English | 日本語

Sieve Tray v0.9.1

A general-purpose observation tool — a desktop front-end for the Sieve series

Combines Sieve Lens's invisible-content observation, Sieve Scope's code similarity detection, and Sieve Referee's text paraphrase detection — plus, new in v0.8.0, Sieve Referee's quantifier audit lane — into a single desktop app. Point it at a folder of submissions and it detects and reports, in one pass, documents containing invisible content or hidden prompts, code pairs with suspicious structural similarity, text pairs that look paraphrased or reused, and sentences that make ungrounded universal claims (hallucination candidates).

---

💡 Concept

Sieve Lens answers "does this document contain something a human can't see but a parser can?". Sieve Scope answers "do these two source files share suspiciously similar logic, beyond coincidence?". Sieve Referee answers "does this text say the same thing as that one, just reworded?". All three are zero-dependency, deterministic observation engines in the Sieve series — but checking a folder of submissions by hand meant running separate tools and cross-referencing the output yourself.

Sieve Tray removes that step. Point it at a directory containing one subfolder per submission, and it automatically routes each file to the right engine: documents go to Sieve Lens, .py files go to Sieve Scope's pairwise comparison, and plain-text files (.txt/.md) additionally go to Sieve Referee's pairwise paraphrase comparison and Sieve Referee's quantifier audit lane (new in v0.8.0, powered by Sieve Referee >=1.3.0). Results land in a single window, with local history so nothing has to be re-run, and a detail view showing exactly which observation axis fired and why — never a bare verdict, only what the underlying engines actually observed.

Like the rest of the Sieve series, this is an observation tool, not a judgment tool. A flagged item means "this deserves a human look", not "this is confirmed misconduct".

---

🧩 The Sieve Series

Sieve is a family of deterministic, zero-dependency engines that share the same design principles (deterministic output, no telemetry, fully local) — each engine observes or masks one specific risk, and they compose:

Package	Role	Connects to
sieve-core-engine	Base logic / plugin host for the ecosystem	—
sieve-lens	Observes invisible content & hidden prompts in documents	→ redact via --from-lens
sieve-scope	Detects structurally similar code pairs	—
sieve-referee	Detects paraphrased / reused text pairs	—
sieve-redact	Masks sensitive regions (text & PNG) with receipts	← lens, → tray
sieve-tray	Desktop front-end that routes files to the right engine	← all engines

Observation → masking in one pass: Lens finds what a human can't see, Redact removes it with a receipt that proves nothing else changed.

🔥 Key Features

• One folder in, everything checked — routes .py files to Sieve Scope, plain-text files (.txt, .md) additionally to Sieve Referee (both pairwise paraphrase comparison and the quantifier audit lane), and all documents (including .html, .docx, .pdf, images) to Sieve Lens, automatically, by extension.
• Background-threaded GUI — scanning never freezes the window, even on large submission sets.
• English and Japanese UI — switch languages from the Settings dialog (English by default). Note: Sieve Referee's own explanation text for a flagged text pair is generated in Japanese regardless of UI language, since the engine's calibration is Japanese-specific; the detail dialog shows it labeled as such.
• Sortable, filterable results — click any column to sort; filter documents by filename and/or "flagged only"; filter code/text pairs by filename.
• Detail dialog — double-click any row to see the per-axis observation states and the underlying evidence/scores/reason the engine recorded, not just the mask string.
• Quantifier Audit lane (new in v0.8.0) — every scanned .txt/.md file is also passed through Sieve Referee's quantifier audit lane, which flags sentences that make universal claims without backing (theorem / enumerated / sample) — the kind of "all X are Y" claim that, with no proof or finite enumeration behind it, is a hallucination candidate. Flagged sentences surface in their own Quantifier Audit tab, one row per sentence, with the signal (RED = unbacked universal, YELLOW = sample-only or unbacked measured), the matched patterns, and a separate Certainty Markers section in the detail dialog showing any epistemic certainty markers (e.g. 「明白に」「例外なく」) co-occurring in the sentence. The lane ships with a bundled default modifier-vocab file; users can override the path from Settings to swap in a custom vocabulary.
• Local history — every scan is saved to a local SQLite database (~/.sieve_tray/history.db); revisit a past run without re-scanning.
• Configurable — default export folder, history retention, UI language, and the audit lane's modifier-vocab file (empty = bundled default) are all adjustable from a Settings dialog.
• File or folder selection (new in v0.9.0) — point Sieve Tray at a whole folder (the original mode) or pick individual files; the audit lane runs on a single document, so file-selection mode is a natural fit for quick one-off checks.
• Self-contained report export (HTML or Markdown) — every result can be exported as a standalone HTML report, or as a Markdown report for plain-text / version-controlled workflows.
• History clear (new in v0.9.0) — the History and Sanitize History tabs each have a "Clear History" button for when you want to wipe the local SQLite store.
• Improved signal colors (new in v0.9.0) — RED and YELLOW rows have stronger pastel backgrounds and the signal column is rendered in a signal-colored foreground, so flagged rows are easy to spot at a glance even on calibrated monitors.
• Sanitize text files — run the bundled Sieve Redact engine over the selected folder's .txt/.md files with a rules file of your choice; every output gets a machine-readable receipt, and the summary always reports the masked item count and the byte-integrity verification.
• Sanitize an image — drag rectangles on a PNG (8-bit RGB/RGBA) in the region editor, preview the real engine output in place, and save a reusable region config; every output gets a machine-readable receipt (pixel integrity is verified; byte integrity is not applicable to PNG re-encoding).
• Sanitize a folder of images — apply one region config to every PNG (8-bit RGB/RGBA) in the folder. The config kind routes the run: region-only configs target PNGs, rule/builtin configs target .txt/.md, and files that don't match are reported as excluded (never as failures). A config mixing both kinds is rejected up front with nothing processed. Every output gets a machine-readable receipt.
• Sanitize history tab — every sanitize run (folder and single image) is listed in its own tab with timestamp, folder, config, output location, files and ok count, masked items, and integrity failures, and refreshes the moment a run completes.
• Runs entirely on your machine — scanning happens 100% locally. Submission contents never leave your device: no cloud upload, no API calls involved.
• Zero-dependency core — the scanning engine (sieve_tray.py) itself uses only the Python standard library plus Sieve Lens / Sieve Scope / Sieve Referee. The GUI is an optional layer (PySide6) on top, kept separate so the core stays consistent with the rest of the Sieve series' "zero dependency, offline, deterministic" design.
• Cross-platform desktop builds — CI produces standalone Windows and macOS (Apple Silicon + Intel) apps via PyInstaller; no Python installation required to run them.

---

📦 Installation

Option 1: Download a pre-built app (recommended)

No Python required. Download the build for your platform from the Releases page:

Platform	Asset
Windows	SieveTray-windows.zip
macOS (Apple Silicon / M-series)	SieveTray-macos-arm64.zip
macOS (Intel)	SieveTray-macos-intel.zip

Unzip and run SieveTray.exe / SieveTray.app. On first launch, macOS will warn that the developer can't be verified (the app isn't code-signed yet) — go to System Settings → Privacy & Security and click Open Anyway. Windows SmartScreen may show a similar warning; choose More info → Run anyway.

Option 2: Install from source

git clone https://github.com/neguseatama/sieve-tray.git
cd sieve-tray
python3 -m venv .venv && source .venv/bin/activate
pip install -e ".[gui]"
python3 sieve_tray_gui.py

Requires Python 3.10–3.13 (PySide6 does not yet support 3.14).

Core only, no GUI (for CLI usage or scripting):

pip install -e .

---

💻 Quick Start

GUI

1. Launch the app.
2. Click Select Folder... and pick a directory containing one subfolder per submission.
3. Click Run Scan.
4. Browse results in the Documents, Code Pairs, Text Similarity, and Quantifier Audit tabs — sort by column, filter by filename, and double-click any row for the full per-axis breakdown (audit rows show a separate Certainty Markers section in the detail dialog).
5. Export a shareable report with Export HTML Report. Past runs are always available under the History tab.

(The interface language defaults to English; switch to Japanese from the Settings dialog. Takes effect after restart.)

CLI

python3 sieve_tray.py path/to/submissions -o report.html

Expects a directory of subdirectories, one per submission:

submissions/
    student_a/
        essay.docx
        solution.py
    student_b/
        essay.docx
        solution.py

As a library

from pathlib import Path
from sieve_tray import run_scan, render

result = run_scan(Path("submissions"), progress=print)
print(result.doc_results)   # path / mask / flagged / h_states / evidence
print(result.code_pairs)    # a / b / mask / h_states / scores
print(result.text_pairs)    # a / b / signal / mask / pattern_name / reason / h_states / scores
print(result.audit_results) # v0.8.0: file / sentence_sha256 / sentence_text /
                           #         patterns / backing / scope / signal /
                           #         pattern_name / reason / certainty_markers /
                           #         modifier_vocab_sha256 / modifier_vocab_count

Path("report.html").write_text(render(result))

---

🖼️ Screenshots

English UI	Japanese UI

Code Pairs	Text Similarity

---

🔬 Testing

pip install -e . pytest
pytest -v

89 tests cover the core scan/render pipeline (including regression tests for a CSS rendering bug and for Sieve Referee's paraphrase detection), the history/settings storage layer (save / load / delete / retention pruning, including migration from pre-Referee databases), and the i18n string tables (English/Japanese key parity), plus the vendored Sieve Redact integration (byte-pin sync and exit-code mapping), and the image sanitize path (region helpers, receipt handling, and editor widgets under an offscreen Qt platform), and the folder-uniform image sanitize flow (config-kind routing, excluded-file reporting, and tree-level config errors).

CI runs this suite on every push (Python 3.10 and 3.12, on Ubuntu), and builds Windows/macOS desktop apps on every version tag.

---

⚠️ Known Limitations

1. Code comparison is Python-only. .py is the only extension routed to Sieve Scope; other languages are treated as plain documents (or ignored) for now.
2. Text-similarity comparison is plain-text only. .txt/.md are routed to Sieve Referee; rich formats (.docx, .pdf) are not currently extracted to plain text for this comparison, even though Sieve Lens still observes them individually.
3. Sieve Referee's reason text is Japanese-only, regardless of UI language, since the engine's similarity thresholds are calibrated for Japanese text. The detail dialog surfaces it labeled as such.
4. Desktop builds are not code-signed. Windows SmartScreen and macOS Gatekeeper will warn on first launch; this is expected for an unsigned build.
5. This is an observation tool, not a judgment tool (inherited from Sieve Lens, Sieve Scope, and Sieve Referee). A flag means "a human should look at this", not "this is confirmed misconduct".
6. Sanitize covers plain text and PNG images only — the folder Sanitize action routes by config kind (region-only configs target PNGs, rule/builtin configs target .txt/.md) in a background thread; the image editor handles single PNGs (8-bit RGB/RGBA, non-interlaced). Both write a machine-readable receipt next to every output. PDF and JPEG masking are not supported by the bundled engine.
7. Quantifier audit is Japanese-calibrated (new in v0.8.0). The audit lane's default universal-scope vocabulary and modifier vocabulary are tuned for Japanese text; sentences in other languages may not trigger the lane even when they make equivalent universal claims. The lane surfaces the vocab sha256 and count in the detail dialog so the calibration is visible, and the modifier-vocab file can be overridden from Settings.
8. Quantifier audit surfaces candidates, not verdicts. A RED signal means "this sentence makes a universal claim with no backing recorded in the tagfile" — it is a hallucination candidate, not a confirmed hallucination. The tagfile is the source of truth for backing labels (theorem / enumerated / sample / unbacked); audit_text() does not populate the tagfile automatically. A human must confirm whether a flagged sentence is actually unbacked or merely un-tagged.

---

📚 Related Projects

• Sieve Lens — invisible-content observation engine (documents)
• Sieve Scope — code similarity detection engine
• Sieve Referee — text paraphrase / plagiarism detection engine

---

🎨 Application Icon

Design: AURA_HYSTERICA

---

📄 License

MIT License. See LICENSE for details.

---

👤 Author

• Kai IWASAKI