Context & efficiencyToken-optimization techniques

18 — Coverage and gap map

Coverage map for quota, multimodal input, latency, fleets, portability, literature, governance, and online quality.

Summary

The core techniques chapters cover text-token economics, caching, retrieval, output discipline, routing, provider features, infrastructure, composition, and validation. This map identifies the decision axes that need separate treatment: subscription quota, multimodal input, latency and human time, fleet behavior, portability, current literature, and runtime governance.

The strongest correction is conceptual: a subscriber constrained by a usage cap should optimize tasks per cap, not API dollars per task. Multimodal input is another distinct token class rather than a footnote to text context. Latency and quality also need explicit accounting because a cheaper token path can still be slower or less reliable.

Question and scope

Which independent cost and quality axes are already covered, and where do material evidence gaps remain?

Method

Coverage was assessed against six independent axes rather than the dossier table of contents:

AxisDecision dimension
Cost metricdollars, subscription quota, wall-clock time, human attention
Token classfresh input, cache write, cache read, visible output, thinking, image, document
Emission surfacesystem prefix, tools, message history, model output, media tool results
Leverstyle, tokenizer, context architecture, caching, retrieval, routing, infrastructure, governance
Scopeturn, session, cross-session, capsule, fleet, organization
Deliveryinstruction, configuration, hook, orchestrator, provider feature

The existing chapters were checked for absent, incidental, partial, or decision-complete treatment. A topic remains a gap when the corpus mentions it but does not supply the measurement or decision rule required to choose an action.

Findings

Current coverage map

TopicCurrent assessmentDecision stakeDetailed chapter
Subscription and quota economicsDollar economics are mature; cap accounting remains partly provider-opaqueCache-heavy levers can rank differently under quota than under API prices19 — Subscription and quota economics
Multimodal inputSeparate measurement required for images, screenshots, and PDFsA screenshot can be cheaper or much more expensive than structured text20 — Multimodal token economics
Latency and human timeToken and dollar models do not decide the time tradeoffFan-out, fast modes, and compression hops can exchange tokens for elapsed time21 — Latency and time economics
Fleet and multi-tenant cachingSelf-host mechanisms are covered; hosted cache scope is not fully observableSpawn timing, dynamic prefixes, and quota attribution change fleet efficiency22 — Fleet and multi-tenant cache
Cross-agent portabilityIndividual mechanisms are documented, but support differs by runtimeA stack tied to one hook or provider is fragile23 — Portability matrix
Literature and product surfaceProvider features and research continue to changeNew mechanisms can strengthen, qualify, or leave existing decisions unchanged24 — Literature and market scan
Governance and online qualityOffline validation exists; runtime budget and drift controls are thinnerSavings are invalid if measurement overhead or quality regressions erase them25 — Governance and online quality
Unresolved mechanismsSeveral ideas remain conditional or provider-blockedFrontier ideas need feasibility gates before entering a stack26 — Frontier mechanisms
Combined verdictLayer interactions and metric choice need one synthesisComposition must count overlap, cache effects, recovery, and quality27 — Stack verdict
Volatile claimsProvider-sensitive claims require an explicit verification contractA current decision must expose its cutoff and uncertainty28 — Provider-sensitive claim verification

Headline findings

  • Quota is a separate objective. Provider caps are not transparently convertible to API dollars, so quota-sensitive conclusions are explicitly bounded where the denominator is unpublished.
  • Media must be measured in its native representation. Image and PDF cost depends on dimensions, preprocessing, page handling, and provider rules; text-token intuition is insufficient.
  • Latency is not a token subclass. It belongs beside token and quality metrics in the acceptance function.
  • Fleet behavior is cache behavior plus orchestration. Prefix identity, dynamic sections, start timing, and provider cache scope determine whether nominally identical agents share work.
  • Portability follows the interception boundary. Instruction-only techniques travel widely; provider betas, proprietary hooks, and runtime-specific storage do not.
  • Governance is part of net savings. Measurement, recovery, retries, and quality monitoring consume resources and must be charged to the technique.

Implications for jackin

For jackin, the coverage and gap map evidence identifies which mechanisms are safe to adopt directly and which still require workload-specific validation.

Limitations and unknowns

Evidence boundaries

The chapters use the measurement conventions in 01 — Economics and measurement and the acceptance harness in 16 — Validation harness. Provider-sensitive facts follow the verification contract. Where providers do not publish quota weights, cache scope, or internal token accounting, the result stays bounded rather than inferred.

The dossier's measured heavy-session profile remains a comparison fixture, not a universal workload. Reach must be remeasured for each capsule because a docs-heavy task dominated by native reads presents different optimization opportunities from a shell-heavy code task.

Sources

Evidence and repository references are cited inline beside the claims they support.

On this page