UK-based independent AI efficiency projectEvidence before claims · 2026

AI Workflow Glossary

AI workflow glossary · terms with context

Plain-English and technical definitions with examples and evidence notes. The KD04 pack marks 11 items as standalone glossary references; the remaining 89 definitions are grouped with their parent guide on the Knowledgebase and do not get separate term URLs.

Term index · 100 terms

  1. Token
  2. Tokenizer
  3. Token budget
  4. Input token
  5. Output token
  6. Cached input token
  7. Reasoning token
  8. Context window
  9. Context utilization
  10. Context overflow
  11. Context compression
  12. Retrieval-augmented generation
  13. Prompt contract
  14. System instruction
  15. Prompt compiler
  16. Binding requirement
  17. Acceptance criterion
  18. Structural retention
  19. Semantic equivalence
  20. Model routing
  21. Fallback route
  22. Escalation threshold
  23. Inference
  24. Prefill
  25. Decode
  26. Quantization
  27. Model digest
  28. Model card
  29. KV cache
  30. Local inference
  31. Cloud inference
  32. API billing
  33. Subscription plan
  34. Cost per accepted result
  35. Total cost of ownership
  36. Useful life
  37. Residual value
  38. Idle power
  39. Incremental power
  40. Energy per inference
  41. Retry
  42. Retry budget
  43. Idempotency
  44. Idempotency key
  45. Checkpoint
  46. Checkpoint integrity
  47. Resume operation
  48. Rollback
  49. WAIT_RESOURCE
  50. WAIT_AUTH
  51. Human active time
  52. Elapsed time
  53. Blocked time
  54. Opportunity cost
  55. Observability
  56. Trace
  57. Attempt ledger
  58. Failure denominator
  59. Independent verification
  60. Artifact hash
  61. Exact-byte verification
  62. Source ledger
  63. Claim ledger
  64. Canonical URL
  65. Canonical content identity
  66. Keyword cannibalisation
  67. Search intent
  68. Internal link graph
  69. Pillar page
  70. Supporting page
  71. Knowledgebase entry
  72. How-to guide
  73. Thin content
  74. Model licence
  75. Data boundary
  76. Least privilege
  77. Rate limit
  78. Timeout
  79. Backoff
  80. Circuit breaker
  81. WordPress REST API
  82. Readback verification
  83. Draft import
  84. Duplicate preflight
  85. PRO100
  86. Plan economics
  87. Parallelism
  88. Concurrency
  89. Fan-out/fan-in
  90. Resource contention
  91. Human-in-the-loop
  92. Verified useful output
  93. Raw throughput
  94. Acceptance rate
  95. Evidence provenance
  96. Measured value
  97. Derived value
  98. Assumption
  99. Unknown value
  100. Matched benchmark

Standalone glossary references · 11

Token

Plain English: A small piece of text a language model reads or produces; it may be a whole word, part of a word, punctuation, or another text unit.

Technical: A tokenizer maps a text sequence to model-specific token IDs; token boundaries vary by tokenizer.

Example: The same sentence can have different token counts with different model tokenizers.

DEFINITIONAL; platform-specific meanings require current primary documentation. · Parent guide key: HT-001

Tokenizer

Plain English: The software/rules that split text into units a model can process.

Technical: A mapping from input text to a sequence of token IDs for a particular model family/encoding.

Example: Use the target model’s tokenizer rather than converting words with a fixed ratio.

DEFINITIONAL; platform-specific meanings require current primary documentation. · Parent guide key: HT-001

Context window

Plain English: The information a model can consider in one request or active conversation, subject to product/runtime rules.

Technical: A maximum context capacity over model-defined tokenized input/output; actual usable capacity may be lower.

Example: More history can crowd out task instructions and current evidence.

DEFINITIONAL; platform-specific meanings require current primary documentation. · Parent guide key: HT-005

Model routing

Plain English: Choosing which model handles a task based on its needs.

Technical: Policy mapping task features/state to model/provider route and fallback transitions.

Example: Use a smaller route for a task only after it passes the same checks.

DEFINITIONAL; platform-specific meanings require current primary documentation. · Parent guide key: HT-015

Quantization

Plain English: A way to store model numbers using fewer bits, often reducing memory needs with possible quality/performance tradeoffs.

Technical: Numerical representation transformation (e.g. lower precision) applied to weights/activations with implementation-dependent effects.

Example: A quantized local model is not automatically quality-equivalent to its unquantized source.

DEFINITIONAL; platform-specific meanings require current primary documentation. · Parent guide key: HT-021

API billing

Plain English: Charges for using a software service through requests, often with usage-based categories.

Technical: Provider-defined accounting and rate application by model, token type, tools, caching, region, or other dimensions.

Example: An API rate card may price input and output differently.

DEFINITIONAL; platform-specific meanings require current primary documentation. · Parent guide key: HT-010

Idempotency

Plain English: A repeated request produces the same final effect as making it once.

Technical: Operation property that permits safe replay under a stable key and controlled state.

Example: Retrying a key should return the original payment receipt, not charge again.

DEFINITIONAL; platform-specific meanings require current primary documentation. · Parent guide key: HT-031

Checkpoint

Plain English: A saved snapshot of enough job state to continue after interruption.

Technical: Versioned durable state with integrity and consistency guarantees at a safe boundary.

Example: Save completed step IDs and artifact references before continuing a long workflow.

DEFINITIONAL; platform-specific meanings require current primary documentation. · Parent guide key: HT-033

WAIT_RESOURCE

Plain English: A workflow state meaning progress is blocked until a named resource becomes available.

Technical: Typed resumable state with resource, reason, entry time, deadline, retry/escalation policy, and checkpoint.

Example: Record which queue/resource the job is waiting for; do not silently sleep forever.

DEFINITIONAL; platform-specific meanings require current primary documentation. · Parent guide key: HT-039

WAIT_AUTH

Plain English: A workflow state meaning valid permission or user authorization is needed before continuing.

Technical: Typed resumable state for expired/invalid/absent credentials or consent; secret material must remain protected.

Example: Pause a provider call and ask the user to reauthorize through the official flow.

DEFINITIONAL; platform-specific meanings require current primary documentation. · Parent guide key: HT-043

PRO100

Plain English: A plan label used here for the ChatGPT Pro 100 scenario; product meaning and limits must be verified from current official terms.

Technical: User-specified plan label; exact entitlement, usage allowance, model access and period remain CURRENT_EXTERNAL verification requirements.

Example: Do not infer API token pricing or throughput from the name alone.

DEFINITIONAL; platform-specific meanings require current primary documentation. · Parent guide key: HT-012

Inline definitions in parent guides · 89

Open a definition at its parent guide. These definitions are not promoted to separate canonical pages.

Go to the Knowledgebase how-to library →

See also the Research Questions, Infographics, and tools.