Experiments, prediction and controls

Supplied explanation

The laboratory has several distinct kinds of computation. Exact counts check a declared representation. Prediction evaluates whether a fixed model can recover unprovided information. Shuffling tests one specified source of regularity. Data-integrity tests check that code and records agree. None is interchangeable with independent evidence of a generative architecture of Mind.

The earliest pair-order tests learned a majority ordering from the first part of selected occurrences and checked later ones. They preserved informative exceptions but used selected families and familiar data. Canonical chapter order is not a demonstrated revelation chronology. A later unseen-to-model slice does not make the human analysis blinded or preregistered.

The Pass 11 continuation test gave 7/18 correct for its fixed context model against 4/18 for the training-majority baseline. The sample is small and includes a target unseen in training. The Pass 12 broader next-word test used 10,583 test positions: 391 majority guesses, 1,706 one-word-context predictions and 2,287 two-word-context predictions were correct. Shuffling the training words within verses sharply reduced the two-context performance. That establishes usefulness of local order under that task, not uniqueness to scripture or semantic understanding.

The antecedent task is different. Pass 14 used thirty manually selected cues. Nearest preceding verse recovered 16 acceptable anchors, word-weighted overlap 7, and character-trigram overlap 10. Many new cases were local, favoring the nearest baseline. Matching the wrong throne expression is an informative failure because identity and possessive context, not mere lexical similarity, are at issue.

The broad operation graphs in Pass 8 saturated: all coarse edge types appeared and similarity became 1.0. This is not evidence of identical domain topology. The Pass 6 code compared feature sets in a fixed printed order, not directed subject-preserving event graphs. Its stronger result is superseded. The original computation remains available as a record of why the representation was inadequate.

Further tests should preserve all candidates, specify training/test separation, prevent author-selected targets leaking into model input, report uncertainty, compare matched non-Qur’anic texts, and distinguish a corpus-level scan from a manually interpreted subset. New physiology or genome claims need their own independent data. A source narrative and a code integrity test cannot substitute for those measurements.

Basis: all pass archives, especially 6, 8, 10–14; prior source-field control design. Earlier experiments are reported here, not rerun in this consolidation.

Linked contribution records

- K088 — Forecast protocols and familiarity - K089 — Pair orientation tests are conditional tasks - K090 — Permissive detector produced747 windows - K091 — Shared flags did not prove shared topology - K092 — Coarse graphs saturated - K093 — Golden positional tests lack general signal - K094 — Small context prediction test - K095 — Source-wide next-word benchmark - K096 — Shuffled-training comparison - K097 — Reference retrieval remains unsolved - K098 — Software integrity is not hypothesis evidence

Continue through the relations

contextualizes · A Mind in space and time. Read the current Y-Theory explanation alongside this supplied research angle.

continues-to · Next category. Continue in the supplied category order.

Where this belongs in the journey

• Qur’an, Mind and Body · all categories →Research collection
• Search this source collection →Original contributions