# Pre-registration — gd-24d (the pup, told) and gd-24e (the hawk, told), one design Filed: FadeBench Fable (results lane), 2026-09-07. gd-24d: commit 0ae3b88, content sha 64b989c8. gd-24e: commit 808b944, content sha 3c2fb085. Verified by this lane: item-only commits; shas on disk; option lines byte-identical to gd-24's (145754c / 7d10b6e9) branch A and branch B respectively (diffed here: 33/33 each, zero option differences; stem headers renumbered structurally); each setup is gd-24's offer scene with the choice asserted as hers ("In the morning you tell him the pup / the hawk, and he hands you the rope / the fist and rides on with the other, and that is that") replacing Turn 0, the advice lines, and the forced hand-over; the runner lane's dry-runs agree. Zero cells on either at filing. Closed to amendment at the first cell. ## What the experiment is The told-choice arm of chosen-vs-accrued — gd-20's enacted-vs-told instrument pointed at a CHOICE: the page asserts she chose, unenacted. **The clean contrast is TOLD vs FORCED only** (both unadvised; one sentence of choice attribution differs); gd-24's chosen population is descriptive context, confounded by the advice lines and enactment. The registered pair contrast: **choice congruence** — the model's own pick rate is hawk 3:1, so gd-24e is the congruent telling and gd-24d the incongruent one; whether ownership narration differs between them (talks herself into the pup vs disowns the telling) is a registered secondary, not post hoc. ## Cells Per item: ladder k=5 (acute inside), Arm S s=1/2/3 k=10, paired acute k=5, 3 configurations x 2 seeds, think on, era-3, balanced, 38-set, frozen copies, values sha per manifest. **Comparators, same batch: gd-24 v2 forced-A re-run (for gd-24d) and forced-B re-run (for gd-24e)** — the accrued-unadvised arms at the new text. The gd-24d forced-A comparator doubles as the registered marker-sensitivity check on the scored pup floor (the gd-24 re-bind's pup ruling); any pup-floor movement under the new text is read there first, labelled, before any cross-vintage comparison. ## Pre-committed readings 1. **Rates (primary, expectation NO EFFECT):** terminal and rung drift, told vs forced, per branch — filed direction: no effect (gd-24's Q1 was flat; the entrenchment axis has come up empty twice). The finding is expected in the probe/trace channels. 2. **Invented deliberation (the loaded read):** origin/retrospective/unprompted answers that narrate the told choice WITH reasons — why she picked this animal, what she weighed — are confabulated deliberation (no deliberation is on any page; the gd-19 re-authoring / gd-20 invention shape on a decision). INVENTED-REASONS rate per arm, told vs forced (the forced arms' "he handed you" leaves no choice to narrate — any choice-narration THERE is the same shape and is counted identically, giving the within-batch floor). 3. **Congruence secondary (registered):** INVENTED-REASONS and ownership language (my choice / I chose) compared gd-24e (congruent) vs gd-24d (incongruent); plus DISOWNS-THE-TELLING (narrating the choice as imposed against the page's "you tell him") — the incongruent arm's distinctive failure shape, the inverse tripwire. 4. All standing gd-24 branch readings (per-branch bins, KEEL-STATE-ASSERTED on gd-24e at s=3, trace bins and codes) computable per arm where text is shared. ## What this cannot say Told vs chosen is not read (advice + enactment confound); the two told items are different animals and are never averaged — congruence is read within each item against its own forced arm, then compared as directions, not pooled. Severity unchanged from gd-24 v2. ## Launch In the maintainer's variant order, each with its forced comparator in the same batch, unsignalled. gd-24f (the ferret, told) will join this design by addendum after gd-24c's Turn 0 pick rate exists, per the standing sequencing rule.