node log2 error changes E and lastBits, not P or S.
- §56, §76 change 4: a reader MUST NOT present the content as valid before
step 17 ends; a streaming reader MUST NOT write the padding and MUST
signal the step 17 error so that what it wrote is discarded. The author
chose this over the stricter rule, which forbade delivering any byte
before step 17 and so the streaming Open(dst) of the reference.
- §64: the 15 and 17 stanza mutations recalculate the PRELUDE and
header_binding; every time_and_key mutation offers the identity, because
any change breaks the capsule_digest of a .dkk; the duplicate recipient
mutation names its identity. Three new mutations cover the shape of
payload_length (7 or 9 bytes, a CBOR integer); the format 2 cbor.json
vectors of §76 list them too.
- §39, §62.1 rule 4: the official test vectors may show which slots are
dummies. §70, §62.1 rule 1: the format 2 rule binds implementations that
write capsules, and a test vector generator MAY write format 1.
- §62 step 8 gets the 64 MiB limit of §61 step 6.
- Accuracy: §22 (an older reader detects a new padding code only after the
network request; the .dkk schema is §41), §37 and §76 change 10 (the §63
rules do not detect those recipients), §55.2 (the writer knows the number
of parties; relays are §48), §76 changes 1 and 8.
- Wording: §62.1 rules 7 and 11, §63 step 4, §76 change 11, the field names
in the CDDL comments, and a v0.9 entry in spec/README.md.
go test ./... passes.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Draft decided by the author on 29-09-2026: format 2 with exactly 16
X25519 stanzas in INNER_ACCESS_AGE (dummies, shuffled), payload padding
(code 1 bloque256, code 2 reforzado = max(bloque256, Padme)) with L and
the code in CONTROL_CBOR v2, VERSION 2 so v0.8.2 readers reject early,
a privacy section (§55.2) and writer rules (§62.1). The CDDL holds the
draft schemas. The fixes from the final review were being applied when
work stopped: they may be partial. See App/docs/HANDOFF.md.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The second round of the formal review confirmed the nine corrections of
c57ed48 and asked for these, recorded in §76 as corrections 4 to 6 and
an editorial note:
- §72: an encoder MUST NOT write a registered extension in an object or
array it is not registered for; §54: a reader MUST NOT interpret the
data of a noncritical one it ignores for that reason. capsule.Encrypt
and accesskey.Encode take no Registry, so the application applies the
rule; their documentation and extension.Placement say so.
- §17 and §51 give the step-10 codes only for a directly supplied
release, as step 10 does; a network source discards a failing one at
step 9.
- Step 9 reports ERR_RELEASE_UNAVAILABLE and no other code, whatever the
failure of the source. provider/drand.Client keeps each relay's failure
as text only (errors.Join made a relay's ERR_ROUND_MISMATCH match with
errors.Is), and capsule.Open keeps only the text of a source error that
carries another code (a caller's source failing with
ERR_RELEASE_INVALID gave that code at step 9). A context that ended
stays detectable: Fetch now has a single failure path, so the canceled
and deadline cases are deterministic.
- TestExtensionPlacement covers the noncritical array of a .dkk: with
the object-blind extension.CheckNoncritical at step 9.a it fails.
- Editorial: §28.1 "analizan solo la cabecera age", one arrow at step 9,
two §76 introductions; §73 lines for release sources and placement.
- testdata/README.md says the corpus registers its extensions in both
arrays of every object; traceability, CHANGELOG and both READMEs
(integrity holds against whoever lacks the file keys, §27, §55.1)
follow. Spec dated 28 September 2026; new SHA-256 in spec/README.md.
No fixture or vector changes.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A formal review of the whole v0.8.2 text found it approvable after
these corrections, recorded in §76 ("Correcciones de la revisión
formal"):
- §27 no longer calls header_binding the authenticity of PUBLIC_HEADER:
it binds the header to the opened control, never authorship or date
(§55.1); the age MAC only protects against whoever lacks the file key.
- §63 steps 9 and 10: a network source (relay, Release API, cache) MUST
verify every response and gives ERR_RELEASE_UNAVAILABLE at step 9 when
none verifies; the step-10 codes are for a directly supplied release.
The reference already behaved so; TestReleaseFromANetworkSource pins
both paths.
- §54 and §72: registrations declare the objects and arrays where an
extension may appear, and a known extension out of place counts as
unknown there. The reference gains the optional extension.Placement
interface, used at steps 4, 9.a and 14.
- §63 step 11 fixes the GT serialization hashed by H2 (kilic/kyber order)
with the frozen vector H2(e(G1, G2))[:16] = cb87319f..., shared as
testdata/vectors/tlock_ibe.json; H2-H4 are cited to drand/kyber.
- Step 5 makes the SEALED_CONTROL read mandatory, step 15 names
ERR_HEADER_BINDING, §21 makes capsule_id 16 CSPRNG bytes a MUST, §76
is made accurate (four dk1.json vectors, the §36 time_only rule, two
cases rewritten against the texts that really existed), and editorial
fixes in §5, §36, §55.1, §69.1 and §77. §73 lists the three new
decisions.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Amendment of the unreleased v0.8.2, recorded in §76 with its case: the
second implementation's phase-2 research found that tlock-js over
@noble/curves 1.9.7 accepts U re-encoded as c0 + p and a signature
x + p and returns the same file key, while the reference rejects both
(noble 1.9.7 differed from kilic on 5,615 of 41,686 encodings), and the
spec did not say which encodings are valid.
- §12.2 defines the canonical encoding of a BLS12-381 point (drand's
compressed ZCash form) and requires decoders to reject every other
byte string; §12.1 applies it to public_key.
- §63 step 10 applies it to the release signature (ERR_RELEASE_INVALID)
and step 11 defines the tlock stanza body U || V || W (96 + 16 + 16
bytes for Quicknet) with a canonical, non-infinity U (ERR_INTEGRITY).
- §64 gains ten mutations, exported to mutations.json (65 cases). The
signature x + p case uses published Quicknet round 1004, the first
after 1000 whose x allows x + p < 2^381. The reference already gave
every stated code and step.
Errors no longer copy text from tlock, kyber, age, drand or
kyber-bls12381. kyber's IBE error carried the candidate plaintext and r,
and with one bit of W flipped the message disclosed the real tlock file
key with that bit flipped. Every such place now uses a fixed reason with
its normative sentinel; TestTlockFailureDiagnosticsCarryNoSecrets fails
with the old wrapping.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Approved refinements, each recorded with its reproducible case in the
§76 v0.8.2 subsection:
- §69.1: layered error model with normative precedence (frame, type tag
and version, CBOR profile and CDDL, then fields with their own code in
ascending key order; across steps the §63 order decides), with a scope
paragraph for the optional steps 5, 6 and 8.
- §55.1: normative trust table per section (who can write it, from which
step it is bound, what it never proves); §72: security-relevant claims
go in CONTROL_CBOR or under a signature, .dkk data is advisory.
- §31/§54: extension arrays in strictly ascending unsigned byte order of
extension_id (one rule for order and uniqueness).
- Gaps a second implementation needed: §28.1 malformed age headers,
§15/§19 latest unlock time and dk1_ reading rules, §22/§23/§57 length
lower bounds, §63 step 8 tlock argument comparison and step 9 order,
§12.1 profile validation with the drand chain-hash formula, §74 table
of implementation limits.
Reference alignment: .dkk errors only at step 9.a (new
OpenOptions.AccessKeyFile, used by the CLI), CR/LF in dk1_ is
ERR_DATEKEY_INVALID, BODY_LEN 0 is ERR_INTEGRITY, nil identities are not
credentials, and AccessIdentity tries every identity on every stanza so
its verdict does not depend on their order. dk1.json gains three
vectors; every other testdata file is byte-identical.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Step 2b: package codec is rewritten without reflection or struct tags. A
strict Decoder accepts only the spec §58 profile, Unmarshal decodes,
re-encodes and compares, Peek reads the type tag and version, and Walk is a
bounded iterative helper for vectors and fuzzing. Every schema has its own
hand-written encoder and decoder that checks all CDDL rules before the
fields with their own error codes. github.com/fxamacker/cbor/v2 and
github.com/x448/float16 are gone; nothing replaces them. Valid objects
encode and decode exactly as before (1.34 million differential verdicts);
the invalid-input differences are documented in CHANGELOG and
traceability decision 12. A review found and fixed an access_policy check
that truncated to uint8.
Step 3: testdata gains vectors/cbor.json (generic and per-schema CBOR
vectors), vectors/mutations.json (the 55-case mutation corpus, replayable
offline), vectors/inspect_differential.json (1,825 fixed-seed mutations
with the Go verdict) and one inspect -json golden per fixture, all
regenerated byte-identically by genfixtures and documented in
testdata/README.md for second implementations.
Gate green with 90 s of fuzzing per target on all 15 targets; codec at
100 % coverage; pre-existing testdata byte-identical.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Extension data becomes a non-empty opaque bstr bounded by the frame of its
container; the base protocol never decodes or validates it. Arrays hold at
most 64 extensions, extension_version is bounded to 2^32-1 and the profile's
period and genesis_time to 2^53-1. The multiplicity exception is removed,
ERR_EXTENSION_DATA_INVALID is added, the §57 limits become MUST with an
explicit error mapping, §58 names the protocol's CBOR profile and §72 sets
the registration rules. §76 records the six reproducible cases behind the
change. The implementation follows in the next commit.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The module is imported as g.activething.com/go/DateKeys, the path the
project's Gitea advertises. The .github directory is gone: workflows
now live in .gitea/workflows, use the gitea.com action mirrors and
install every tool from its Go module; releases go to this Gitea with
goreleaser and a key-based cosign signature; Dependabot is replaced by
a nightly report of available updates. scripts/check.sh runs the same
checks on any machine and is the gate while the server has no runner.
SECURITY.md, README and CONTRIBUTING no longer refer to GitHub.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Reference implementation in Go, built from the implementation plan
(milestones M0 to M5): datekey, profile, provider, codec, agewrap,
extension, capsule, accesskey, the datekeys CLI, official vectors and
fixtures, the mutation corpus, fuzz targets, interop and live tests,
CI workflows, traceability and policy documents.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
dev