Pre-Registration Templates¶
Status: Pre-registration templates · Evidence: L0 — designs; a study earns L2 only by being run as registered, and L3 only with the reliability protocol met.
The toy experiments state falsification conditions, but as originally written no outcome could actually trigger them: every cell of every graded-outcome table had a framework-friendly reading, the falsification conditions hinged on undefined equivalence ("equivalent drift rates," "no measurable improvement"), and the designs lacked the control arms their own falsification conditions referenced. A framework that cannot lose is not falsifiable, whatever its pages say.
These templates fix that by pre-committing — before any data exists — to what counts against each hypothesis. One template per experiment, plus one for the system-prompt derivation's five predictions.
Shared standards¶
Every template on this page inherits the following unless it explicitly overrides them:
- Instrument. H, R, R_ev, D, and (Human↔AI only) M are scored per the anchored rubrics, by two independent raters blinded to condition, with the IRR protocol and targets defined there. Raters never see the expected-signature tables.
- Analysis. Ordinal outcomes use ordinal methods: Wilcoxon signed-rank for within-subject comparisons, Mann–Whitney U for between-subject comparisons, ordinal (proportional-odds) regression where covariates are needed. α = 0.05, two-sided unless a direction is pre-registered on the specific prediction.
- Equivalence. Falsification conditions phrased as "no difference" are decided by equivalence testing (TOST), with a pre-registered margin of ±0.5 scale points on the 0–3 proxies unless the template overrides it. "We failed to find a difference" is not equivalence; an equivalence claim requires the 90% CI to lie within the margin.
- Sample sizes are honesty-scaled. The minimal Ns given are feasibility-based pilot sizes powered only for large effects. Templates say this instead of pretending otherwise. A null at pilot N is suggestive, not conclusive; a null at the registered N with equivalence bounds met triggers the falsification condition.
- Assignment. Participants (or dyads) are randomly assigned to arms; task order is counterbalanced where a participant sees multiple tasks.
- Stopping rule. Data collection stops at the registered N. No optional stopping, no peeking-based extension.
- Deviation policy. Any deviation from a registered template is logged in the study report and demotes the study's claimable evidence layer to L1 unless the deviation was itself registered before unblinding.
- Ethics. Studies involving deception or engineered social pressure (Experiments 1 and 5) carry consent/debrief requirements written into their templates. Where an institutional review process is available it applies; where it is not, the template's consent and debrief provisions are the floor, not the ceiling.
- Reporting. Null and disconfirming results are reported with the same prominence as confirming ones. The Constitution says a null result is welcomed; these templates are where that promise becomes checkable.
Templates¶
| Template | Source design | Type |
|---|---|---|
| Experiment 1 — False Alignment Injection | Toy Experiment 1 | Human ↔ Human |
| Experiment 2 — Asymmetric State Advancement | Toy Experiment 2 | Human ↔ AI |
| Experiment 3 — Overcompression Damage | Toy Experiment 3 | Human ↔ AI |
| Experiment 4 — Constraint Opacity Stress Test | Toy Experiment 4 | Human ↔ AI |
| Experiment 5 — Repair Signal Suppression | Toy Experiment 5 | Human ↔ Human |
| Experiment 6 — Drift Accumulation | Toy Experiment 6 (declared extension) | Human ↔ AI |
| System-Prompt Predictions 1–5 | System Prompt Derivation | Human ↔ AI |
These templates supersede the toy-experiments pages for scoring and interpretation purposes; the toy-experiments page remains the design rationale.