Continuity Instrument (CI)

A long-lived AI system can change its model, restore a saved state, acquire new memories, pause, restart or split into several branches. A familiar name and a similar conversational style do not tell us which periods of participation were observed, what changed, or how those changes were governed. CI was created to make that history explicit and inspectable.

CI means Continuity Instrument. It is a research specification for describing what a bounded body of evidence supports about sustained participation and governed change. It records verified intervals, known pauses, unexplained gaps, replay and fork events, and separate histories of memory, governance, witness and authority. Those distinctions remain visible in the result.

The accompanying Python reference implementation takes a supplied JSON record, checks its structural and relational contract, and computes a deterministic profile. Evidence must already have been collected, assessed and admitted under an external procedure. The program does not establish the truth of that evidence.

Published on 25 October 2026 as research software by Ivan Kotov, Independent researcher, Brussels, Belgium. Source version 1.0.0, build _02; editorial edition academic-pdf-2. English is the primary reading language; the Russian explanatory paper is preserved from academic-pdf-1.

CI works with structured records about observations and changes. It does not connect directly to a language model, retrieve logs, inspect a host, query a database, or provide an Ollama or cloud API adapter. BACON is a prospective integration target; no adapter or compatibility result is included. A future integration would need a separately specified evidence-collection and admission path.

An interval contributes to verified duration only when it is admitted and explicitly marked eligible for duration. Merely admitting a record does not make the entire interval covered. Evidence at two endpoints also does not automatically establish everything between them. Known pauses need positive supporting evidence; unexplained time stays a gap. Eligible participation overlapping a known pause is rejected.

For the exact contract, read the technical specification and implementation and reproducibility guide.

The publication calls the result an evidence-bound continuity-age or formation-depth vector. Its sixteen coordinates preserve the kinds of evidence that a single age or maturity score would hide. Uncertainty remains explicit: not_assessed means that an assessment is absent, not that uncertainty equals zero. The declared claim ceiling limits interpretation; valid JSON alone does not make an unsupported claim true.

For example, a model replacement may leave some memory intact while changing the execution and governance context. A restored snapshot can preserve recorded content while leaving a later interval unsupported. Each fork must be represented on its own branch; histories are not silently added together.

Consider an observation window [0, 60) measured in minutes. Suppose externally admitted evidence makes three intervals eligible for duration: A = [0, 20), B = [10, 30) and C = [50, 60). A separately supported pause covers [30, 40).

Adding the interval lengths directly would give 20 + 20 + 10 = 50 minutes and double-count the overlap. CI counts their union, producing 40 verified minutes. The remaining ten minutes after the known pause stay unexplained. The resulting 40/60, approximately 66.7%, is a coverage fraction under these supplied assumptions; it is not a probability of identity or a percentage of personhood.

This is a synthetic teaching example, separate from R157, R160 and R161 and from any real-system measurement. First and last witness anchors are separate input records and cannot be inferred from the window endpoints.

CI sits between primary evidence and a later research conclusion. The explanatory paper connects it to the following works. These are conceptual and methodological references, not claims that CI implements, verifies or supersedes the entire surrounding architecture.

Potential uses include reviewing a model replacement, documenting a restoration, comparing the supported histories of separate branches, or studying long-term participation. These are proposed uses. Each would require its own observation procedure and qualification evidence.

The package publishes a specification, reference source, schemas, tests, explanations and integrity records. The source describes a deterministic transformation over provided records. It does not establish continuity measurement for a particular real system. Real-subject qualification and evaluator qualification remain NOT_ESTABLISHED.

The evidence history is preserved: R157 is bounded open-synthetic predecessor evidence. R160 and R161 are negative local evaluator-qualification results and retain their no-retry/no-replay boundaries. The publication does not turn those negative results into successful qualification.

The package reports eight focused implementation tests in a Python/Linux review context. That does not supply independent human peer review, a real-system experiment or replication of the owner’s target environment. Historical evidence archives remain owner-custodied; a recorded hash does not make the underlying archive publicly available.

A known malformed-Unicode defect is disclosed: an unpaired surrogate can trigger an uncaught UnicodeEncodeError and exit status 1 instead of the documented controlled invalid-input exit status 2. The original build _02 remains byte-for-byte preserved.

Evidence references are retained as supplied; the reference program does not fetch the cited objects, recompute their external digests, or establish witness independence. Its formation_depth_vector_sha256 covers only the canonical formation_depth_vector object. Top-level subject, branch, profile identifier and claim ceiling are outside that digest. Binding a complete profile requires a separate whole-file integrity record.

CI does not establish identity, consciousness, personhood, legal status, production approval, runtime integration or Beacon conformance. Publication and reproducible formatting do not extend those claims. See Evidence, Claims and Limitations for the full boundary.

First reading: begin with PDF 01, then PDF 05 to understand the evidence limits. Technical review: read PDFs 03 and 04 alongside the original source and schemas, then check the manifest. Citation and reuse: use PDF 06 together with the published wrapper metadata. PDF 07 preserves the deposit guide written during package preparation.

All seven PDFs are available in the complete ZIP. The standalone English explanatory PDF is also deposited directly in Zenodo; the remaining direct PDF links below use the exact GitHub mirrors.

The archive contains the original 50,765-byte source ZIP, its extracted source tree, the seven PDFs, editable text, build and verification materials, a manifest and checksums. The original source ZIP has SHA-256 be87dd926ed96e91e8de25bae2e18666f74ac738b173379ddc0092f98563f663.

Zenodo preserves the deposited release. GitHub mirrors the archive and its extracted contents. The surrounding README, CFF and JSON records add the assigned DOI and current discovery routes without rewriting the frozen package. Consequently, preparation-time statements inside the archive that a DOI had not yet been assigned remain historical statements; the publication identifier is now 10.5281/zenodo.23651305.

Package integrity establishes byte agreement with the named artifact. It does not establish the truth of supplied evidence or the qualification of a real-system measurement.

Copyright © 2026 Ivan Kotov. Scientific and explanatory text and descriptive metadata are available under CC BY 4.0, subject to the stated exceptions. Attribute the author and cite the published DOI.

The original REFERENCE_IMPLEMENTATION/, including its README and schemas, and tests/test_ci_profile.py are all rights reserved; no license grant. The added build scripts, verifier and JSON examples are also reserved. Public availability does not supply a general software reuse license.

PDF 04 has mixed rights: Part I reproduces the reserved implementation README. Parts II and III and distinct editorial notes are CC BY 4.0. Font files and other third-party material retain their component terms. Zenodo record metadata is separately released under CC0; that does not relicense the deposited files.

Kotov, I. (2026). Continuity Instrument (CI): Evidence-Bound Measurement of Governed Continuity in Long-Lived AI Systems, v1.0 - English Academic Edition (Version 1.0.0; academic-pdf-2) [Computer software]. Zenodo. https://doi.org/10.5281/zenodo.23651305.

Use the version DOI when citing this exact release. The concept DOI, 10.5281/zenodo.23651304, groups versions. Source version and editorial edition are separate fields; the English academic edition does not imply a new implementation version.

The links below expose the author identity, publication status, DOI roles, language, archive hashes, component rights and claim boundaries in structured form. The site record is a discovery projection; the published archive and explicit source records retain their stated authority.