Evidence

What was measured, on what, and what was not

Three rules govern the project. Every claim is labelled as measured, designed for or hypothesised. No mechanism enters the specification without a published catalogue that demands it. Every prohibition names the counted failure it prevents. This page applies the first rule to the project itself.

measured designed for open

01 Conformance

The test suite is the specification's referee

A tool is conformant when the corpus says so. On 2026-11-15, at commit 967e9f3, the reference validator passed 822 of 822 vectors in 18 families and re-verified both digests of 4,483 objects in the 5 bundles this site serves.

  • jcs22
  • modality21
  • parameter22
  • tailoring19
  • attestation32
  • facet29
  • reference56
  • lifecycle58
  • tier10
  • dsse27
  • composition11
  • conditional8
  • risk73
  • kernel97
  • scale18
  • check205
  • advice74
  • author40

The released 1.0.0 has 157 vectors in twelve families. risk, kernel and scale came with 2.0.0-rc.1, the 205 check vectors with 2.0.0-rc.2, the 74 advice vectors with 2.0.0-rc.3.

validate_core.py

measured at build

The reference: one Python file of 3,535 lines, with jsonschema as its only dependency. It includes canonicalization, both digests, the tailoring law, Ed25519 and DSSE verification, bundle composition, risk coverage and an I-Regexp engine.

validate_core.ps1

measured 2026-11-12

The twin: 4,836 lines of Windows PowerShell 5.1 with no dependency, for the machine an auditor already has. On a hosted Windows runner it returns the same counts as the Python reference on all 822 vectors, the five bundles, the worked pack and the example folder. A workflow repeats that run on every change.

A third, in the browser

measured

The app's Validate page runs a JavaScript port of the kernel rules against all 18 vector families. The app's readme lists what it leaves to the Python validator.

Line counts use the counter of the gate-4 report: non-blank lines that do not start with #.

02 Corpora

Eleven conversions measured, five passed on

Each converter writes a coverage report that states every rule with its count and reports defects of the source without repairing them. Five bundles may be redistributed under their publishers' terms and are served here. Open any of them in the app.

CorpusObjectsTypesLicence of the content
AU.ISMInformation Security ManualAustralian Signals Directorate1,7311,150 Requirements · 581 RequirementSets1,150 controls; classification and Essential Eight baselines as SetsCC BY 4.0Open in the app
DE.BSIGrundschutz++ AnwenderkatalogBSI, Germany1,2891,000 Requirements · 289 RequirementSetsGerman prose, statement grammar and protection goals as facetsCC BY-SA 4.0Open in the app
US.CSFCybersecurity Framework 2.0NIST226185 Requirements · 41 RequirementSetsFunctions, categories and subcategories as nested SetsCC0 1.0Open in the app
US.IFA-GoodReadOSCAL example system (SSP, plan, results, POA&M)NIST, fictional content166 Components · 3 Implementations · 2 Findings · 2 Requirements · 2 Risks · 1 AssessmentThe lifecycle types: Component to Attestation, with two RisksCC0 1.0Open in the app
US.SP800-53SP 800-53 Rev 5.2.0 with the 800-53B baselinesNIST1,2211,196 Requirements · 25 RequirementSetsStatement-scoped parameters; four baselines as membership SetsCC0 1.0Open in the app

Object counts measured at build. No publisher produced, reviewed or endorsed these conversions. The Apache licence of the project does not cover the bundles; the licence note beside each bundle does.

Measured, not passed on

Six further corpora were converted and measured for 1.0.0. Their bundles were removed from the repository on 10 November 2026 after a review of the publishers' terms. The converters and coverage reports remain, so anyone entitled to the source file can reproduce the bundle.

  • CIS Controls v8.1 and the CIS Ubuntu 24.04 LTS Benchmark: the terms forbid derivative works.
  • BSI C5:2026, BSI C3A and the Belgian CyberFundamentals 2025: no licence statement was found.
  • FedRAMP CR26: a U.S. Government work whose repository states no licence.

The 1.0.0 measurement 22 July 2026

  • Eleven bundles, 6,914 objects, both digests re-verified per object.
  • 251,681 leaf values in the sources, every source path with a declared destination.
  • OSCAL 1.2.2 export of ten catalogue bundles: valid against NIST's release schema, 5,886 of 5,886 objects equal by semantic digest after re-import.
  • Size against the sources: ISM 29 % smaller, Grundschutz++ 50 % smaller, CR26 55 % larger. The format is smaller where a source is redundant and larger where a bespoke format was terse.
The gate-4 measurement report

03 Gates

Nothing advances on narrative

mid-2026

Gate 1: the census done

The ISM (1,150 controls, 8,958 props), Grundschutz++ (13,315 props) and FedRAMP CR26 (246 rules, no props) inventoried before the architecture was drawn. The kernel is what all three were measured to need.

July 2026

Gate 2: executable done

The normative JSON Schema with closed, mutually exclusive shapes, seven standard-library facet descriptors, the conformance corpus and the reference validator.

22 July 2026

Gate 3: the customer test done

SP 800-53 Rev 5.2.0 with the four 800-53B baselines, CSF 2.0 and a full lifecycle set of SSP, assessment plan, results and POA&M. No change to the kernel schema was needed. The corpus exposed two defects in the reference validator, which were fixed.

22 July 2026

Gate 4: the engines, and release 1.0.0 done

DSSE signature verification with a dependency-free Ed25519, bundle composition, the conditional rule, the PowerShell twin and the bidirectional OSCAL export. The specification went through adversarial review in the repository and an external pass; every finding was adjudicated on the record. One fix recovered 239 ISM taxonomy sets that a converter had silently merged.

10 Nov 2026

2.0.0-rc.1: Risk as the tenth type

All 6,914 objects of the 1.0.0 corpus validate unchanged under the new schema. No census authority publishes a risk record, because a risk register is the consumer's data. Three standards require it: ISO/IEC 27001, BSI Standard 200-3 and NIST SP 800-30.

12 Nov 2026

2.0.0-rc.2: Check as the eleventh type

Proposed on 11 November with schema, validator, 156 vectors and a census of two open check corpora: 481 of 510 rules of the CIS Ubuntu 24.04 profiles in ComplianceAsCode fit, and no comparison in either corpus needs a regex outside I-Regexp. Accepted the next day with six decisions. The promotion rule gained a clause for types that catalogue authorities do not publish; Risk and Check both enter under it. 205 vectors now.

12 Nov 2026

2.0.0-rc.3: advice packs, and two typed dates release candidate, not tagged

A vendor's recommendations as one published file, accepted with eight decisions on the record: no new type, eleven rules, 74 vectors and a worked pack with a signed attestation. Run over the shipped bundles, the rule on capability-ref found one edge in the converted NIST example that cited a capability its own component does not declare. On the same day the author closed the four points that were left before the tag. Two fields that 1.0.0 left as free text are now typed, the date a deviation was opened and the due date of an action.

13 Nov 2026

2.0.0-rc.4: who decided, and when release candidate, not tagged

The comparison with five neighbouring formats found two places where one of them was ahead, and the author decided both the same day. Who decided and when is now spelled one way, provenance, and is optional on a tailoring and each of its entries, on a risk and its ratings, on an implementation and on a finding. A deviation may say when it was decided and when it is due for review. A validator reports what is missing, one line per object and rule, and reports a review date that has passed against the date of the run. A missing author fails a bundle only when the caller asks for it; a review date that has passed never does. No new type and no new primitive: 40 vectors for the two reports, and nineteen deliberate defects in the reference each turn one red. Three times that accepted any string are typed.

04 Limits

What is not proven

These are the open points on 2026-11-15. Each of them is also a place where outside work counts for more than the maintainer's.

No third-party user open

A search on 10 November 2026 found no mention of JASCON outside its own repository. No authority publishes in it and no vendor reads it. The measured conversions and the defects found in published catalogues can be used without adopting the format.

One author behind both validators open

The Python reference and the PowerShell twin were written by the project author with AI assistance from the same sources. Their independence is limited to language and runtime. The gate measures whether the specification determines an implementation. It does not measure a stranger's weekend.

Path coverage is not value arrival open for 5 corpora

The coverage reports count source paths with a declared destination. Whether each value arrives is traced on 6 corpora, with a ledger that follows every source value through the converter: of 214,455 leaf values, 122,891 are carried (71,259 copied, 51,632 by a named rule), 91,564 are no content by a rule the ledger checks (a name that only selects a field, a namespace, an id that can be rebuilt exactly, a resource nothing refers to), 0 are dropped and 0 are unaccounted. A converter run with an unaccounted value fails. Every path map of these corpora had read 100 %, and the first ledgers found values that were lost all the same; they are carried now. A ledger holds a rule to what it wrote, not to whether the rule is right, and it does not trace back what a converter writes on its own. The workflow reproduces these conversions from their pinned source files on every change. The other 5 corpora are unmeasured.

CorpusSource valuesCopiedBy a named ruleNo content, by a checked ruleDroppedUnaccounted
AU.ISM36,1615,89615,66514,60000
DE.BSI49,52122,67911,82215,02000
FedRAMP-CR267,5325,5451,987000
US.CSF4,7261,9369051,88500
US.IFA-GoodRead835762621100
US.SP800-53115,68034,44121,19160,04800

The first Risk corpus stays in its first state open

OSCAL's risk data has no owner, no single level and no acceptance date. Risks converted from it cannot leave the state identified, so risk coverage on real converted data shows every requirement as uncovered. Coverage with a treated risk is shown on the example bundle only.

Check closes its shape on two platforms open

The fit of the Check type was counted on Linux hosts and Kubernetes. No check against a cloud API was counted, nothing in the repository runs a collector, and the worked example ran against a stub. 29 of 510 counted rules do not fit the declarative form and stay manual. The state in full.

2.0.0 is a candidate, not a release open

Nothing is left to decide, and both validators are green on all 822 vectors of the candidate. The two changes the author accepted on 13 November 2026 are in it as rc.4; the tools that write the new fields are still to follow. The tag is the author's step and has not been taken. Until then the released version is 1.0.0, and the shapes of 2.0.0 can still change.

Not every 1.0.0 bundle is a 2.0.0 bundle decided

Until rc.2 the 2.0.0 schema accepted everything 1.0.0 accepted. Since rc.3 the date a deviation was opened must be a date, and the due date of an action a typed point in time or duration. A 1.0.0 bundle with free text in either field has to convert it. Since rc.4 the same holds for the date of a mapping's provenance, the time of an assessment and the timestamp of an attestation. All five shipped bundles and the example folder conform. A pattern with a Unicode category follows the Unicode database of the machine it runs on, so a code point assigned later can give two verdicts on two machines; an author who needs one verdict writes the range.

Advice packs rest on one example file designed for

No catalogue authority publishes vendor advice. For such data the promotion rule asks for a standard that requires the record, or for two open implementations whose content was counted. Advice packs have neither. The clause-level field on a capability is backed by a single fictional component definition from NIST's examples, with three recommendations, two clause-level entries and two suggested parameter values; the product-version field is backed by nothing counted. Both entered the kernel on the author's decision against that rule, and the specification records it that way. No vendor has published a pack, the pack builder imports OSCAL component definitions, and NIST's own examples come through without loss of anything a pack can hold, but no component definition a vendor published has been run through it. That census is still owed.

Who decided is recorded, not enforced in rc.4, backlog 65 and 66

The comparison of 13 November 2026 found two places where a neighbour was ahead: a Deviation had no expiry, where a security exception in CISO Assistant has an expiration date and a risk in OSCAL a deadline; and no tailoring operation recorded its own author, where CISO Assistant logs each change. Both are in rc.4. The fields are optional: a validator reports a missing author, which fails a bundle only when the caller asks for it, and a review date that has passed, which never does. A Deviation past its review stays in force until someone acts. The author of a version is not a log of every change. Since 14 November 2026 the tools write the fields. The app's pages write the author its user has set, and converters and imports pre-fill author and date from the source, which whoever imports may overwrite. A source gives little: OSCAL names who is responsible for a document, not who wrote one of its statements, so one of seven records in the converted NIST example gets an author, and seven such report lines remain.

Governance is one person open

The handbook describes a federated registry and a steward for the standard library. Neither exists. Today there is one maintainer in personal capacity, a public decision register, and the conformance corpus.

The full queue with counts: the spec-feedback backlog. Items enter with evidence and leave through an entry in the decision register.

05 Prior art

What the neighbouring formats miss

Two passes. The first five rows were checked on 13 November 2026 against the schema or the code at a named version: OSCAL 1.2.3, Gemara 1.6.0, CycloneDX 1.7.2, CISO Assistant and XCCDF 1.2 with SCAP 1.4. The rows below them come from the survey of 10 November, read from public repositories and, where a primary site could not be reached, from secondary sources. No neighbour's data was converted, so a cell says what a format provides, not what a deployment does with it.

FormatWhat it coversWhat it misses
OSCAL (NIST)Catalog, profile, component definition, system security plan, assessment plan and results, POA&M, and since 1.2.0 control mapping. A complete chain from control to an evidence file with a hash; deadlines on risks; a vendor model that cannot modify a catalogA risk exists only in assessment results and POA&M, never beside the selection it would justify. No profile change records an author. No construct carries a second language. No native signing: inline signatures were declined upstream (issue 245), which leaves byte hashes. Extension props carry a namespace string that nothing validates, and core constraints written for SP 800-53 bind other frameworks (#2118, #2112)
Gemara (OpenSSF)Layered catalogues for guidance, controls, threats, risks and capabilities; mapping documents with typed relationships, strength and confidence; evaluation logs whose evidence carries a digest; schemas in CUE. Used by FINOS Common Cloud Controls and the OpenSSF OSPS BaselineNo prose in a second language. No due date or expiry on a finding or an exemption. No implementation record: the model names a layer for it and ships no schema. A risk is accepted without an approver or a date. Signing covers the published bundle's bytes and is done by tooling, outside the schema. The schemas are closed: a CUE profile can narrow them, not add a field. The policy and risk schemas are marked experimental
CycloneDX Attestations (OWASP, ECMA-424)Standards as data; requirements mapped to claims and counter-claims, evidence with author and expiry, an affirmation with signatories. Signatures per object over canonical JSON, which survive re-indentationNo risk object, no record of tailoring, no language map. Evidence attachments carry no hash. The signature exists in the JSON form only: the XML schema and the protobuf definition have no place for it. Neither signer nor affirmation carries a date. Extensions are name/value strings whose registry is optional
CISO Assistant (intuitem)In the tool: risk scenarios linked to controls, a Statement of Applicability export, risk acceptance with approver, date and expiry, security exceptions with an expiration date, an audit log. In the library format: more than 200 frameworks, reference controls, risk matrices, set-theory mappings, a translations map per objectRisk acceptance, exceptions and the Statement of Applicability stay in the application; they leave it through a backup export, not through the library format. No signing: byte hashes for deduplication. A library update can rewrite existing assessments. No check logic attached to a requirement. The importer ignores library keys it does not know
SCAP, XCCDF, OVALChecklists and check languages in XML. A Tailoring file changes a vendor's benchmark without editing it; an override records authority, time and reason; titles and descriptions repeat per language; a data stream can be signed and countersignedNo risk record. No due date, expiry or waiver type. No hash on a collected item. No author on a single tailoring change, and a severity can be lowered with no reason given. Re-indenting a signed stream breaks the signature. Conformant checks are OVAL and OCIL; a script in any language runs only through an OpenSCAP extension. Elements in a foreign namespace are checked only if a schema happens to be at hand
Regulator formatsFedRAMP's consolidated rules as JSON with a JSON Schema; BSI C5:2026 as YAML; Grundschutz++ and the ISM as OSCAL catalogues and profilesOne dialect per publisher, each structuring modality, deadlines and applicability its own way. C5 ships German and English as separate files; Grundschutz++ keeps its vocabularies in props that no validator reads
Mapping semanticsThe relationship styles of NIST IR 8477; the STRM of the Secure Controls Framework; OWASP OpenCREIR 8477 defines relationships, not a file format. SCF publishes its mappings as spreadsheets and PDF and exports only its catalogue to OSCAL. OpenCRE is a hub-and-spoke requirement graph, not a catalogue format
in-toto and SCITTSigned attestations with vetted predicates for provenance, verification summaries, test results and SBOMs; IETF transparency services over signed statementsNo in-toto predicate attests a control or a requirement. SCITT is an architecture for transparency, not a compliance model
Heimdall Data Format, OCSF, SARIFResults of checks and scansThey carry what a check found, not what a standard requires
NIST IR 8286Risk register and risk detail record schemasNo edge from a risk to the selection of controls
OpenControl, Compliance MasonryYAML standards and component narrativesNo meaningful activity since 2023
ISO/IEC 27001 and 27002, IEC 62443, PCI DSSThe standards themselvesNo official machine-readable edition; third-party renderings exist in SCF and CISO Assistant

What JASCON takes from them

Most of JASCON is not new. Each feature below has a precedent in a neighbouring format.

JASCON featureNearest prior art
Structured modality per statementThe modal-verb vocabulary of Grundschutz++; the keywords of the FedRAMP rules; the OpenSSF OSPS Baseline
Prose as a language mapThe translations map of the CISO Assistant library format; xml:lang on repeated titles and descriptions in XCCDF
Baselines as membershipOSCAL profiles; XCCDF profiles
Tailoring as a closed list of operationsOSCAL profile alter and set-parameter; XCCDF refine-value
Typed cross-framework mappingsNIST IR 8477; the STRM of the Secure Controls Framework; the OSCAL 1.2 mapping model; Gemara mapping documents
Typed, schema-bound extensionsTyped custom fields inside CISO Assistant; CUE profiles that narrow Gemara's closed schemas, as the OSPS Baseline does; OSCAL props with namespaces, untyped; XCCDF elements in foreign namespaces, validated laxly
Component to finding chainThe OSCAL lifecycle models; CycloneDX declarations
Signed attestation over requirementsCycloneDX affirmations and signatures over canonical JSON (ECMA-424); signed and countersigned SCAP data streams; the DSSE envelope of in-toto
A risk recordThe OSCAL risk assembly, after the fact; the Gemara risk catalogue; risk scenarios and risk acceptance inside CISO Assistant; NIST IR 8286
Vendor advice kept apart from the controlsOSCAL component definitions; the XCCDF Tailoring file, which leaves a vendor's benchmark untouched
A recorded exception with an approverThe XCCDF override with authority, time and reason; security exceptions in CISO Assistant; the FedRAMP deviation props on OSCAL

What no neighbour was found to combine

  • Two digests per object, one over the delivered bytes and one over the canonical content, so that an attestation survives repackaging.
  • A partial order of modality against which tailoring is checked, with weakening routed through a recorded Deviation.
  • A risk record joined to control selection, so that the requirements resting on no risk are computed.
  • Extensions that fail closed: an extension declares which computations it changes, and a tool that does not understand it must stop.

How far that claim goes

Each of the four has partial precedent. The claim concerns the combination and is only as strong as the survey. Gemara and CycloneDX Attestations are the closest neighbours, and a reader from either community will ask why JASCON does not extend theirs; the handbook answers in appendix F.10. The attestation gap is narrow: in-toto has no predicate that attests a requirement, and a JASCON attestation expressed as one would reach an existing verifier ecosystem. The ten features that matter to users and auditors are compared on the home page.

PRIOR-ART.md, with sources

Reproduce the verdict in one line

Clone the repository and run the validator. It executes the conformance corpus, then re-verifies every object of every shipped bundle with both digests. It needs Python and jsonschema.