Applicability Note

The invariants and predicates defined in this section SHALL be applied to all market survey executions from Version 2 onward.

They SHALL NOT be used to reinterpret, correct, or invalidate Version 1 results, which remain a valid historical snapshot executed under the definitions and evidence available at that time.

---

I. Capability Definition Invariants (Authoritative)

All capability meanings are fixed by: Cognitive Memoisation: LLM Systems Requirements for Knowledge Round Trip Engineering.

No platform documentation, analyst interpretation, or inferred behaviour SHALL override these definitions.

The following requirements are locked:

Semantic Artefact Ingress
Durable Artefact Egress
Context Determinism
Transport Independence
Governance Compatibility
On-Site / Embedded Suitability (deployment classification only)

Symbol semantics are fixed as:

✓ (satisfies requirement)
✗ (fails requirement)
? (insufficient evidence to determine)

Successful CM round-trip (emit → capture → re-ingest → assert → govern) SHALL be treated as decisive evidence for:

ingress,
egress,
determinism,
and transport independence.

---

II. Survey Execution Invariants

Each platform surface (Hosted UI, API, Self-hosted) is evaluated independently.
No behaviour observed on one platform or surface SHALL be projected onto another.
No speculative downgrades are permitted.
Absence of evidence SHALL be recorded as '?'.
Later surveys record deltas; they do not invalidate earlier versions.
Training, fine-tuning, or vendor “memory” features are non-authoritative.

---

III. Evidence Acceptance Rules

Permitted evidence:

Direct CM artefact testing (MWDUMP, TMLDUMP, XDUMP)
Author-observed behaviour
Publicly available platform documentation

Excluded evidence:

Marketing material
Roadmaps or intent statements
Community anecdotes
Unverifiable benchmarks

---

IV. Search Predicate Invariants

The following search predicates SHALL be reused unchanged in future survey executions unless explicitly amended by the curator.

They are expressed parametrically using {PLATFORM} and {SURFACE}.

Core Capability Predicates

"{PLATFORM} file upload supported formats limits"
"{PLATFORM} semantic document ingestion"
"{PLATFORM} read uploaded file context behaviour"
"{PLATFORM} deterministic prompt reproducibility"
"{PLATFORM} context reset new session behaviour"
"{PLATFORM} system prompt governance control"

Artefact Egress Predicates

"{PLATFORM} export generated file"
"{PLATFORM} save model output as file"
"{PLATFORM} return generated document"
"{PLATFORM} download assistant output"
"{PLATFORM} file output api"

Transport and Round-Trip Predicates

"{PLATFORM} re-upload generated file"
"{PLATFORM} reuse uploaded document new session"
"{PLATFORM} file upload reassert context"
"{PLATFORM} multi-session document persistence"

Surface Differentiation Predicates

"{PLATFORM} hosted ui vs api differences"
"{PLATFORM} api file handling"
"{PLATFORM} batch inference file input"
"{PLATFORM} cli or sdk document ingestion"

Deployment / Embedded Predicates

"{PLATFORM} self hosted deployment"
"{PLATFORM} on premise llm deployment"
"{PLATFORM} open weights license"
"{PLATFORM} edge or embedded inference"
"{PLATFORM} hpc inference deployment"

---

V. Embedded Suitability Clarification (V2+)

On-Site / Embedded Suitability SHALL be evaluated as architectural feasibility of curator-controlled deployment.

It SHALL NOT be interpreted as:

a measure of CM correctness,
a requirement for SaaS platforms,
or a downgrade of CM-capable hosted systems.

A ChatGPT-like system deployed on curator-controlled infrastructure with CM governance SHALL be marked embedded, even if the original SaaS offering is not redeployable.

---

VI. Change Control

Any modification to:

capability definitions,
symbol semantics,
search predicates,
or execution rules

MUST be recorded as a new curator amendment and versioned separately.