The Runbook Shelf

An event-relay operator asks one question: How do I recover incident ERR-0713? The answer lives in a runbook shelf with these fixed properties:

• 800 Markdown runbooks, one incident code per file;
• one recovery procedure per runbook;
• distinctive incident codes shared by queries and headings; and
• a declared 200,000-token window budget.

demo.py builds the shelf deterministically in a temporary directory. Its token denomination is stated in code: one whitespace-delimited fixture word is one token. This is not a claim about a production tokenizer. It is a controlled unit that makes the before→after attributable to one admission decision.

The runbook is available, but the corpus cannot become resident whole. The checked fixture contains 1,293,600 tokens across 800 documents (demo.py, run 2026-08-07), more than six times the declared window budget. Frontloading therefore fails before the operator's question is answered.

The obstacle is corpus size, not uncertain need. Deferring the same fixed 1.29-million-token payload until ERR-0713 appears only moves the impossible load later. The fixed-payload ladder has run out; the payload must become addressable.

Apply Reference Data: leave the shelf on disk, search it by the incident code already present in the question, then admit the matching runbook.

The baseline and the shift hold the shelf, question, answer, window budget, and fetch unit fixed. One decision changes: admit the complete corpus, or search before admission. No summary, embedding, rewrite, or child execution is introduced, so neither distillation nor subagent offload can claim the delta.

The baseline treats the shelf as one payload:

episode starts → admit all 800 runbooks → answer ERR-0713

The admission guard rejects it:

The answer exists on disk and remains unusable. Availability has been coupled to residency, so a corpus larger than the window might as well not exist.

The shifted arm treats the shelf as reference data:

episode starts → search disk for ERR-0713 → admit one matching runbook → answer

The exact lexical scan reads the current files and returns one path. Only that path's document crosses the admission seam. The runbook arrives verbatim; the search step selects source bytes and mints nothing.

Run it from the example directory:

python3 demo.py

The script builds the shelf, measures both arms, asserts the budget boundary, asserts a single match, and checks the returned recovery procedure. The generated files disappear with the temporary directory after the run.

The checked run changes feasibility without changing the answer:

The query-shaped arm admits 0.125% of the corpus. The remaining 799 runbooks stay available on disk without occupying the window. The benefit is not that search made the corpus smaller; it made corpus size cease to be the admission unit.

The deterministic checks are executable:

• the full shelf exceeds the declared window budget;
• the exact query returns one and only one document;
• the admitted document fits the same budget; and
• its Recovery: line equals the expected procedure.

Those checks prove the placement claim. They do not prove that a model will choose the right query, that lexical search reaches conceptually phrased questions, or that the returned procedure is true of a live system. The fixture makes vocabulary match perfect on purpose. Misses and broad-query floods remain Reference Data's probabilistic residue; this example isolates the earlier boundary where frontloading is not an available posture at all.

A too-big corpus is not a too-big answer. Once admission is query-shaped, the shelf can grow without forcing every runbook through the window. What must stay bounded is the returned slice.

The incident codes are furniture. The decision travels to policy libraries, schema catalogs, standards collections, and any authored corpus whose queries share vocabulary with answer-sized source units.