UK-based independent AI efficiency projectEvidence before claims · 2026

Token Optimisation

AI efficiency · measured
Discuss a diagnostic

AI Workflow Infographics

Visual versions · evidence labels stay with the graphic

Accessible, responsive diagrams for the measurements and reliability questions Token Optimisation is investigating. Each visual has a version ID, evidence boundary and link to its source article or question.

A short response can follow a larger AI requestOne bounded GPT-6 Luna auth canary reported 13,289 input tokens, zero cached input and ten output tokens; this does not attribute input to individual components.ONE BOUNDED CANARY · OBSERVEDProvider-reported input13,289tokenscached input: 0 · one run, not a benchmarkVisible output10tokensexact reply: KD_OPENCODE_AUTH_OKTotal usage does not identify each wrapper/tool/history contribution.
Request boundary: measure provider-reported input, cached input and output—not only the visible reply. Read the evidence and limits.Visual ID KD10-VIS-REQUEST-BOUNDARY-01 · v1.0 · OBSERVED single run
Four signals needed to understand AI workflow progressProcess liveness, provider activity, durable stage progress and verifier acceptance answer different questions and should be recorded separately.PROGRESS IS NOT ONE SIGNAL1 · PROCESSIs it alive?2 · PROVIDERActivity/status?3 · STAGEDurable progress?4 · VERIFYAccepted?A heartbeat answers liveness; it does not reset a semantic-progress deadline.
Progress signals: separate liveness, provider activity, stage artifact and verifier result. Read the incident observation.Visual ID KD10-VIS-PROGRESS-SIGNALS-01 · v1.0 · scoped observation

12 concurrent Luna Low sessions · one-machine observation

Owner-provided ranges: CPU ~39–71%, RAM ~25.6–26.0 of 31.8 GB, free ~5.8–6.1 GB, GPU ~10%, disk negligible. This is OBSERVED ON ONE MACHINE / ONE WORKLOAD, not a concurrency ceiling.

Read the research question, hardware context and full limits →Visual ID KD10-VIS-CONCURRENCY-12-01 · v1.0 · linked to the owner-provided observation

21 article infographics

Each visual pairs with one live article: view article ↔ view infographic. Observed, calculated, synthetic and unknown evidence stays labelled at the visual boundary.

  1. 12 Luna Low Sessions: One-Machine Resource SnapshotArticle companion · evidence scope retained
  2. Acceptance: A Heartbeat Is Not a Verified ResultArticle companion · evidence scope retained
  3. API Cost Scenario: Keep Inputs, Rates, FX Date and Limits VisibleArticle companion · evidence scope retained
  4. Auth Failures: The Event Sequence Does Not Prove CauseArticle companion · evidence scope retained
  5. Cached Input: Preserve the Provider’s Separate Usage FieldArticle companion · evidence scope retained
  6. Cheaper-Model-First Routing Needs a Quality GateArticle companion · evidence scope retained
  7. Context Budget: Allocate for the Task, Do Not Fill by DefaultArticle companion · evidence scope retained
  8. Cost per Verified Job: The Denominator Comes FirstArticle companion · evidence scope retained
  9. Heartbeat vs Progress: Four Different Workflow SignalsArticle companion · evidence scope retained
  10. Human Time vs Machine Time: Keep the Clocks SeparateArticle companion · evidence scope retained
  11. LLM or Deterministic Tool? Route by Uncertainty and AcceptanceArticle companion · evidence scope retained
  12. Local AI vs Cloud AI: Compare the Same Accepted WorkArticle companion · evidence scope retained
  13. Measure at the Final Request BoundaryArticle companion · evidence scope retained
  14. Provider-Free Resume Test: What the Harness Can ShowArticle companion · evidence scope retained
  15. Request Boundary: Why a Short Reply Can Follow a Large InputArticle companion · evidence scope retained
  16. Shared AI Operating Rules: Consistency Without Savings ClaimsArticle companion · evidence scope retained
  17. Stall Detection: Check, Reconcile, Then DecideArticle companion · evidence scope retained
  18. Token Optimisation: Measure Useful Verified WorkArticle companion · evidence scope retained
  19. WAIT_AUTH vs WAIT_RESOURCE: Two Different Recovery PathsArticle companion · evidence scope retained
  20. WAIT_RESOURCE: Recovery Outside the Model TurnArticle companion · evidence scope retained
  21. Workflow Tools: Question → Inputs → Transparent OutputArticle companion · evidence scope retained