Privacy Policy & Data Rights
Last updated: September 20, 2026 • Effective immediately across all SiftrCode software distributions.
SiftrCode V2 is architected from the ground up as a 100% offline, local-first context optimization engine. We believe intellectual property and codebase confidentiality are inviolable. SiftrCode does not upload your files, does not maintain remote telemetry, does not train machine learning models on your syntax, and operates zero network proxies.
SiftrCode never trains, fine-tunes, or adapts machine learning models on your syntax, prompts, or proprietary code.
Processed strictly in volatile RAM. No raw source code or tokens are ever written to disk logs or remote storage.
100% of candidate recall, ranking, bundling, and budget solving runs on local CPU. 0 outbound network sockets.
In local evaluation telemetry, files not exposed to the agent are labeled unknown, never falsely assumed negative.
Raw telemetry is isolated behind a RightsFilter gate. No training dataset builder can operate directly on raw tables.
Full lineage ties every training row to upstream observations, ensuring complete GDPR/CCPA traceable deletion.
systemConfigurationHash binds sessions, plans, and observations cryptographically.
Raw rows cannot enter persistent training storage without sanctioned texport_ export IDs.
1. Source Code Execution & Processing
When you execute siftrcode context, siftrcode audit, or invoke MCP tools through your coding agent, all syntax tree analysis, multi-channel candidate recall, transparent ranking, and knapsack budget solving occur exclusively on your local device's CPU using SiftrCode's air-gapped runtime.
- Zero Network Sockets: The parsing and optimization engine opens 0 network connections during code analysis.
- Memory-Only Syntheses: Context bundles are generated in volatile RAM and emitted directly across local standard input/output (
stdio) pipes to your agent client. - No Remote Parsing Queues: No source files, directory paths, git commits, or AST representations are ever streamed to external cloud servers.
2. Telemetry & Analytics
The SiftrCode Open Source CLI (npx siftrcode) contains no usage tracking, no user telemetry, no analytics beacons, and no crash report uploads. We do not track what repositories you optimize, which files you prune, or what AI agents you pair with SiftrCode.
3. Model Context Protocol (MCP) Boundaries
The SiftrCode MCP Server communicates with coding agents (such as Claude Code, Cursor, or Antigravity) strictly over local Unix standard streams (stdio). All 10 native MCP tool calls (siftr_context, siftr_optimize, siftr_rank, siftr_expand, siftr_outcome, siftr_session, siftr_skeleton, siftr_batch_skeleton, siftr_pack, siftr_audit) execute as direct child processes of your IDE, preserving your IDE's established security boundary.
4. Learning Plane, RightsFilter Boundary & Deletion Traceability (Sections 50–52)
For self-hosted enterprise or research environments utilizing SiftrCode's optional learning plane to refine ranking heuristics, strict algorithmic firewalls and traceable lineage protect your intellectual property:
-
Authoritative Operation Rights & Storage Sanitization: Durable SQLite persistence (
.siftr/observations.sqlite) is governed authoritatively byoperationRights.*.retention.local. Under default customer rights, all raw source code strings, snippets, and prompts are automatically stripped prior to insertion. Legacy flat flags serve strictly as backward-compatible fallbacks. -
Fail-Closed Remote Processing & Credential Isolation: External provider processing is strictly fail-closed: unless explicitly enabled with granular permissions (
createJevPermittedDataRights), all remote operations are denied. In addition, deprecated clients disable environment-key inheritance, guaranteeing that legacy tools (such assiftr_pack) cannot leak credentials or bypass V2 data rights. -
Enforced Egress Security Gateway: External provider calls (rerankers, embeddings, JEV) are structurally wrapped by
EnforcedEgressGateway, which checks customer DataRights, filters secrets, and strictly guarantees that execution callbacks receive only sanitized, redacted contents. -
Strict RightsFilter Export Boundary: The local telemetry store operates behind an enforced pipeline:
Raw Observation Store → RightsFilter → Training Export. Candidate decision observations are rights-sanitized before insertion, and training set generation pipelines are strictly prohibited from querying raw observation tables directly. -
Tri-State Evidence & Null Training Labels: Downstream outcomes are captured as tri-state verification records (
verifiedSuccess: boolean | null). Unexposed or unobserved candidates are labeled asUNKNOWNwithlabel: null, mathematically preventing unobserved code from being penalized as negative examples. -
Source Provenance & Default Review Gate: Every source file and repository is stamped with provenance permissions (
ALLOWED,FORBIDDEN, orREVIEW). Per Section 51, any code with ambiguous, unspecified, or proprietary commercial licensing automatically defaults toREVIEWand is unconditionally blocked from training exports. -
Multi-Dimensional Training Evidence Lineage: Exported training records (
TrainingEvidenceRecord) retain deep lineage metadata (DerivedDataLineage) linking feature vectors back to observation record IDs, license state, feature transform versions, and retention windows without leaking source code. -
Traceable Erasure (GDPR Art. 17 / Right to be Forgotten): When a repository is purged or access rights are revoked, the
DeletionManagercascades across all derived training rows, purges all associated downstream records, and records an immutable, verifiable audit record (DeletionAuditRecord). -
Canonical AgentEnvironment Lineage & Migration 13: All runtime stores (
sessions,context_plans,candidate_decision_observations, andjev_shadow_judgments) enforce exact environment lineage derived fromsystemConfigurationHash. Mismatched agent environments throwAGENT_ENVIRONMENT_MISMATCH, preventing poisoned or uncurated state. -
Sanctioned Exporter Boundary: Raw training rows and evidence records cannot be written to persistent storage without a sanctioned export ID (
texport_/texport_ev_) issued byTrainingExporterwith verified multi-dataclass data rights. Direct raw inserts are blocked at the database boundary.
Data Classifications & Lifecycle Boundaries
To eliminate ambiguity regarding what SiftrCode reads, transforms, or stores, our architecture strictly delineates five formal data classifications:
| Classification | Examples | Processing State | Storage / Egress Gate |
|---|---|---|---|
| SYMBOL_METADATA | Function names, line ranges, symbol types, export signatures | Local Volatile RAM | Persisted strictly in local SQLite (.siftr/observations.sqlite). Never sent to remote servers. |
| AST_STRUCTURE | Syntax tree node types, parent-child edges, reference topologies | Local Volatile RAM | 0 disk writes. Discarded immediately after bundle composition. 0 network egress. |
| CODE_EXCERPT | Discrete method body slices (BODY resolution isolating target method) |
Local Pipe Only | Emitted directly across local stdio pipes to your authorized agent. Never retained in logs. |
| RAW_SOURCE | Complete raw source file contents on local disk | Local Disk Workspace | STRICTLY PROHIBITED: rawSourceRetentionAllowed: false. Never persisted in telemetry or sent off-device. |
| REDACTED_CONTENT | AWS keys, Stripe secrets, Bearer tokens, private certificates | Sanitized Instantly | Intercepted by EnforcedEgressGateway and replaced with [REDACTED_SECRET] before any downstream handler. |
Structured Zero-Egress Gateway & Shadow JEV Air-Gapped Privacy
SiftrCode v0.2.1 introduces the Structured Zero-Egress Gateway (EnforcedEgressGateway) and rigorous shadow-mode boundaries for experimental judgment evaluation:
All data egress paths enforce automated credential scrubbing using strict regular expressions. AWS Access Keys (AKIA[0-9A-Z]{16}), Stripe Secret Keys (sk_live_[0-9a-zA-Z]{24,}), and Bearer authentication tokens are scrubbed and replaced with deterministic redaction tokens before any provider callback can execute.
When Judged Evidence Validation (JEV) runs in shadow mode or under SiftrBench validation (500 bounded shadow calls across 25 audited real tasks on Express, FastAPI, and SiftrCode), it operates as a silent shadow provider. Deterministic ContextRank remains the sole, unmodified decision-maker, guaranteeing 100% normalized decision plan invariance (SHA-256 match across units, resolutions, allocations, and exposures). Shadow observations are air-gapped from production bundles.
7. Website & Team Trial Inquiries
When you visit our marketing website (siftrcode.com) or request a Team Pro trial via our activation form, we may collect:
- Work Email Address: Provided voluntarily to issue team trial credentials and setup guides.
- Team Size & Primary Agent Stack: To provision appropriate team cache resources.
- Standard Web Server Logs: Standard IP addresses, timestamps, and browser user-agent headers generated during HTTP requests, retained strictly for server security and rate-limiting.
We never sell, rent, monetize, or trade contact details with third-party advertising networks.
8. Security Disclosures & Inquiries
Because SiftrCode is open source, our AST engine and local daemon are completely transparent and verifiable. If you have questions regarding security or wish to report a vulnerability, contact our engineering team directly at:
Ready to test it on your codebase?
Run a 5-second audit or register SiftrCode directly with your agent.
npx siftrcode audit