Product & research / October 10, 2026
Introducing TracerYX Cloak
Cloak is TracerYX’s first-party PII redaction and tokenization product, built to advance the state of the art in local privacy protection.
Complete-span coverage
Character F1
Protect data before export
Latest primary cohort: 797 retained English documents from Nemotron and Gretel; 790 annotation-valid. TracerYX at threshold 0.20. Removal outside annotations: 3.64%.
AI applications need useful context. That context often includes names, contact details, addresses and identifiers that should not follow a trace into every downstream system.
TracerYX brings detection and transformation into the customer’s environment. It finds supported private spans, replaces them with opaque aliases or redaction markers, and protects supported trace payloads before export. Teams can keep the relationships they need for debugging while reducing the sensitive values exposed in telemetry.
Our work focuses on complete identifiers and useful text together. On the latest primary evaluation, TracerYX covered 95.79% of complete reference spans, compared with 67.60% for the tested Presidio configuration. The difference is 28.19 percentage points on this cohort. Results across other document families, and the cost of unnecessary removal, are reported below.
01 / Comparative evidence
Cover the complete identifier.
Finding part of an address is not enough. Our complete-span metric requires every annotated character to be protected. On the latest primary cohort, TracerYX had the highest observed pooled complete-span coverage among the six tested configurations.
797 retained English documents (790 valid), Nemotron and Gretel. Settings were frozen before test acquisition. Full spans and character F1: higher is better; outside removal: lower is better.
Scroll the table horizontally to compare all metrics.
| System | Full spans | Character F1 | Outside removal | Errors |
|---|---|---|---|---|
| TracerYX | 95.79% | 90.36% | 3.64% | 0 |
| Presidio .01 | 67.60% | 77.15% | 3.42% | 0 |
| Gravitee BERT .30 | 68.65% | 81.96% | 0.86% | 0 |
| OpenMed 44M .10 | 92.12% | 93.45% | 2.15% | 0 |
| NVIDIA GLiNER .50 | 83.26% | 87.78% | 0.83% | 0 |
| OpenAI Privacy Filter | 64.76% | 75.96% | 0.79% | 53 |
Errors are processing failures, separate from incomplete detections. TracerYX uses threshold 0.20 in the latest and external views; sensitivity presets are 0.05, 0.20 and 0.40.
How to read the comparison
Coverage is class-agnostic reference-span coverage, not typed-entity F1 or a privacy probability. Character F1 balances masked-character precision and recall. “Outside removal” measures masking outside annotations; omitted gold identifiers can inflate apparent excess removal.
Unresolved annotations are tracked separately. Processing errors stay in the coverage denominator with zero detection credit. Thresholds were frozen before acquiring test data; external systems are comparator dependencies only.
The latest primary cohort has no finance test records after acquisition exclusions. Its 797 retained rows contain seven unresolved Nemotron annotations; 124 valid released documents have at least one reference miss. OPF’s CPU context window was 2,048 tokens. Integration errors affect external scores and do not establish those models’ best achievable performance.
Built around practical workflows
Protect the values.
Keep the context.
Privacy protection has to do more than mask every suspicious sentence. An instruction like “Keep it refrigerated” should remain useful when the adjacent customer name is protected. TracerYX combines field context, identifier boundaries and learned discrimination to make that distinction.
Agent & application traces
Protect supported text and attributes in OTLP HTTP JSON traces before they reach a collector. Preserve validated trace relationships, structural IDs and timestamps.
Support & operational context
Replace repeated contact values with consistent aliases within the same batch, so a reviewer can follow an interaction without needing the original email or name.
Domain-specific evaluation
Use your own representative examples to choose sensitivity. Our published studies cover general synthetic text, finance, conversation and separate legal and external-document diagnostics.
These are integration use cases and evaluation directions, not claims that every format or domain is qualified. The current local SDK supports Node.js 22 and strict OTLP HTTP JSON traces; protobuf, gRPC, compression, logs and metrics are outside that integration’s current contract.
02 / First-party technique
Context, boundaries and policy
working together.
TracerYX combines two first-party sparse contextual taggers with deterministic patterns, contextual extraction and a gradient-boosted corrector. Competitor models are used in offline comparisons, not in the product’s inference path.
Validate & map
Validate policies and resource limits. Normalize text while preserving mappings back to original character offsets.
Detect & discriminate
Combine contextual margins with 92 numerical features in a 160-tree corrector. Mandatory rules and configured known values remain active.
Resolve & protect
Repair supported identifier boundaries, combine findings and apply redaction, typed reference replacement or sensitive-value discard.
Choose a sensitivity that fits your workload.
The sensitivity study evaluated High coverage (0.05), Balanced (0.20) and Preserve text (0.40). Explore these fixed operating points in the benchmark selector above. Lower thresholds increase coverage at a measurable text-preservation cost. Scores and thresholds are not privacy probabilities.
03 / Local transformation
Useful traces. Opaque identifiers.
The installed SDK protects supported OTLP HTTP JSON traces before export. Repeated exact values of the same type share an alias within a call or batch; later batches receive new aliases. Validated structural trace/span IDs and timestamps are preserved.
Synthetic input
customer_name: "Jordan Patel" user_email: "jordan.patel@example.com" backup_email: "jordan.patel@example.com" customer_id: "CUST-99821" authorization: "Bearer synthetic-secret" handling_note: "Keep it refrigerated"
Observed behavior / IDs abbreviated
customer_name: "[PERSON:…]" user_email: "[EMAIL:same-token]" backup_email: "[EMAIL:same-token]" customer_id: "[ACCOUNT:…]" authorization: "[discarded-sensitive]" handling_note: "Keep it refrigerated"
The field behaviors above summarize the installed SDK E2E fixture; the bearer value is abbreviated for presentation. SDK aliases are ephemeral pseudonyms, not a reversible vault. The separate gateway reference store adds tenant/owner/scope and transaction controls; those guarantees do not apply to the SDK alias map. A token cannot protect an identifier the detector missed.
04 / Calibration research
Test statistical protection
alongside usefulness.
Our offline document-level conformal experiments score every annotated private character and use separate calibration data to choose expansion thresholds. This examines both residual reference misses and the cost of extra masking.
Historical TracerYX + conformal 5% experiment, on a 303-document cohort. Outside-annotation removal was 54.26%, with no held documents. The gain in coverage carries a substantial usefulness cost.
In the latest 1% experiment, zero released-document reference misses accompanied 400 withheld inputs out of 797 and 81.83% removal on valid released text. No tested joint policy passed all source-level limits for misses, average removal, removal-tail risk and withholding.
Conformal is an offline evaluated prototype, not a certified SDK release policy. Customer-adjusted thresholds have no statistical certificate. The supplied healthcare manuscript’s population re-identification model, patient-level betting bounds and review optimizer were not implemented. Annotation coverage is not a re-identification guarantee.
05 / Release evidence
Qualify the complete path.
Privacy tests passed
Retained-engine implementation suite rerun on the droplet; behavioral checks are separate from population accuracy.
Native parity documents
Zero threshold-decision mismatches against the Python reference across the recorded development cohort.
Installed SDK assertions
CLI and exporter flows checked tokenization, redaction, schema preservation and safe failures in the October 2 E2E run.
The controlled E2E run used a built account console, offline package installation, a real Node OpenTelemetry exporter, the packaged sidecar and a local collector. Malformed input was rejected without forwarding. Upstream failures received protected payloads with no raw fallback.
Release boundaries and current scope
The latest experimental candidates did not pass all release guards; the existing TracerYX detector was retained. Research and E2E execution used curavoice-droplet. Payment responses and release approvals in commerce tests were synthetic fixtures. Production sales require separate operator configuration, signing trust and legal/model redistribution approval.
Residual detection misses, benign-text false positives, long-input limits and domain differences remain. In the latest sensitive-value diagnostic, 190/192 credential/payment references were fully detected and 154/192 fully discard-typed. These results do not establish perfect redaction, universal leadership, multilingual support, clinical readiness or a compliance certification.
Availability & next steps
Evaluate TracerYX Cloak
on your workload.
The local processing preview and integration evidence are available for review. Contact our team about evaluating representative traces in your own environment and selecting a coverage-preservation operating point.
Commercial release approval is separate from benchmark performance. Signing trust, legal/model redistribution approval and production operator configuration must be completed for the distributed release. This introduction does not announce unrestricted production availability or a certified privacy guarantee.
06 / Benchmark register
A versioned evidence trail.
Earlier results remain historical evidence. Some cohorts were reused; document counts across studies must not be added as independent tests.
| Study | Evaluation | Role |
|---|---|---|
| Initial smoke / first owned model | 42 authored cases; initial and revised 400-record Nemotron runs | Baseline history |
| Early conformal / Gretel transfer | 400 Nemotron; 1,000 Gretel test documents | Calibration research |
| Context v2 / three datasets | 1,650 Nemotron, Gretel and PII-TRACE units | Historical comparison |
| Discrimination experiments | Rejected retraining candidates; 506-document narrow confirmation | Regression and release gates |
| Context-corrector comparison | 303-document comparison plus 32 TAB judgments | Context-corrector improvement |
| Sensitivity and tokenization study | 311 synthetic; 32 TAB; 90-document correction confirmation | Sensitivity and tokenization |
| Latest TracerYX evaluation | 797 primary; 200 RedactionBench | Current benchmark headline |
| Local SDK E2E | Account, package, exporter and collector checks | Integration verification |
Public dataset sources
This page summarizes existing internal engineering reports; it is not an independent audit or an official dataset leaderboard. English finance was included in earlier comparative studies. The human-annotated TAB adaptations differ from official subject-level scoring. RedactionBench diagnostics are not the official R-Score.