Library / Knowledge-Corpus Access Engineering Principles Framework
Jump to passage
In this reading

Link to current text

Published source confirmed at last check

Source changed 2026-10-03 14:36:52 UTC · snapshot created 2026-10-03 14:38:14 UTC · last check 2026-10-03 14:50:08 UTC

KCAE.DELIVER:4 - Solution

KCAE.DELIVER:4.1 - Specify the receiving contribution and reader

Identify what the recipient will do: choose a source, understand an explanation, perform a method, check a conclusion or maintain the access service. State the preparation already available and the operations the runtime actually supports. For a human, available time, domain competence and access permissions matter. For a model, include current instructions, model capability, tool access, residual window and output limits. Total historical tokens consumed and free space in the active context are different resources.

A search packet can contain a candidate’s identity, edition, brief reason for relevance and source address. An application packet usually needs the method’s recognition conditions, required inputs, consequential exceptions, operating instructions, stop conditions and any joining operation not already known to the reader. This is a sufficiency question, not a fixed field count. Derive the necessary material from the requested use and inspected source, not a universal 600-token allowance.

KCAE.DELIVER:4.2 - Recover the required reading closure

Start with the selected passage and follow dependencies needed to interpret or apply it: definitions, qualifications, tables, footnotes, named prerequisites and relevant adjacent exceptions. Expand across documents when a release note or external criterion changes the meaning. Stop expansion when the intended reader can obtain the receiving result without guessing an essential relation; unrelated background need not travel.

Keep original passages distinguishable from the engineer’s explanation. A generated contextual prefix is discovery support, not a quote from the author. A translated passage is a translation with a stated role, not automatically the governing edition. If a missing source dependency prevents adequate delivery, return that specific limitation. Do not manufacture a complete-looking method from fragments.

For difficult or consequential use, test a question that depends on the vulnerable distinction. Can the recipient explain why protocol v1 selects manual reconciliation while v2 may permit a resend? Recognition of a document title would not test that contribution. KCAE.EVAL supplies a bounded reader-use comparison when the design depends on this capability.

KCAE.DELIVER:4.3 - Read complete bounded units through a pinned sequence

When source reading is paginated, the reader needs an edition or snapshot identifier, a bounded unit, a continuation token or explicit next position, and a completion indication. Pin the edition for the sequence. Record which portions were actually returned, and detect a gap, duplicate or changed generation before combining them. A source hash can confirm the obtained bytes; it cannot prove that a truncated tool message conveyed those bytes to the reader.

Use the actual output limits. Request a smaller portion when a response is truncated, verify the continuation boundary, and continue until the needed unit is complete. Do not take “command succeeded” as an end-of-document signal. A tool may store a complete file while showing only its beginning. Its reader interface must distinguish the saved result from the shown portion.

A method can span more than one call, provided the recipient can retain or revisit the dependencies it needs. Very long sources may require an external notebook of source-addressed intermediate results and selective rereading. That notebook is itself a lossy representation. Preserve the exact return path and reread the relevant original when its omitted distinction becomes live.

KCAE.DELIVER:4.4 - Choose a delivery arrangement the runtime can perform

Direct complete reading is suitable for a manageable dossier and a capable recipient. Retrieval followed by selective reading spends context on material selected for the current use. An isolated reader can inspect a larger or private portion and return a bounded result, provided the platform actually creates an independent context and the result retains enough grounds for its consumer. Controlled history reconstruction can retain current instructions and selected state while removing obsolete search output, but only when the runtime exposes that operation.

An instruction such as “forget the preceding search” does not establish deletion or reclaim a context window. A skill or tool that returns a concise message does not establish an isolated reader. Verify these runtime properties through documentation and a relevant observable test. Where isolation is unavailable, keep search material outside the main dialogue as far as the interface allows, use smaller returns and plan the necessary reading before filling the window. Where a fresh context is used, explicitly supply its governing instructions, source permissions, question and necessary prior results; independence is not inherited knowledge.

Separate conversation contexts can still share a filesystem, credentials, tools or network access. Check those actual boundaries before assigning a private dossier to another reader. Context separation alone is neither a permission grant nor a security sandbox. Supply only the authorized material and capabilities needed for the bounded contribution.

A cached long-context source can be a credible alternative when the platform supports it and the workload amortizes cost. Cache storage does not guarantee reliable attention to every needed item. Compare its actual delivery performance and current charging semantics with selective reading. A cache holding an old edition remains an old edition.

KCAE.DELIVER:4.5 - Budget discovery, application and explanation together

Reserve room and time for the selected source, necessary connections, answer and uncertainty before spending the entire allowance on candidate discovery. Include the cost of parallel calls and external readers in total work, even if they save main-context space. A cheaper main dialogue can conceal a much more expensive service.

Use a runtime observation of remaining capacity where available. When it is unavailable, label the estimate and leave a margin for interface overhead and output. Cumulative billed tokens across calls do not reveal the active window’s remaining space; cached, repeated and external-reader input can be counted differently. Do not present an inferred residual as a measured runtime fact.

For example, if the active window has room for 12,000 further tokens, and source closure plus explanation is estimated at 7,000, a 9,000-token candidate dump is infeasible even if the model has consumed only a small fraction of its account quota. Either shrink discovery output, use a genuine external reader, obtain more capacity or return a narrower justified result. These invented numbers illustrate reservation, not a recommended product limit.

Check data boundaries on every transfer. Public retrieval can use a sanitized question; local case facts can remain with a local assessor. A remote provider’s retention and logging behavior is part of the chosen interface, not implied by calling the input “context.” Redact with attention to meaning: if removing the decisive condition makes assessment impossible, obtain an approved local route or a permission decision rather than guessing.

KCAE.DELIVER:4.6 - Confirm receipt and expose what remains to be done

Return the contribution at the level the recipient needs: conclusion, grounds, conditions and next operation. For execution, distinguish a recommended action from an action already performed. For later use, include sufficient edition and dependency information to detect staleness. For human adoption, a readable explanation or small first useful exercise can reduce the burden; dumping every engineering detail may make correct material practically unusable.

Confirm actual receipt where the use requires it. Delivery to an API, display in a user interface and human comprehension are different observations. Unknown receipt remains unknown. If the recipient cannot recover a needed distinction, repair the representation, reader preparation or access support indicated by the failure, rather than adding unrelated text.