Trust, privacy, and model governance (1.0)¶
Status: gate closed for programme purposes (0.9.7 drafts; Hub-card residual dropped 2026-08-07)
Programme: pre_release_roadmap_1_0.md §13
Related: release_severity_triage_1_0.md, dependency_audit.md, runtime/models.md, NOTICE
Mandatory gate before 1.0. Missing licence/privacy truth for shipped models is a release blocker; incomplete polish of notices may be a known limitation only where legal/privacy risk is absent.
Checklist¶
Third-party model and dataset licence inventory (matrix below — 0.9.5)
Model download origins and gated-model requirements (rows below)
Voice embedding and speaker-identity privacy wording (
VOICE_PRIVACY_USER_NOTICE, notice v2, Settings Speakers panel — 0.9.7)Confirmation that no telemetry or remote processing occurs unless explicitly configured (see § Telemetry)
Secrets and absolute-path audit (secrets_check + hygiene; live paths cleaned 0.9.5)
Dependency vulnerability checks (release CI pip-audit / clean-env / image pip-check — ongoing); licence NOTICE published (0.9.7)
AI output labelling (Local AI badges on Insights/Overview LLM surfaces — 0.9.7)
Model, prompt, and analytical-semantics identity in artifacts where needed (provenance badges + existing artifact provenance fields)
Explicit definition of what “reproducible” means for stochastic LLM output (see § Reproducibility)
Hub-card /
owner-verifyresidual dropped 2026-08-07 — ship on documented Hub/profilelicence=fields + NOTICE; no further owner Hub-card pass tracked
Telemetry / remote processing¶
TranscriptX analysis runs locally on the user’s machine (or local Docker). There is no product telemetry and no remote analysis SaaS path in 1.0 scope.
Remote network activity occurs only when the user (or install profile) explicitly enables it, for example:
Hugging Face / spaCy model downloads on first use
Optional Ollama (typically local; user-configured host)
Optional dependency / image pulls during install
Default Docker Compose binds the GUI to loopback — see SECURITY.md.
Reproducibility (LLM)¶
For stochastic LLM modules (determinism_tier T2):
Byte-identical transcript reproduction is not claimed across runs, models, or temperatures.
Artifacts should carry provenance where available: prompt version, model id, provider.
Deterministic modules (T0/T1) remain the basis for comparable longitudinal metrics.
“Re-run with same settings” is best-effort for LLM surfaces; treat outputs as drafts requiring human review (action items banner).
AI output labelling¶
User-facing LLM surfaces show a Local AI badge (and action-items keep the human-review caption). Deterministic summaries are badged Deterministic on Overview when they win precedence.
Voice / speaker-identity privacy¶
Authoritative copy: VOICE_PRIVACY_USER_NOTICE in transcriptx.core.speaker_profiles.voice.privacy (pinned by PRIVACY_NOTICE_VERSION = voice_privacy_notice.v2). Shown in Settings → Speakers before enable. Embeddings stay on-host under speaker_profiles/voice; revoke deletes evidence.
Model / dependency matrix¶
Sourced from runtime/models.md defaults and pinned HF profile licence= fields in hf_text_classification/profiles.py. Not legal advice; identifiers follow documented Hub/profile fields.
Component |
Licence |
Download / gated? |
Privacy notes |
1.0 posture |
|---|---|---|---|---|
spaCy |
spaCy model licence (see spaCy / model card) |
Docker preinstalls sm/md/lg; others download when enabled |
Local NLP; no remote call once installed |
Ship |
spaCy |
spaCy model licence |
Download on first use when enabled |
Same |
Optional |
|
Apache-2.0 (Hub card / typical) |
Hugging Face Hub download when enabled |
Local embeddings |
Ship |
|
Apache-2.0 (Hub card / typical) |
Hub download when enabled |
Local embeddings |
Optional |
NRCLex vocabulary ( |
See NRCLex / NRC emotion lexicon terms |
Via |
Lexicon association; not a neural model |
Ship |
Contextual emotion profile |
Apache-2.0 (profile |
Hub download; pinned revision; experimental |
Local classifier; experimental channel |
Ship (experimental) |
Fine-grained emotion profile |
MIT (profile |
Hub download; pinned revision; experimental |
Local classifier; experimental channel |
Ship (experimental) |
Sentiment VADER (default) |
NLTK / VADER terms |
Bundled NLTK data |
Lexicon; local |
Ship |
Sentiment transformers option |
See Hub model card |
Hub download when |
Local classifier |
Optional |
Dialogue acts (heuristic default) |
N/A (rule/heuristic) |
None |
No model download |
Ship as heuristic |
Topic modeling LDA/NMF (sklearn) |
sklearn / SciPy stack |
None beyond install |
Local |
Ship |
BERTopic stack |
See BERTopic / UMAP / HDBSCAN deps |
Embedding model as above |
Local; heavy |
Optional ( |
Ollama LLM modules (optional) |
Model-dependent (user-chosen Ollama models) |
User pulls models into Ollama |
Local when Ollama is local; label AI outputs |
Optional — document |
Voice / speaker-match embeddings (ECAPA / |
See SpeechBrain / model cards |
Optional extras; may download |
Speaker-identity sensitive — privacy notice v2 |
Optional ( |
NOTICE¶
User-facing third-party notice: repository root NOTICE (0.9.7; Hub-card residual dropped 2026-08-07). Licence identifiers follow documented Hub/profile fields.
Known limitations for users: known_limitations.md.