Deklarix itself no longer depends on the Anthropic API — that was a
separate API key/billing relationship from Claude Code (used to develop
Deklarix), which the user did not intend to take on for the product
itself.
Consideration (Gegenleistung) is no longer guessed from text — it's a
required form field now, since only the submitter actually knows
whether a business relationship existed. A keyword-only system can't
tell a covertly-paid post from a genuinely organic one; they read
identically. What internal/extract *can* still determine reliably and
deterministically from the caption: whether a disclosure keyword is
present (werbung, anzeige, bezahlte partnerschaft, paid partnership,
#ad, #werbung, #anzeige, #sponsored, #sponsoredby, #sponsoredpost —
case-insensitive), its exact original-case wording, and whether it sits
before the platform's "mehr anzeigen" truncation point (~125 chars
Instagram, ~150 TikTok — rough estimates, platforms change these without
notice, verify before real customer use).
internal/extract's Anthropic HTTP client and tool-use schema are gone
(client.go/api.go deleted), replaced by engine.go — a stateless Engine
with no network calls. extract.Result/ParsePayload keep the exact same
JSON shape as before (gegenleistung/kennzeichnung_vorhanden/
kennzeichnung_wortlaut/kennzeichnung_vor_kuerzung), so internal/store and
internal/dossier needed no changes at all — only extract itself, the web
form/handler (new consideration field), and main.go (no more
ANTHROPIC_API_KEY requirement) changed.
Trade-off the user was told and accepted: without an LLM, the system can
no longer independently catch undisclosed paid content that carries no
recognizable keyword at all — that now rests on the submitter's honesty.
Creative or implicit disclosure phrasing outside the keyword list also
won't be recognized.
Verified against a real running instance with zero API keys configured:
register -> check (real rule engine, correctly triggered WK-004 for a
disclosure placed 130 characters in, past the Instagram threshold) ->
archive -> PDF dossier download, all against real Postgres.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
This is the piece that turns a Pre-Publish-Prüfung into an actual
archived, provable record instead of a one-off form response.
extract.Client.Extract now returns Result{Facts, RawJSON} instead of
just Facts — RawJSON is the model's exact, unmodified JSON, which is
what belongs in extraction.payload (the audit trail), not a re-encoded
view through our own Facts struct. extract.ParsePayload reconstructs
Facts from a stored payload later, reusing the same parsing/validation
path Extract uses (including the enum guard), so a previously-saved
extraction can be read back exactly as it would have been the first
time.
internal/dossier.BuildContent no longer requires AssetHash: most checks
right now are caption-only (no image/video upload wired yet), and
inventing a placeholder hash for a nonexistent asset would itself be an
integrity problem in an evidence tool. Content shows "kein Asset
hinterlegt" instead.
internal/web gains a narrow Store interface (mirroring the Extractor
pattern — only the methods these handlers use, not the full
*store.Store) so its test suite stays network/DB-free via an in-memory
fake:
- POST /pruefen persists submission + extraction + findings and marks
the submission "checked". A needsClarification result persists the
extraction (there's something worth keeping) but no findings and no
status change, and the template omits the archive option entirely.
- POST /veroeffentlichen re-derives Facts from the stored payload, hashes
the canonical submission+facts+findings metadata, gets an RFC-3161
timestamp, generates the PDF dossier to disk, and persists the
evidence_package row before marking the submission "published".
- GET /dossier/{id} serves the generated PDF.
Tested end-to-end offline: a fake Timestamper builds a real, structurally
valid self-signed RFC-3161 response so the full check→archive→download
flow runs against an in-memory store, verifying the downloaded bytes are
an actual PDF and the dossier file lands on disk — without hitting a
real database, TSA, or the Claude API.
cmd/deklarix/main.go now wires store.Store, evidence.NewHTTPTimestamper
(TSA_URL, default FreeTSA), and DOSSIER_DIR (default "dossiers") into
web.NewServer alongside the extractor and rule set.
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>
