Skip to content

Jura Trace v1.0.0

Latest
Latest

Choose a tag to compare

View all tags
@juralabs juralabs released this 18 Jun 07:32
v1.0.0
8c07e91

Jura Trace — Know What's Real.

Jura Trace v1.0.0

📥 The canonical download surface is juralabs.org/download — same installers, plus the getting-started guide, FAQ and a practice pack of sample images. This Releases page is a stable mirror for direct binary access.

First-time pilot tester? Read the Getting Started guide — install + your first Content Credentials verification in 10 minutes.

Jura Trace examines images, videos and documents and tells you what it finds — honestly, and without certainty where none exists. Twelve forensic detectors plus full C2PA Content Credentials verification, running entirely on your device. No cloud, no accounts, no telemetry.

Download

If you have… Download Signed by
Mac (Apple Silicon — M1/M2/M3/M4) JuraTrace-1.0.0-macOS-AppleSilicon.dmg Apple Developer ID (notarised)
Windows (64-bit) — recommended JuraTrace-1.0.0-Windows-x64.msi Azure Trusted Signing
Windows (64-bit) — NSIS installer JuraTrace-1.0.0-Windows-x64-setup.exe Azure Trusted Signing
Linux (x86_64) — AppImage JuraTrace-1.0.0-Linux-x86_64.AppImage Self-signed (AGPL source build)
Linux (x86_64) — DEB package JuraTrace-1.0.0-Linux-x86_64.deb Self-signed (AGPL source build) — install with sudo apt install ./<file>

Verify your download: grab SHA256SUMS.txt and run shasum -a 256 -c SHA256SUMS.txt (macOS / Linux) or certutil -hashfile <filename> SHA256 (Windows).

Full install + first-test walkthrough: GETTING_STARTED.md.

Technical changelog

8 June 2026 — rc.30 cut: pre-launch documentation hardening, copyright sweep, build-script verification gate, Tauri updater URL fix

Six commits across three calendar days (6–8 June) preparing rc.30 for the v1.0 launch on Monday 22 June 2026. Closes the post-rc.29 punch list: build-script architecture bug that caused rc.29's first notarisation to be rejected, Tauri updater URL typo that broke the auto-updater fallback, copyright-sensitive training-data attribution in user-facing documentation, the v10 retrain decision, and 31 of 46 findings from the 8 June documentation deep review.

Build script Phase 0.5 + Phase 6.5 — eliminates the rc.29 notarisation failure (7 June, commit 8aac1d8)

Diagnosed root cause of rc.29's first notarisation rejection (Apple notarytool reported 600+ nested .so/.dylib files plus the PyInstaller bootloader jura-sidecar as "binary is not signed with a valid Developer ID certificate"). The build script's Phase 5a per-file-signed every Mach-O inside .app/Contents/Resources/sidecar-bundle/ correctly. But Phase 6 (cargo tauri bundle --bundles dmg,updater) re-copied from src-tauri/sidecar-bundle/ into the .app, overwriting Phase 5a's signed files with the unsigned source. The .app's top-level re-seal then baked hashes of unsigned nested files into the new signature. The DMG passed codesign --verify but failed Apple notarisation. rc.29 was rescued by a 30-minute manual recovery: extract .app from rejected DMG, per-file sign 611 binaries, rebuild DMG with hdiutil, re-notarise.

Two new phases in scripts/build-local-mac.sh:

  • Phase 0.5 — sign the source at src-tauri/sidecar-bundle/ BEFORE any cargo tauri bundle invocation. Both Phase 1 and Phase 6 then copy already-signed files, preserving signatures in the final DMG. Content-based Mach-O detection (file -b | grep Mach-O), not extension-based, so the PyInstaller bootloader is caught naturally without a hard-coded one-off sign line. Deepest-first ordering for correct sealing semantics. Surfaces codesign failures visibly (no stderr redirect, unlike the CI workflow bug we are fixing in parallel).
  • Phase 6.5 — post-bundle verification gate. Walks every Mach-O in the final .app's Resources/sidecar-bundle/ and asserts every signature bears TeamIdentifier=Y82C4P9L7F. Build fails fast at this step if any nested binary is unsigned or mis-signed, instead of silently producing a DMG that wastes 25 minutes at notarisation.

Phase 5a kept as belt-and-braces (now functionally redundant since the source is signed pre-bundle, but cheap defensive insurance against future Tauri bundler behaviour changes). rc.30 and v1.0 should notarise on first attempt without the 30-minute manual recovery required for rc.29.

Tauri updater fallback URL + release-notes licence typo (6 June, commits db84af3, 9d2ad8d)

Two real shipping bugs caught during the post-rc.29 audit:

  • src-tauri/tauri.conf.json:40 — the updater endpoint fallback URL was https://github.com/juralabs/jura-trace/releases/latest/download/latest.json (lowercase org). The real GitHub org is Jura-Labs/jura-trace (capital J, hyphen). The lowercase URL 404s. Auto-updater fallback was non-functional in every build through rc.29. Fixed to https://github.com/Jura-Labs/jura-trace/releases/latest/download/latest.json.
  • .github/workflows/release.yml:123 — the release-notes template footer still quoted PolyForm Noncommercial 1.0.0 as the licence. The project switched to AGPL-3.0-or-later on 2026-05-06. Every auto-generated release-notes default body (rc.27, rc.28, rc.29) carried the wrong licence. Replaced with AGPL-3.0-or-later plus a link to COMMERCIAL.md for the dual-licence path. Same typo fixed in the truncation-footer CHANGELOG link (line 132) and two cosmetic comments (lines 23, 141).

Copyright-sensitive training-data attribution removed from public documentation (7-8 June, commits 195064a, df74850, b68b1e3)

Public-facing documentation named specific external corpora and AI-generator brands as authentic and synthetic training sources. Public attribution of these references carries copyright-exposure risk independent of how the underlying training data is actually licensed or sourced.

This sweep removes the public attribution. The training data on disk is NOT touched. Reproducibility anchors remain at the model-card metadata JSON shipped alongside each release.

Three classes of treatment:

  • In-app help model card rewritten (compiled into the desktop bundle): ui/src/routes/help/model-cards/+page.svelte — both training-data sections rewritten across both model cards on the page. Authentic sources reduced to "real camera DCIM photos, Wikimedia Commons photographs (curated, non-art)". AI-generated sources reduced to "diverse imagery across the diffusion and GAN landscape spanning commercial and open-weights model families; specific generator names withheld from public documentation; per-generator recall is reported in the model-card metadata JSON shipped alongside each release".
  • Four high-risk docs moved out to internal: docs/decisions/option-c-corpus-strategy.md, docs/fairness/corpus-demographic-profile.md, docs/TEST_CORPUS_BRIEFING.md, docs/testing/real-world-test-plan.md moved to ../jura-labs-docs/jura-trace-internal/ per the existing project_repo_doc_hygiene discipline. .gitignore extended with explicit entries for the four moved paths so they do not creep back into the working tree of any clone. Empty docs/testing/ directory removed.
  • Ten medium-risk docs lightly redacted with named generator labels replaced by categorical descriptors. The model-card metadata JSON (with the actual per-generator recall numbers) stays as the reproducibility-anchor reference. Affected: docs/calibration/univfd-v10-multi-format-augmentation-plan.md, docs/calibration/univfd-v10-screenshot-retrain-plan.md, docs/calibration/univfd-v10onnx-divergence-fix.md (the public model card cited from docs/methodology.md on the release repo; per-generator-recall table retained with generic "Generator family A/B/C/D" labels), docs/calibration/univfd-v9-onnx-validation.md, docs/calibration/univfd-v9-platform-augmentation.md, docs/calibration/s28-jpeg-ghost-weight.md, docs/decisions/splice-benchmark-longterm.md, docs/av-corpus-methodology.md, docs/design/exif-injection-detection.md, docs/development-workflow.md.

Net 2,179 lines deleted / 54 added across 14 files in the source repo. Out of scope for this sweep (separate follow-up): the agent scripts at scripts/agents/crawl_authentic_images.py still contain code-level references to specific corpora (function names, HuggingFace dataset URLs). The script's behaviour is functional code, not documentation, and redacting it would break the corpus-rebuild path; to be addressed by relocating those scripts to internal if needed. Also out of scope: docs/backlog.md progress notes, which name specific generators in historical context — sweep planned for the next cleanup pass.

Pre-launch documentation deep review — 31 of 46 findings closed (8 June, commit b68b1e3)

Documentation deep-review agent identified 46 findings across four categories: inconsistencies between documents (12), backlog promises and future-version claims (14), internal tracker references in public copy (13), and detector language too specialist for non-technical readers (7). Overall RAG: Amber. Audit report saved to ../jura-labs-docs/jura-trace-strategy/documentation-deep-review-2026-06-08.md.

Top 5 P0 items fixed across release repo, in-app help, and the wiki:

  • C-01 GETTING_STARTED.md:59 — "Intel support returns at v1.0 via a universal binary" (false at v1.0) replaced with "this build requires Apple Silicon (M1 or later); a universal binary is planned for a future release". v1.0 ships Apple Silicon only.
  • T-01 / T-02 / T-03 ui/src/routes/help/methodology/+page.svelte:789 — JPEG Ghost calibration <dd> block rewritten without "Sprint 28 calibration sweep (S28-FU9)", "CASIA v2" (named external corpus, also flagged by the copyright sweep), or "backlog item #10". Honest disclosure of the calibration limitation retained (Berkeley Protocol §6 reproducibility commitment).
  • C-06 ui/src/routes/help/compliance/+page.svelte:480 — security status callout "0 open items as of v0.9.0-rc.1" (factually wrong, four months stale, understated the audit scope) replaced with "all findings from the most recent audit (May 2026) have been resolved; 0 open items".
  • C-04 wiki Home.md:24, 42 — three instances of "Juralabs CIC" (one word) corrected to "Jura Labs CIC" (canonical abbreviated form with space). Full legal name "Juralabs Community Interest Company" at line 5 left unchanged per the canonical form.
  • B-12 / B-13 GETTING_STARTED.md:315, 317 — "Thank you for piloting Jura Trace" pilot framing in the public release footer replaced with "Thank you for using Jura Trace"; stale "27 April 2026" date updated to "June 2026".

P1 fixes batched in the same sweep across in-app help (methodology + compliance + forensic-detectors pages), the GitHub Wiki (Home + Methodology + Glossary), and the public release repo's methodology overview:

  • C-02 detector-count consistency: "12 automatic forensic detectors" replaced with "ten automatic (thirteen in total, with three available on demand)" at all occurrences across GETTING_STARTED.md and the in-app methodology page.
  • C-03 database-path consistency: jura.db (incorrect) replaced with jura_archive.db (canonical, matches src-tauri/src/db.rs) at four occurrences.
  • C-05, B-11 Linux platform claim: removed from the v1.0 platform list (wiki Home + Format-Support). Amended to "macOS (Apple Silicon), Windows (x64); Linux installer planned for a future release".
  • C-09 audit-report contact email: consultancy@juralabs.org replaced with security@juralabs.org for security-audit requests and licensing@juralabs.org for commercial-licence enquiries (canonical per SECURITY.md and COMMERCIAL.md).
  • T-04 NPR Known Limitations block rewritten without internal "Sprint 28" and "Content-authenticity-expert" references. Tan et al. (AAAI 2024) paper citation kept; framing improved.
  • T-05 "Demoted Sprint 28: on-demand investigation tool only..." replaced with "On-demand: investigation tool only..." in the signal weighting table (NPR, shadow consistency, splice boundary rows).
  • T-06 / T-07 / T-08 Sprint 19 + Sprint 14 references in the compliance audit section replaced with "25 March 2026" date references or "the previous audit cycle".
  • T-09 / T-10 removed (tracked as JTV-156) and (JTV-188) ticket references from rendered forensic-detectors copy.
  • B-02 FFmpeg install section in GETTING_STARTED.md gained an explicit note that video/audio analysis is planned for a future release: "FFmpeg can be installed now; the features will activate automatically when they become available".
  • B-03 docs/methodology.md:21 (release repo) — "reactivated in v1.0.2 (October 2026)" softened to "planned for re-enablement in a later release". No version + date pin.
  • B-04 docs/methodology.md:79 (release repo) — "audio and video deepfake analysis are deferred to v1.0.1 (target Monday 4 August 2026)" softened to "planned for a subsequent release, timed to the EU AI Act Article 50 transparency obligations (binding 2 August 2026)". Drops the v1.0.1 version pin and the specific 4 Aug target Monday; keeps the legitimate Article 50 binding-date anchor.
  • B-05 / B-06 / B-07 "Pro tier feature" softened to "feature-flagged off in v1.0 and planned for a later release" across wiki Methodology + Glossary entries for Conformant Mode + Signing Mode. The Pro tier has not been announced at v1.0 launch.
  • C-11 / C-12 AGPL-3.0 SPDX identifier extended to AGPL-3.0-or-later at five occurrences across the wiki and in-app help.

Category 4 plain-English rewrites in the wiki Methodology page (the flagship transparency page that journalists, fact-checkers, museum curators, and solicitors will read on launch day):

  • 5b GBM AI Generation Detection: "extracts an 84-feature vector across six classes: noise statistics (LSB randomness, LSB entropy, LF/HF ratio, anisotropy); spectral decay patterns; Local Binary Pattern (LBP) texture descriptors; Grey-Level Co-occurrence Matrix (GLCM) contrast measures; demosaic inter-channel coherence; and PRNU sensor pattern consistency" (six unexplained acronyms in one sentence — discipline-specific signal-processing jargon) preceded by a plain-English opener: "the classifier examines statistical properties of the image that are invisible to the naked eye: the distribution of noise grain, the way fine texture repeats across the image, the consistency of colour-channel data, and the spatial pattern of sensor noise that real camera hardware imprints on every photograph it takes". Technical detail preserved in trailing parenthetical for specialist readers.
  • 5c Copy-Move Detection: "extracts SIFT (Scale-Invariant Feature Transform) descriptors from image patches and performs nearest-neighbour self-matching with Lowe's ratio test... RANSAC geometric verification and DBSCAN clustering" rewritten in plain English: "breaks the image into small overlapping patches and generates a compact fingerprint for each patch; those fingerprints are compared across the whole image; if two patches in different parts of the image are nearly identical, they are likely copies of each other; the detector verifies that the matches form a geometrically coherent group before flagging them". SIFT/Lowe/RANSAC/DBSCAN technical names preserved in trailing parenthetical.
  • 5d Colour Temperature: "Converts the image to the CIELAB perceptual colour space" — CIELAB defined inline: "(a standard that aligns colour distance with human visual perception)".

15 findings deferred to v1.0.1 per the audit's own prioritisation: HTML-comment sprint references not rendered to users (T-11/12/13), wiki Glossary + Format-Support version pins (B-08/09), compliance page Video Deepfake numbering inconsistency (C-10), in-app forensic-detectors PRNU residual energy roadmap reference (B-10), and three of the seven Cat 4 readability notes (5a fixed via T-01, 5e/5f noted as no-change-needed templates, 5g handled via T-04).

v10 retrain decision — skip, v10onnx + GBM v4 IS the v1.0 launch model (7 June)

Third (and final) re-check of the corpus delta against the 18 May reference date. Zero new authentic samples since 11 May v10onnx training (the one "new" file in corpus/training/authentic/google_photos/ is a _c2pa.jpg test artefact — a Jura-Trace-signed copy of an existing image used for verify-pipeline testing, not corpus data). Zero new AI samples. Zero new platform-forwarded augmentation data. Decision locked: skip the refresh-only retrain. v10onnx (0534a9e80e352a5b…) + GBM v4 (512def7ec62cbeb0…) is the v1.0 launch model.

Rationale documented in memory project_v10_retrain_committed.md: retraining identical algorithm on identical data produces a numerically identical model with no quality benefit and exposes the SHA-pinned model identity (already cited in docs/methodology.md, the rc.29 release notes, the press release v2.2 Technical reference block, and the sidecar runtime identifier univfd-probe-v10onnx) to needless re-publishing risk. The actual retrain trigger is JTV-127 (Track 3 Global Majority handset corpus, ~1,000 outstanding photos); when those land post-launch, schedule UnivFD v11 against the GM additions specifically — a v1.0.1 / v1.1 deliverable, not a pre-launch fire-drill.

Companion change in the public release repo: docs/methodology.md §2.3 (commit c9a1614 on Jura-Labs/jura-trace) updated from stale v9 numbers to current v10onnx (training corpus 39,016 → 50,710 samples; AUC-ROC 0.9933 → 0.9929; FP rate 4.12% → 3.87%; AI recall 95.70% → 95.77%; SHA ed691b45… → 0534a9e8…; model card path univfd-v9-platform-augmentation.md → univfd-v10onnx-divergence-fix.md; new paragraph explaining the "onnx" suffix as a PyTorch↔ONNX preprocess divergence fix retrained against the production sidecar's PIL + ONNX runtime path). Closes funder-audit finding D4 (Medium).

Codeberg source mirror catch-up sync (8 June, codeberg commit 0bf2f23)

The post-AGPL clean history on codeberg.org/jura-labs/jura-trace diverged from Jura-Labs/jura-archive over the rc.30 prep cycle. Cherry-picks failed on different-ancestry conflicts because the clean-history split means several jura-archive commits reference files (the calibration / decisions / fairness docs) that do not exist on codeberg's tree at all.

Resolved via end-state sync: six files brought to current state from jura-archive HEAD and applied as a single squashed commit on codeberg. Equivalent to cherry-picking commits 8aac1d8, 195064a, df74850, b68b1e3 from jura-archive but applied without the cherry-pick conflicts.

Synced files: scripts/build-local-mac.sh (Phase 0.5 + Phase 6.5), ui/src/routes/help/model-cards/+page.svelte (training-data redactions), ui/src/routes/help/methodology/+page.svelte (JPEG Ghost block + NPR + signal table + AGPL identifier), ui/src/routes/help/compliance/+page.svelte (audit status + Sprint references + contact emails), ui/src/routes/help/forensic-detectors/+page.svelte (JTV ticket removals), and .gitignore (precautionary entries for the four high-risk files that exist on jura-archive but not on codeberg).

All four remotes now in sync for v1.0 launch on Monday 22 June 2026:

Remote Head Purpose
Jura-Labs/jura-archive (private source) b68b1e3 Full dev history + canonical source
codeberg.org/jura-labs/jura-trace (public source mirror) 0bf2f23 Clean post-AGPL history; goes public on launch day 22 June
Jura-Labs/jura-trace (public release) 8c07e91 Installers + READMEs + methodology overview
github.com/Jura-Labs/jura-trace/wiki 572f39e Mirror of Codeberg wiki content (live now)

Test posture post-sweep

  • 543 Rust lib tests pass unchanged
  • cargo clippy --all-targets -- -D warnings: clean
  • cargo fmt --check: clean
  • svelte-check: 0 errors across the help-page surface
  • Sidecar test_health.py: 3/3 pass
  • Pre-commit hook ran cleanly on each of the documentation commits (195064a, df74850, b68b1e3)
  • npm ci --dry-run passes against the restored lockfile (a working-tree drift of the same class that killed rc.28 was caught and reverted in the pre-flight check on the evening of 8 June)

Open items carried into rc.30 + v1.0 launch

  • rc.30 tag-cut scheduled for Tuesday 9 June 2026 from current HEAD b68b1e3. Pre-flight diagnostic confirms all blockers resolved: lockfile valid, Apple Developer ID cert valid through April 2031, notarytool keychain profile jura-trace-notary authenticates, Tauri updater key at ~/.tauri/jura-trace-v10.key (mode 600), Phase 0.5/6.5 build-script fix in place, all four models present on disk, Python 3.13.5 + PyInstaller 6.19.0 ready, external cargo target (/Volumes/MAC SSD) has 730 GB free, no hung processes.
  • press@juralabs.org alias (JTV-166) — not yet active. Press release v2.2 uses hello@juralabs.org as the catch-all per the canonical contact policy.
  • juralabs.org/download page (JTV-242) — scheduled deploy Friday 19 June 2026. HTML template at ../jura-labs-docs/jura-trace-strategy/juralabs-org-download-page.html.
  • Codeberg repo public flip — scheduled launch day 22 June 2026 (currently private).
  • rc.29 release on public repo — to be marked as draft when rc.30 publishes. rc.29 carries the stale model card with named external corpora; users downloading from /releases/latest should land on rc.30.
  • 15 of 46 documentation deep-review findings deferred to v1.0.1 per the audit's own prioritisation; sweep planned for the Article 50 compliance launch (target early August 2026).
  • Detector plain-English rewrites for the in-app methodology help page — the equivalent of wiki Cat 4 5b/5c/5d (GBM, Copy-Move, Colour Temperature) applied only to the wiki; the in-app methodology page received the equivalent fix only for the JPEG Ghost block (5a). The non-blocking 5b/5c/5d in-app fixes can be batched into the v1.0.1 cycle if not done before launch.
  • scripts/agents/crawl_authentic_images.py still contains code-level references to specific corpora that the copyright sweep did not touch. To be addressed by relocating those scripts to internal if the external attribution risk warrants it.
  • docs/backlog.md historical progress notes naming specific generators in context — sweep planned for the next cleanup pass.
Assets 12
Loading