# BHI Research Note RN-002 — Why Cloud Core Refused to Produce a Composite Score **Series:** BHI Research Notes — short, dated, reproducible research statements. Not marketing. **Status:** PUBLISHED v1.0.1 — revised 02.09.2026 (opening sentence aligned with the provenance column; dated revision note at the foot of the note; v1.0 published 26.08.2026 (drafted 22.08.2026; publication delayed two days past the 24–25.08 cadence, caught by the 26.08 plan review). Permanent URL: https://blackholeindex.com/research/rn-002-cloud-core-identification-boundary **Author:** Ivan Savich (ORCID 0009-0003-9542-7022) · independent researcher **Dataset/version:** Cloud Core frozen snapshot S1, SHA-256 `6565778b84b38a9420c41fec37219303548cf6b7fb52f88237c6677b949712fe` (publicly fixed 22.08.2026, blackholeindex.com/research-snapshots.txt; two independent RFC 3161 timestamps of the digest). --- ## The result A rigorous application of the Black Hole Index to cloud infrastructure — AWS, Azure, GCP, Oracle Cloud, one provider-neutral frame, one stipulated Reference Customer — reaches an identification boundary before reaching a composite score. Eight of eleven structural parameters carried determinations for all four providers — each published with a provenance tag stating how much of the score the measurement frame itself determined (three parameters are provider-observed, four combine a frame baseline with provider-documented movement, one is stipulated by the frame). Three were not scorable from provider-side evidence at all — and the refusal to compute B from the remaining eight is the finding, not a gap in it. ## Why refusal, mechanically The BHI protocol's 11/11 rule forbids computing, displaying or publishing any B until all eleven parameters score. This is not caution for its own sake: the extended parameters multiply capture, so omitting one is arithmetically an assertion of minimum capture, not an abstention. When the rule was adopted, re-computation against the published release showed the omission it forbids would have moved the zone classification of 70 of 184 platforms (38.0%), with a worst-case distortion of ×1.86. The scoring pipeline enforces the rule: the scorer first reproduces all 184 published release scores from the pinned dataset, then refuses computation while any determination is Not Scorable — the refusal output, with its reasons, is part of the published result. ## What could not be scored, and why that is informative - **Momentum (`t`)** requires velocity of capability change relative to competitors — a common unit over time. First-party release channels exist symmetrically, but their units are editorial: one vendor bundles dozens of changes per dated entry, another lists one feature per entry, a third mixes both, the fourth's page describes itself as covering only "some of the major updates". Counting entries would measure documentation policy, not platforms. The missing common unit is itself a standardisation gap. - **Closeness (`c`) and Human Fallback (`h`)** are properties of the customer's state — contact patterns, retained capability — which no provider documents. Deriving them from the measurement frame fails the deductive bar that legitimately settled one other parameter: near-certain is not deductive. The deepest regulator record in the domain measures switching at market-average specificity — evidence that the constructs vary, not a measurement of a stipulated customer. Each Not Scorable determination is recorded with the searches that failed to close it, so a reviewer can attack the refusal itself. ## The design point An instrument that produces a number wherever a number is wanted measures its audience, not its object. The valuable output here is the decomposition: which dimensions of structural dependence are identifiable from provider evidence (eight, published with provenance tags stating how much of each score the measurement frame itself determined), which require customer-state or longitudinal evidence (three, with closure paths registered as research issues), and what a regulator can therefore expect provider-side evidence to establish. A composite B would have compressed exactly the information a policy reader needs. ## Limitations Single evaluator; inter-rater reliability not established; declared anchoring exposure (the evaluator had seen earlier published release figures). Scores are statements about first-party documentation read in a declared window against a stipulated Reference Customer. The refusal is relative to this boundary and evidence class — a customer-state study (registered as a research issue) could close `c`/`h`; a defensible capability-transition unit could close `t`. ## Reproducibility Snapshot S1 (hash above) contains the complete corpus: boundary, evidence rules, comparability matrix, four ledgers, construct audit (32/32 rows), comparative profile, prediction register, research issues, enforcement scripts. The hash was publicly fixed on the freeze date and independently timestamped twice (RFC 3161, FreeTSA + DigiCert, 22.08.2026). The enforcement scripts re-verify the corpus structurally on every audit run. --- **Revision v1.0.1 (2 September 2026).** The opening sentence previously read "were determined from first-party provider documentation, for all four providers"; that compressed the provenance column — one of the eight is frame-stipulated and four combine a frame baseline with provider-documented movement, as this note's own "design point" paragraph already stated. Wording aligned with the provenance column. No value, determination or conclusion changes.