inference/research
verifiedhigh · 0.95minorupdated 2026-10-10

Typed fields cannot distinguish legal vs commercial vs disclosure vendor silence, nor definitional vs sourcing vs verified-negative null bandwidth without schema discriminators

[minor] measurement

Fields

id
gotchas/w11-schema-discriminators-for-vendor-silence-and-null-bandwidth
Class
measurement
Severity
minor
Concepts
schemadiscriminatorvendor-silencenull-bandwidthprovenanceverified-negativedefinitional-nullsourcing-nullexport-controlpaywallaggregator-onlyno-config-json
Symptom
Two independent audits discovered identical compression traps across disparate layers: in data/sources/, four distinct vendor silences (legal export control, commercial paywall, vendor disclosure practice with no spec table, and packaging lack of config.json) all collapsed into an ambiguous 0.3-0.5 confidence score; in data/interconnect/, three distinct null bandwidth reasons (definitional transport without intrinsic link rate, sourcing aggregate without per-link divisor, and verified-negative confirmed absent spec) all collapsed into bandwidth_gbps: null. In both cases, downstream consumers and query tools were forced to parse load-bearing free-text notes or basis fields to determine why a value was missing or low-confidence, and could not query or filter records cleanly by mechanism.
Root cause
The schemas lacked optional discriminator enums to make explicit the distinct real-world mechanisms documented by gotchas w10-vendors-who-publish-nothing-export-control-paywall-and-aggregator-only and w7f-non-nv-am-vendor-conventions. Without source.vendor_silence ('export-control', 'paywall', 'aggregator-only', 'no-config.json') and interconnect.null_bandwidth ('definitional', 'sourcing', 'verified-negative'), every unclosable negative result or unmeasurable transport looked identical to an unresearched draft or an unfinished audit.
Workaround
Use the schema discriminator fields instead of free-text parsing. Future data waves can backfill source.vendor_silence on low-confidence sources: 'export-control' on nv-h20-flopper-specsheet (permanent legal restriction), 'paywall' on nrg2-uptime-institute-survey-paywalled (commercial member-only document), 'aggregator-only' on cn-cambricon-mlu590 and cn-iluvatar-bi-v150 (vendors publish no spec table), and 'no-config.json' on hf-embed-voyage-code-3 (packaging artifact on Hugging Face). Similarly, backfill interconnect.null_bandwidth on null-bandwidth fabrics: 'definitional' on roce, ethernet-uec, and cerebras-wafer-mesh; 'sourcing' on biren-blink, groq-lpu-interconnect, and tenstorrent-warp100; and 'verified-negative' on iluvatar-bi-v150-and-bi-v250-fabric.
notes
Added in wave 11 schema enhancement to bridge gotchas w10 and w7f. Per repository rules, schema discriminators are declared first without mass-backfilling data/ records. Existing records remain valid and untouched with null discriminators until dedicated data migration waves populate them.

Notes

Added in wave 11 schema enhancement to bridge gotchas w10 and w7f. Per repository rules, schema discriminators are declared first without mass-backfilling data/ records. Existing records remain valid and untouched with null discriminators until dedicated data migration waves populate them.

Cross-references