Docs
31 synthesis documents written from the records, citing record ids inline. Every quantitative claim in a doc traces to a record you can open from the doc itself. docs/00-index.md is generated bytools/index.py and is not rendered here.
Decision docs
Hardware selection by workload01-hardware-selection · 36,576 charsFLOP map: which operations dominate, and what that means for hardware02-flop-map · 44,010 charsEngine selection: which engine for which hardware03-engine-selection · 33,678 charsQuantization: weight-only-4bit vs W4A4 are different regimes04-quant-selection · 34,748 charsKnown traps, grouped by severity05-known-traps · 48,600 charsGlossary06-glossary · 14,191 charsBenchmarking methodology and its pitfalls07-benchmarking-methodology · 16,811 charsCost per output token08-cost-per-token · 29,156 charsModel memory envelope: weights, KV, and the weight-vs-KV streaming crossover09-model-memory-envelope · 51,973 charsTesting methodology: differential oracles that actually catch silent corruption10-testing-methodology · 18,886 charsInterconnect decision table: when is tensor parallel worth it over a given fabric?11-interconnect-decision · 16,242 charsServing workflows: where the time actually goes, and why it moves12-serving-workflows · 23,839 charsModel × engine compatibility: which architecture needs which kernel, and which engine has it13-model-compatibility · 54,381 charsAudit findings14-audit-findings · 38,109 charsCompilers and kernel DSLs: what actually generates the kernel15-compilers-and-kernel-dsls · 9,860 charsMigration and upgrade: when a new GPU generation is worth it16-migration-upgrade · 9,702 charsDeployment shape: single GPU, multi-GPU, or multi-node17-deployment-shape · 10,213 charsReliability: what actually goes wrong in production serving18-reliability · 16,428 charsProduction operations: what to run, what to alert on, what is unsolved19-production-operations · 21,844 charsObservability: which metrics to alert on, and which ones lie20-observability · 16,063 charsDeployment topology: control plane, data plane, and what a cold start costs21-deployment-topology · 16,609 charsSafety and security, from the serving system's side22-safety-and-security · 24,381 charsChoosing a quantization format for a given hardware target23-quantization-selection · 44,071 charsChoosing a metrics surface: what each engine can actually measure24-metrics-surface-selection · 21,917 charsCitation integrity: when a source does not contain the number it is cited for25-citation-integrity · 16,632 charsKnown defects: the scope-boundary ledger26-known-defects · 28,788 charsAffordable 35B-class fit matrix27-affordable-35b-fit-matrix · 259,311 charsPaper reading map: the 282 paper records by topic and adoption27-paper-reading-map · 43,487 charsValue: rent vs own, $/Mtok, joules per token, and what the corpus can actually rank28-value-rent-vs-own · 6,516 charsWorking notesNOTES · 1,009 charsReading guideREADME · 3,274 chars