Foray 410 · Essai 03 · Fictional analytical rehearsal

What must we reread?

A short certificate can justify a sentence. Finding that certificate may take longer. The distinction matters before we claim a saving in review work.

From a dependency list to a stopping rule

The board has one recommendation left to settle. Each request for evidence takes effort. Which fact should it commission first? Stop when every situation consistent with the evidence so far gives the same verdict. At that point another check cannot change the answer under the stated assumptions.

Our sentence is “A is the cheapest admissible option in the current register.” A costs 8 fictional units. B and C may be admissible and cheaper. The register must also be complete and current. “Reopen” means this sentence is not justified; it does not automatically establish that A is a bad choice.

The completeness check is an actual obligation in this model. Unlisted options in the real world cannot be made absent by a convenient database assumption.

Try a review

Choose a known fictional situation, then reveal its facts in any order. The suggested question minimises expected checking cost under the agreement setting below. All 64 situations remain possible. Costs are invented units.

Facts to check

    Why can we stop — or why not yet?

    Temporary rehearsal only. No report is saved or changed. Changing the situation or restarting clears the revealed facts and cost; refresh also resets.

    When does changing the order pay?

    Start with A. If A passes and B is ruled out, should the team check the register or investigate C? Checking C and its price costs 3 on average and settles the verdict one quarter of the time. Doing that before the register costs 9 on this branch. Checking the register first costs 8 plus 3 times the chance it is current. The cheaper order changes when agreement exceeds two thirds.

    A careful human reviewer can follow the adaptive rule. This is a construction of a defensible stopping rule, not evidence that software beats a competent person. The measured comparison here concerns 64 invented states and declared costs.

    The tempting smaller number is the cheapest certificate chosen after knowing all the answers. A real reviewer cannot choose it using facts they have not read. The interactive trace charges the search.

    What changes in the practical enquiry?

    Selective review needs more than source links. It needs a truth condition for the sentence, a bounded set of possible changes, a way to establish completeness, and a stopping rule that accounts for information still missing. Against whole-packet rereading there may be room to save work. A fixed order ties the adaptive rule across agreement probabilities from one third to two thirds. From two thirds to seven eighths the expected saving from switching order is (3q − 2)/4, where q is the agreement probability. This model gives a reason to measure the relationship before spending money on a new review process.

    This narrows the useful implementation target: expose why a claim can remain, and why another fact would still matter. Do not promise saved human effort from the length of a hindsight certificate. Discovering the dependencies of free prose, reliable source extraction, policy maintenance and real review time remain untested.

    The general question of which check to commission next and when to stop checking passes to 220 · Programme decision sequences.

    To review the wording of a full board paper, open Which sentences can stay? To prepare and record your own wording for the Fellbridge example, open the steering-pack assistant. Each example has its own evidence.

    Source, construction and proof

    Gupta and Mumick(1995), materialized-view maintenance distinguishes maintenance problems by available information, modifications and query features, including negation. That motivates treating “no cheaper admissible option” and register completeness as dependencies. The adaptive questioning algorithm here is our finite construction, not a theorem claimed from that survey.

    The recurrence is V(s)=0 if all compatible worlds agree; otherwise weight the costs of both possible replies by their conditional probabilities and choose the question with the smallest total expected cost. Each question removes an unknown bit, so induction on unknown bits proves optimality within this declared information and cost model. Changing the prior, correlations, fallible checks or admissible questions changes the problem.

    Executable model · Independent checks · Complete calculated result. The tests compare 729 partial states with an independent three-valued Boolean evaluator, execute every policy path, enumerate 720 fixed orders with conditional skipping and distinguish hindsight from prospective cost. At both endpoints of each stated interval the same policy attains its optimum. Every fixed decision tree has expected cost linear in agreement, so the endpoint comparisons certify every intermediate setting.