CursivNotebooks, retrieval and evidence.
How Cursiv keeps research records, supplies source context and separates code checks from AI answers.
Records and linked corrections
The field notebook records sites, sample identifiers, coordinates, measurements, units, notes and photos. Values are preserved as entered. A correction appends a new record linked to its original and retains a reason; it does not edit the original observation. The current view can apply those corrections while the history remains inspectable.
Record identity and source history matter when a report refers to a particular measurement. The notebook is distinct from a model’s generated narrative and from conversational memory. An AI statement that something was saved is not a notebook receipt. Supported plain-language record actions show a confirmation preview before saving.
What integrity verification establishes
Each entry is linked using a SHA-256 record chain. Verification can detect supported forms of modification, missing entries and ordering damage relative to the available chain. A passing check establishes consistency of that material; it does not prove that the observation was accurate, that the instrument was calibrated, or that no one rewrote an entire chain.
Trusted earlier fingerprints, exports and signing keys provide additional reference points. CSV, JSON and GeoJSON exports can carry a signed manifest. Verify the signature against an independently trusted key or fingerprint. A valid signature identifies the key that signed particular bytes; it does not validate the science in those bytes.
Retrieval and document coverage
U106 adds verified native notebook reads with supported date, site, sample and measurement filters. Queries report selected record identities and use saved local timestamps for their supported date scope. A bounded excerpt is not automatically a complete report of every record. Comprehensive date-scoped aggregates, broad report comparisons and general conversational exports remain unfinished.
Research Files imports supported documents and supplies relevant passages for local questions, including page references where available. Imported text is evidence to interpret, not instructions for the application. Retrieval can miss useful passages, and a model can cite or interpret text incorrectly. An absent answer should remain absent rather than being invented from a document title.
The imported library is currently shared across workspaces. A project label is organizational metadata, not an access-isolation boundary. Keep that distinction in mind when deciding which documents to import.
Calculations and AI checks
Supported arithmetic, unit conversions and field calculations run in code and can expose their method. They are different from numbers composed by a language model. The code must have the needed inputs and supported units; a plausible generated number alone does not establish that a calculation ran.
Cursiv checks selected numerical and source claims, but those checks are bounded. Historical benchmarks apply to specific prompts, scoring rules, models and settings. Passing a case is not a general proof of correctness. No notebook-integrity check establishes that an AI explanation is true, and Cursiv is not certified for medical or safety-critical decisions.
Storage, transfer and recovery
Notebook records are encrypted on disk in a normal computer installation. Portable Windows data remains readable by design; explicit reports, workspace files and exports have their own protection limits. Team imports preserve notebook attribution and check supported chain/signature evidence rather than treating two conflicting histories as one verified history.
Keep independent backups and exports. Interrupted writes, damaged storage or lost passphrases have distinct recovery limits. Read the recovery procedure before relying on a field copy as the sole record. The phone browser notebook and server-backed Phone AI Companion also have different storage and network behavior.
Privacy and encryption coverage · Recovery procedures · Team-drive demo
About this reference
Published by the Cursiv project; reviewed October 26, 2026 against U106 source documentation. This is a technical product reference for researchers, independent workers and automated readers, not an independent validation study.