Create/Get for submission (including the one legitimate status
transition — submission is not append-only, unlike the other three),
Create/GetLatest for extraction and evidence_package, Create for
finding plus ListCurrentFindings which applies the anti-join documented
in the migration (a finding referenced by another row's `supersedes`
is not "current").
Tested against real Postgres, including that the append-only trigger
still rejects UPDATE on evidence_package via this new code path, and
that ListCurrentFindings actually hides a finding once a correction
supersedes it.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Routing, html/template layout/content pattern, and the Pre-Publish
check flow: POST /pruefen runs extraction (Stufe 1) then rules.Evaluate
(Stufe 2) and renders the result as an htmx fragment. Nothing is
persisted yet — that's the next step (wiring internal/store in).
The needsClarification case is rendered explicitly as a request for
more information rather than "no findings", matching the core
principle. Every result carries the legal-advice disclaimer required by
CLAUDE.md's guardrails.
Server depends on a narrow Extractor interface rather than *extract.
Client directly, so tests inject a fake instead of calling the real API
— internal/web's test suite never touches the network. htmx is vendored
locally (internal/web/static/htmx.min.js) instead of loaded from a CDN,
keeping the UI usable without runtime internet access.
cmd/deklarix/main.go now wires all of this together: reads
ANTHROPIC_API_KEY (required) and RULES_DIR (default "rules"), builds
the extract client and loads the rule set, and serves web.Server instead
of the old inline health-only mux.
This exposed the same crash-loop risk fixed earlier for DATABASE_URL:
postinst's start guard only checked DATABASE_URL, so a fresh install
would now crash-loop on a missing ANTHROPIC_API_KEY instead. The guard
checks both. scripts/build.sh also now ships rules/*.yaml into the .deb
under /usr/share/deklarix/rules (not a conffile — rules are updated via
the release pipeline, never hand-edited on a server), and
deklarix.env.example points RULES_DIR there by default.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Content assembly (BuildContent) is separate from PDF drawing (Render),
so the actual business logic — what goes into the evidence dossier, in
what form, with which mandatory fields — is unit-testable without
parsing PDF bytes. BuildContent refuses to produce a dossier missing
its evidentiary fields (timestamp token, asset/metadata hash, platform)
rather than emitting one with silently empty proof sections. Every
dossier carries the "this is not legal advice" disclaimer required by
CLAUDE.md's guardrails.
Uses github.com/go-pdf/fpdf (actively maintained fork of jung-kurt/
gofpdf, no dependencies beyond the Go stdlib) for rendering. Its core
fonts use cp1252 internally, so a small cp1252.map (copied from the
fpdf module, embedded via go:embed) drives UnicodeTranslator — German
umlauts render correctly without needing an external font file at
runtime, keeping Deklarix a single binary. Verified visually with
pdftotext/pdfinfo against a generated sample.
Tests build a real, structurally valid RFC-3161 token offline (a
throwaway self-signed cert + timestamp.Timestamp.CreateResponse), so
BuildContent/Render/Generate are fully tested without hitting a real
TSA — unlike the network-gated integration test in internal/evidence.
Also adds evidence.TimestampTime(), extracted from the parsing logic
already used by the TSA client, since the dossier needs to show the
timestamped time to a human reader.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
HTTPTimestamper builds a Request directly from a precomputed SHA-256
digest (Request{HashAlgorithm, HashedMessage, Nonce}.Marshal(), not
CreateRequest(io.Reader) which would hash the input itself), POSTs it to
a TSA, and validates that the parsed response's hash algorithm, hashed
message and nonce actually match what was sent before accepting the
token — a response that doesn't match the request isn't a valid
timestamp for that hash, regardless of whether it parses.
Uses github.com/digitorus/timestamp for the RFC-3161/ASN.1 encoding
rather than hand-rolling it.
TSA choice (open point in CLAUDE.md): FreeTSA.org for now — free,
RFC-3161-compliant, verified end-to-end with a live smoke test, but
not eIDAS-qualified. Documented as needing an upgrade to a qualified
provider (D-Trust, Bundesdruckerei, ...) before real customer use, same
treatment as the open legal questions in rules/OPEN.md.
Tests cover request/response validation without network (hash size
guard, a fake server that parses and checks the incoming request,
non-200 and malformed-response handling) plus a real integration test
against FreeTSA gated behind DEKLARIX_TSA_INTEGRATION, mirroring the
DATABASE_URL-gated pattern in internal/store.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
SHA-256 over raw asset bytes and over the canonical JSON form of
metadata (encoding/json already sorts map keys and preserves struct
field order deterministically, so no separate canonicalization library
is needed for our own fixed types). A Timestamper interface stands in
for the RFC-3161 timestamp step — no concrete implementation yet, since
which TSA to use is an open decision (CLAUDE.md "Offene Punkte") that
directly affects the archive's evidentiary weight, not a purely
technical choice to make silently.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Rule now carries a required Jurisdiction (land) field, and Facts a
matching Jurisdiction supplied by the caller (like Platform — no
legal jurisdiction can be read off a caption or image, so the model
never guesses it). Evaluate() only lets a rule fire when its
jurisdiction matches the facts' jurisdiction.
This is the structural half of "deutsche Rechtslage zuerst, Struktur
für Österreich und Schweiz vorgesehen": a future AT/CH rule set can be
added as plain new YAML files without touching existing DE rules, but
no AT/CH content is added now — that needs its own legal research
first, same as WK-001/WK-004 needed for Germany.
WK-001 and WK-004 are tagged land: DE, all golden fixtures carry
jurisdiction: DE, and extract.Input passes Jurisdiction through
unchanged into the returned Facts.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Stufe 1 from the core principle: Client calls the Claude Messages API
directly over net/http (no SDK dependency, stays consistent with
"Go-Standard-Library wo möglich") and forces tool-use with a strict JSON
schema instead of parsing free text. Extract() returns rules.Facts
directly rather than an intermediate DTO, since producing exactly that
is the point of this stage. Platform is supplied by the caller, never
guessed by the model.
Every failure mode returns an error instead of a zero-value Facts:
network errors, non-200 API responses, a missing tool_use block, and —
critically — a gegenleistung value outside the four allowed enum
values, which would otherwise get silently coerced into a wrong fact.
Tested entirely against an httptest fake server, no real API calls.
Building this surfaced a real gap in internal/rules: Evaluate() treated
Consideration=="unklar" the same as any other value, i.e. it just
produced an empty finding list — indistinguishable from "everything's
fine". That contradicts the core principle (uncertain extraction should
trigger a user clarification, never a judgment). Evaluate() now returns
(findings, needsClarification), with a golden case covering it.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
internal/rules implements Stufe 2 from the core principle: the LLM
extracts facts, this deterministic engine judges them against versioned
YAML rules. Facts/Condition/Rule/Finding types, a loader that refuses to
load on a missing id/version or a duplicate rule id rather than silently
skipping a bad file, and Evaluate() matching facts against rules.
Two real rules grounded in verified research (see rules/OPEN.md for the
open questions that surfaced along the way):
- WK-001: no disclosure at all despite consideration (§ 5a Abs. 4 UWG,
§ 22 Abs. 1 MStV)
- WK-004: disclosure present but hidden behind a "mehr anzeigen" cut
(§ 5a Abs. 4 UWG, Leitfaden der Medienanstalten, LG Köln 12.05.2026)
The two are deliberately disjoint (WK-004 requires disclosure_present=
true) so a post with no disclosure at all doesn't double-fire both
rules. Golden suite in testdata/golden/ covers both rules plus two clean
cases; it's this suite, not the UI, that's the actual asset per
CLAUDE.md.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
internal/store connects via pgx and runs golang-migrate migrations
embedded in the binary (go:embed), so Deklarix stays a single binary
despite the move to Postgres. Schema covers the five MVP tables
(submission, asset, extraction, finding, evidence_package, participant).
extraction, finding and evidence_package are append-only by design: a
Postgres trigger rejects UPDATE/DELETE outright, since a corrigible
evidence archive isn't an evidence archive. Corrections to a finding are
new rows whose supersedes column points at the row they replace (set at
INSERT time on the new row, since the trigger blocks UPDATE on the old
one) — "currently valid" findings are the ones no other row supersedes.
scripts/test.sh now spins up a disposable Postgres container so the
store's integration tests (including the append-only guarantee) actually
run on every test.sh/release.sh invocation instead of silently skipping
for lack of DATABASE_URL.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
