DATALOGIX CAPABILITIES
The full inventory of DataLogiX: its eight strengths, three more areas built on them, what has been measured, and where the limits are.
AI-assisted, never AI-decided: analytics, reliable reporting and document control for regulated organizations.
6,593 of 6,593
contract fields accepted without a person were right
12+ digits
agreement with NIST reference statistics
662 of 664
CMS home health grouper test cases matched
91 of 91
planted unsupported sentences caught
117 of 120
questions found their passage in the top ten
0
planted patient identifiers reached a log
From tests and recorded runs. Chapter 12 lists the measurements.
01 · EVERY NUMBER VERIFIED
Models propose plans, queries, facts and wording. Code checks each proposal, computes every number, computes it again a second way, and refuses what it can’t prove.
31 rules
checked against every plan before any data is read
94 to 98 of 100
planted SQL disagreements told apart by a small distinguishing example, against a bar of 90
1,000
random tables the suppression audit tried to recover a hidden count from, with none recovered
A model reads the question and drafts the plan, the queries, the extracted facts and the wording. Every decision that could change a number belongs to code that can be tested.
Code checks each plan against 31 rules before any data is read. A person reviews it where policy asks, always for an organization’s first analyses, and the approval names the plan’s hash, so a plan that changed can’t be approved by mistake.
Every query is one read-only statement over allowlisted tables, written in the source’s own dialect and validated there first. Three candidates are written; when they disagree, small tables built to tell them apart show where their readings differ.
Metrics, dimensions and joins are defined once and compiled into each database’s dialect. A published version never changes, and publishing needs passing checks on the target database.
No model writes a number. Every figure in a report is a binding to a registered value with its lineage, and the model that writes never sees a data row.
A critic model reads each sentence beside the evidence it cites, and code refuses a sentence the evidence doesn’t support. A level can never stand in for a rise or a fall.
Before release, every value and published table is computed a second, independent way. A mismatch blocks release; a value with no second way waits for a named person.
Each value is read back from the Excel, PowerPoint, Word, PDF and web files and must match what was registered. Bar and column axes start at zero, and a file that can’t be read stops delivery.
A value is computed, insufficient, conflicting or degraded. When sources disagree, copies count once and your precedence rules apply; values are never averaged.
“Trend”, “outlier” and “significant” appear only where a test supports them, with every comparison in a report adjusted together. Small counts are suppressed and checked by an attacker audit.
Growth, contribution, price-volume-mix, regression, difference-in-differences, survival curves and Cox models, logistic regression, clustering and forecasting with intervals, each checked against an independent implementation.
Lean 4 references for key functions run against the product every night, and rule catalogs are checked for contradictions by SMT solvers. Each method is also checked against NIST reference values and a second, independent implementation.
One number or a small table in seconds, still guarded, registered and computed twice. A question that needs more becomes a full report with its data and dates kept.
Choropleth and point maps have bound legends, print in every format and never use online geocoding. Click into a bar, a slice or an area, and every number you reach is registered and recomputed.
Golden questions run every night on your own data, edits to reviewed reports are counted, and a weekly report card shows recomputation, citation support, calibration and abstention with 95% intervals. A go-live check reads them.
Releases, signatures, disputed facts and any figure without a second computation are a person’s call, and every file prints whether a person reviewed it.
02 · FULL TRACEABILITY
Each sentence, number, chart and marking traces to its origin in a signed trace, and every Office file and PDF you send carries an appendix printed from it.
Every element
of every delivered file has a trace record, located in the file and compared with its printed text
Any byte
changed in a signed file fails verification
Every element has a record: its origin (a query, a document page, a function, a rule or a person), whether it is anchored, the checks it passed, any review by a person, its marking and a hash of its printed text.
Each element is located by its place (Excel cells, PowerPoint shapes, Word bookmarks, PDF destinations, web paths) and its printed text compared. A change outside a known element is caught too.
Every Office file and PDF ends with an appendix printed from the trace: coverage, what isn’t anchored, and each element’s origin, checks and source, linked back to the element.
Every number is a chip that opens its query, rows and data snapshot. Inspect mode marks each element as anchored or not by line style, never by color alone.
Two versions side by side: values changed, added or removed as each version printed them, sections added or reworded, and a line-by-line diff of the text.
PDFs are signed with PAdES and a trusted timestamp. Excel, PowerPoint and Word files carry a C2PA manifest that marks them as made with AI assistance. Signing comes last, after every check.
Anyone can check a file without signing in, on the trust center, through the API or from the SDKs: valid, changed, not signed by us, or unsupported. Adobe Acrobat and Content Credentials tools read the signatures too.
Rows keep their file, locator, parser and hash. Queries keep their exact SQL and data snapshot. Values keep their second computation. Facts keep their quote, page and box.
Every file says whether a person reviewed it. After a review chain signs, every file is rendered and signed again, naming the signer’s role and day.
04 · ANY DATA
DataLogiX reads data where it lives, in almost any format and protocol, checks what arrives, and keeps the provider’s rules with it. No migration project comes first.
117 of 120
questions found their passage in the top ten of document search
Every time
a hostile test corpus of bombs, malware samples and encrypted archives is held before any parser reads it
SQL Server, PostgreSQL, MySQL, MariaDB and Oracle on premises through an outbound-only relay; Snowflake, Redshift, Databricks and other cloud warehouses in place; Microsoft 365, Google Drive, accounting systems, Salesforce and REST APIs; document databases and vector stores.
Browser upload up to 5 GiB a file, an SFTP drop, an email-in address, watched folders and Azure Files shares.
CSV in any dialect, fixed-width, Parquet, Avro, JSON and XML; workbooks with merged cells and pivots; PDFs and scans, Word, PowerPoint, email and archives; audio. HL7 v2, FHIR, C-CDA, X12, CMS assessments, DICOM metadata and clinical-trial formats. Bank statements, e-invoices and EDI, ledgers and XBRL.
An e-invoice failing an EN 16931 rule, an unbalanced ledger entry, a bank statement whose entries don’t reach its closing balance and an X12 835 that doesn’t balance are held whole.
Workbooks load only when every total and subtotal reconciles exactly. A text layout never seen before gets a parser drafted, checked by code against the source and confirmed by a person.
Columns are typed by deterministic detectors first. Metric definitions are drafted from the schema, validated on the database and reviewed by a person before they are published.
A text pass and a vision pass read each document; every value needs its quote verbatim on its page, and a person reviews disagreements beside the PDF page. Scans are read by OCR.
Password-protected files open with a password supplied once or a standing rule per source. OpenPGP, S/MIME, ZTDF and Purview files open with your keys or grant, and signatures are checked against the senders you register.
Truncated PDFs, broken archives and damaged Office, text, JSON and XML files are repaired automatically, and what was lost is counted and shown. A PDF that can’t be repaired is read page by page from its images.
Types are decided by content, never by extension; a malware scan fails closed; every archive is walked within limits; macros are held. Held files explain themselves and are deleted after 30 days.
Health and personal data, government identifiers, secrets, marking lists, prohibited terms and lists of known-leaked records. Per rule: alert, hold, quarantine, reject, redact, drop the rows or pause the source. Findings are kept as counts, never values.
Markings are read from labels, banners and portion marks and never fall below the file they came from. Labels join through every step; flow policies decide every query, prompt, export and delivery; model policies decide which models may see each kind of data.
Purposes, territories, recipient kinds, expiry, AI use, retention and attribution, attached to a source, a dataset or a document, enforced in every run and printed in every file.
Drift, freshness and quality checks without data values in the alerts; restated periods detected; withdrawn and corrected documents leave citations; one report arriving by two feeds counts once; units convert only under declared factors and dated rates.
Precedence rules by field, metric or entity kind, effective-dated, and scheduled reconciliation pairs that report where two systems agree, differ or miss each other.
Each company’s systems are read where they are, with their own metric models, so “revenue” means the same across systems that store it differently. The same customer, patient or supplier is matched across systems, and a person’s “different” is final.
Meaning and keywords fused by rank, with each hit checked against its marking. Your Qdrant, pgvector and Pinecone collections are searched in place. Related passages are labeled similar, never a match.
Recordings in English, Spanish and other languages are transcribed, with every quote tied to its time in the recording and speakers told apart. Translations are graded A to D.
Text in photographs and scans becomes searchable, with plates, VINs and serial numbers located by box. Captions and tags are labeled as a model’s description, similar images are found offline, and a photo’s time and place are read as the file states them.
Security telemetry in OCSF; SAS, SPSS and Stata files; GeoJSON, KML and GPX; web archives; streams and change data; sensors; and healthcare and HR vendor connectors.
05 · SUBJECT SEARCH
Find people, organizations, places, vehicles, vessels, accounts and identifiers across everything you hold, see how they connect, and be told when something new appears. Every search is governed, logged and explainable.
Hashes only
the search index holds keyed hashes, never values, so the index itself reveals nothing
Purpose first
every search states a listed purpose and a reason, and is logged before it is answered
People, organizations, places, identifiers, vehicles, vessels, aircraft, wallets and any kind a sector pack adds, across documents, extracted fields, records, images and vector stores. Names match through aliases, nicknames, transliteration and look-alike spellings; two people with one name stay separate.
Search by an area drawn on a map, a radius, a country or region, and by a range or a prefix checked in the record’s own typed column.
Up to 1,000 rows matched in one request. A matched list or a saved query is matched again about every hour against new data, and its owner is told what is new.
Link charts, places and timelines with each link’s evidence and whether it was inferred; communities, paths between two subjects, and ranking by degree, betweenness and eigenvector centrality over the reader’s own view.
Packs declare relationships such as “employs”, “owns” or “registered to” and the rules that produce them. A new connection to a watched subject, or a link far outside the usual pattern, raises an alert with its evidence.
Incidents, tickets, security events and transactions are searched by their own fields and joined to the subjects they mention on one chart.
Text in images is searched, each mention with its box, along with captions, similar images and time and place. Face matching runs only against your own consented gallery, never for uses the law prohibits, always with two people, and every match is a candidate a person confirms.
One inquiry’s searches and items with evidence, link chart, map and timeline views, note threads and a marking approved by a cleared, trained person. A case downloads as GraphML, i2 ANX or CSV through the dissemination gate.
A report question can name subjects. Its plan always waits for a person, and for a second person where the rules ask. Off by default.
A second person for sensitive identifiers and kinds; conflicts of interest escalated or refused; deconfliction when two people search for one subject; daily limits; clearance, compartments and need to know; U.S. person masking; oversight of who searched and what they opened, also sent to your SIEM.
06 · EXPERT PACKS
A sector’s rules are data an expert can read, test and approve, not code buried in the product, and a later layer can never weaken a protection an earlier one set.
7 layers
merged in order, from the platform to a single product, with locks that only allow stricter
Every pin
approved by a second administrator, with a one-step rollback
Domain, metric, collection, product, style, tradecraft, marking, access, records, workflow, function and model policy: terms and rule tables, formulas and suppression, report types, writing rules, analytic standards, markings, authorization, dissemination, review chains, approved calculations and allowed models.
Platform, framework, sector, organization, client, product type and product. Review gates, the dissemination tier, subject-search limits and budgets can only be tightened by a later layer.
Type-checked CEL expressions, effective-dated rule tables and access policies. Every pack is validated and its golden tests run before it can be built, and the merged configuration is frozen into each run.
Each pack version is a signed, deterministic artifact with update metadata that refuses rollback and expired versions. Offline mirrors carry packs to disconnected sites.
A candidate pack can be evaluated against your stored reports before it is turned on: markings, banners, verifier rules, terms, workflows and small-count rules.
Home health and hospice payment rules with every row citing its CMS source; analytic standards (ICD 203 and 206, UK PHIA, NATO AJP-2.1, IPCC, GRADE, Europol 4x4, LEAS); marking schemes; records rules; report products; nine citation style guides; subject kinds.
Typed claims checked for dangling links and cycles; graded sources where copies count once; likelihood words printed by code from a stored probability; confidence computed with what limits it; key assumptions varied; alternatives, calibration, indicators and a scorecard.
Reviewers approve, challenge, replace, dissent or waive per claim. Release waits while a challenge is open, and a reviewed product is rebuilt without any model call, verified and signed again.
Rules are shown as forms with plain labels, where each setting comes from and what a lock still allows, with Test, Preview and “Try on recent reports”. Administrators see every pinned pack and approve or withdraw each change.
Every pack version is certified before it is signed: validation, golden tests, determinism and, for code, stress and sandbox-escape probes. Organizations and outside publishers can sign their own rule sets.
Customers read every rule in a rule set, try it on their own recent reports and turn it on with a second administrator, while publishers see counts, never a customer’s name, report or data.
Every report’s cost is metered, model calls and computing included, and usage reports show the cost per verified deliverable.
07 · HEADLESS AND ASSISTANT-READY
Teams keep their tools and assistants and still get verified results. Systems call a documented API, assistants call DataLogiX as their verified tool, and people get a complete web app built on the same API.
540+
API operations in more than 45 areas, each decided by a policy rule
28
tools for AI assistants, signed in as the person
Ask and read, review facts beside their page, connect data, share and release, search subjects and run cases, answer rights requests and administer the organization. Outside recipients open links without an account.
Every screen is checked against WCAG 2.2 A and AA in light and dark themes at phone, tablet and desktop widths, with keyboard paths and clear states for loading, errors and missing access.
OpenAPI 3.1, API keys with scopes or your own sign-in, an authorization rule for every operation, one error shape, signed paging cursors, idempotency keys and long operations you follow to the end.
Each analysis streams its progress with replay from the last event. Webhooks are signed at send time, retried, and replayable from a dead-letter queue.
TypeScript and Python, generated from the API and checked against it in every change, with helpers for event streams, paging, report verification and offline transmittal checks.
Each person signs in with their own account, and the assistant can do what that person may do and no more. Tested live with two enterprise assistant platforms, plus a conformance and attack suite.
A reason, an acknowledgment or a confirmation your rules require is taken in the web app through a one-time link. An assistant can prepare a share; only the person sends it.
Single questions, reports of 2 to 8 questions (6 by default) with a table of contents, pack-defined report products with a review chain, monthly or weekly subscriptions with backfill, and dashboards.
The web, Excel with live formulas, PowerPoint fitted from font metrics on your master, Word in your style guide, and PDF/A-2b. A visual check fixes layout without changing a value, and citations follow 17 styles.
In-app notices built from fixed templates with typed slots, so no free text can leak; email from your own Microsoft 365, mail service or SMTP server; Slack, Microsoft Teams and webhooks.
A second wall per client, with its own key, dataset, people and usage, and reports, subscriptions and dashboards under the client’s own packs.
08 · SECURITY AND DEPLOYMENT
Isolation, encryption, audit and model routing are built into the product and tested in every change, so security reviews go faster.
Not found
is the answer every endpoint gives another organization’s credentials, in a suite run on every change
0
planted patient identifiers that reached a log or telemetry export
Row-level security forced on every tenant table, a signed organization context in every transaction, its own cloud project, keys and service identities. Every endpoint is tested with another organization’s credentials.
For providers: a signed client context, a per-client key and dataset, and client-scoped people, keys and usage. Offboarding a client destroys its keys.
Sensitive columns are sealed per organization, keys rotate every 90 days, and leaving destroys an organization’s keys so its data can no longer be read, checked by a cloud test.
Health data goes only to models and services cleared for it, through an input and output screen that fails closed. Logs and telemetry are scrubbed, and a canary run requires zero planted identifiers anywhere.
Models are named by alias in one lineup; changing one needs its evaluations to pass; prompts are versioned and locked; a contract can exclude a model family.
Every authorized request writes exactly one audit event, and a failed audit write fails the request. Events go to Splunk, Microsoft Sentinel, a signed webhook or files, and an access review lists who can reach what.
Dependencies at least seven days old and pinned; signed images verified by signature and software bill of materials; a security gate that fails on any high finding; a prompt-injection corpus; model-written code only in no-network jobs.
A managed service in the United States, with recovery copies in a second US region; your own cloud or data center, with your own sign-in and key store; a signed offline install with local models and no call-home; a government enclave; or a forward node at the tactical edge. Under each, an outbound-only relay reads databases and file shares that stay on your network.
Each organization’s model policy decides which models may see each kind of data, and a contract can exclude a model family. Health data reaches only models cleared for it.
Clearances, compartments and need to know decide every search and every copy, markings are computed for every portion, and every pack change needs a second administrator. Enclave installs sign people in with smart cards and use FIPS-approved algorithms only.
28 CFR Part 23 rules for criminal intelligence, access records kept at least 365 days, and a purpose, reason and oversight for every subject search.
Internal objectives for API success, analysis success and time, and event delivery, with burn-rate alerts; uptime checks every minute; recovery targets of 15 minutes of data and 4 hours for the control plane; resilience tests under outages.
In the managed cloud, in the United States, with recovery copies in a second US region. In every other deployment, on your own infrastructure.
Support never needs health data: an analysis id, a request id or an error code is enough. Staff have no standing access to an organization’s data.
09 · STREAMS
Built for questions that can’t wait for a report. Items arrive one at a time, each verified on its own, and judgments and stories revise as evidence arrives, with “verified within minutes” stated per feed.
3 tiers
of checks per item: code on every item, a second model on material items, a person on items that would move a judgment
Feeds, cloud message topics, signed webhooks, Kafka and database change data, with the customer’s own authentication, delivered at least once in and exactly once in effect.
Schema, quote found in its source, number bindings, source independence and a valid marking on every item; every claim carries the time it was last verified.
New evidence makes a new revision with an explanation, every earlier revision stays readable, and a retraction re-evaluates everything that cited it.
Server-sent events and WebSocket with replay, webhooks, Atom and JSON Feed, STIX 2.1 for security topics, and assistant tools, filtered by topic, subject, confidence and state.
Tables kept current by change data say the time they are current to; triggers start a report or an alert over a window; backpressure pauses rather than drops.
Changes to the eCFR, the Federal Register, Regulations.gov dockets and CMS memos are matched to the obligations that cite them, with an impact report.
10 · SENSING
SensorLogiX adds sensing, image and video search, and analysis on your own GPUs to DataLogiX.
Provisional
describes every detection a model makes, until a person or a second sensor confirms it
Video, imagery, moving-target radar, tracks and telemetry from drones, aircraft, ground vehicles, robots, ground sensors, ships and unmanned vessels, over any link.
MISB motion imagery, NITF and SAR imagery, STANAG 4607, Cursor on Target, AIS, ADS-B, ASTERIX, Remote ID, MAVLink, ROS 2 and SAPIENT.
Recorded and live video, still images and imagery, searched by object and attributes, area and time, text in the frame or a similar picture, where your rules allow. A question in plain words becomes a structured query, shown to you and recorded in the audit.
Tracks from different sensors fused by deterministic rules into situation products, alerts and your TAK, confirmed items only.
Vehicle, vessel and aircraft attributes come from closed vocabularies and show only where their measured rate passes. The same vehicle or vessel across cameras is a candidate link a person confirms. Change between images of one area comes with its areas and counts computed twice, and copies for public release blur faces, plates, bystanders and chosen areas while the original is kept.
Edge collectors buffer through lost links. A forward node on rugged servers built to military and maritime standards runs the analysis fully disconnected, with smart-card sign-in, local models and a FIPS-mode stack hardened to DISA STIGs, every component and license published. It syncs with the rear by link, encrypted drive or your cross-domain solution, under both ends’ rules.
11 · COMPLIANCE AND GRANTS
Obligations are kept with their text and citations and given verdicts from evidence, grants are managed from opportunity to post-award, and community evidence is built without pooling health data.
Never a guess
a verdict whose check couldn’t run is degraded and holds release until a named person overrides it
Each obligation is kept with its citation, text versions and applicability, and gets a verdict (met, not met, insufficient evidence, not applicable or degraded) computed by rules from evidence. Security packs import and export OSCAL.
Survey windows, monthly re-screens and attestation chains (who signs, in what order, with which evidence) on one calendar.
Staff and vendors screened monthly against the OIG exclusion list and SAM.gov; every near-miss goes to a person, and nothing is cleared without one.
Hospice and home health conditions of participation, survey procedures, HIPAA, 42 CFR Part 2 and OIG guidance, with a survey-readiness binder mapped to survey tags.
A deterministic CMMC SPRS score from your own evidence, and the Uniform Guidance and HHS rules with SF-425 reconciliation and single-audit readiness.
Opportunities found and scored for fit, funding notices turned into reviewed rule sets, SF-424 forms and budget math, every mechanical reason for a return caught before a person submits. People write the narrative and sign; a mock review is never shown as a predicted score.
Addresses become census tracts inside your own project, public data carries its data year, and benchmarks across agencies come from pooled aggregates only.
Each member’s data in its own compartment with its own key; sharing agreements enforced as data; small cells suppressed with an attacker audit; publication after every member validates its own figures.
12 · MEASURED QUALITY
From live or recorded-live runs, from oracle tests, and from gates that run on every change. Offline replays don’t count as measurements.
| AREA | MEASURE | RESULT | GATE |
|---|---|---|---|
| Statistics | NIST reference values, lower and average difficulty | 146 values, at least 12.2 correct digits | 9 digits |
| Health rules | CMS home health grouper test cases | 662 of 664 | — |
| SQL | BIRD benchmark subset | 23 of 30 (76.7%) | 70% |
| Planning | Golden plans | 21 of 22 complete; 8 of 8 traps caught | 90%, every trap |
| Extraction | Contract fields, digital documents (F1) | 0.999 | 0.97 |
| Extraction | Contract fields, scanned documents (F1) | 0.9903 | 0.92 |
| Extraction | Contract fields accepted without a reviewer that were right | 6,593 of 6,593 | — |
| Extraction | Clinical document fields (F1), 30 documents | 0.984; 7 of 1,469 accepted fields wrong | — |
| Search | Right passage in the top ten | 117 of 120 | 0.95 |
| Images | Photos whose plate, VIN or serial label is found with its box | 29 of 30 | 27 of 30 |
| Drafting | Expected metrics and dimensions drafted from a schema | 32 of 32 | — |
| Tradecraft | Sentences typed as fact, assumption, judgment or estimate | 397 of 400 | 392 |
| Tradecraft | Planted fragile assumptions flagged | 38 of 39; 8 of 69 falsely flagged | 90%; at most 15% |
| Forecasting | Beats a naive forecast, 30 series | 28 of 30; 80% interval coverage 0.761 | 0.8; 0.7 to 0.9 |
| Engineering | Authorization p99; event delivery p95; webhooks delivered once | 2.5 ms; 10.4 ms; 100 of 100 | 5 ms; —; all |
On constructed audit sets, with 95% intervals. A high false-alarm rate sends work to a person, never past one.
| VERIFIER | HIT RATE | FALSE ALARM |
|---|---|---|
| Plan critic | 96.2% [89.3, 98.7] | 40.5% [26.3, 56.5] |
| SQL pairwise critic | 95.6% [87.8, 98.5] | 0.0% [0.0, 5.3] |
| Document second read | 100% [97.3, 100] | 0.0% [0.0, 2.8] |
| Entity pair read | 100% [94.0, 100] | 0.0% [0.0, 6.0] |
| Claim check | 100% [95.9, 100] | 2.0% [0.4, 10.7] |
13 · LIMITS AND SCOPE
What is off until you turn it on, what isn’t offered, and the limits you will meet first.