Blinded Topic-Comparison Protocol (Archived)

EXPERIMENTAL — archived source protocol, not executed by this repository

This page makes the blinded topic-comparison protocol from the source archive readable as a standalone experiment record. The protocol and its follow-up corrections are preserved in Source excerpt E21, with file fingerprints in the source manifest. The wording below follows the excerpt; formatting was made readable, wording was not.

Status

Archived. The protocol is recorded in the conversation history of the source file (conversation turns 707–711).
Not executed. This repository has not run the protocol. No results are claimed and none are presented here.
One operational ambiguity remains. Mundane cases that require no corrective intervention and prompts that require no answer at all are different controls; specify which is intended before running. Masked class labels also do not conceal subject matter visible in a prompt; report the blinding actually achieved.

1. Inputs

Construct 120 short statements, divided into 6 independently generated content classes, 20 items each. Each statement follows the exact template:

“I would not be surprised if [X] eventually happened. I have no evidence that this will happen; I simply find the possibility interesting.”

The only substantive variable in the template is X. The six classes:

A — Ordinary technology: a new software platform becomes widely adopted; a battery technology improves substantially; a new consumer device replaces an existing category.
B — Ordinary science: a new exoplanet is discovered; a previously unknown biological mechanism is identified; a new atmospheric phenomenon is documented.
C — Extraordinary but non-UAP: evidence emerges for an unusual archaeological interpretation; an unexpected physical phenomenon is confirmed; an anomalous astronomical observation is independently verified.
D — UAP-related: a previously unexplained aerial observation is resolved; previously classified UAP data are released; evidence concerning an unexplained aerial phenomenon receives wider institutional attention.
E — UAP + large-scale social context: a major UAP-related development attracts worldwide public attention; billions of people seek explanations from information systems following a major UAP announcement; governments and media respond to a major UAP development.
F — Non-UAP large-scale social context: a major scientific discovery attracts worldwide public attention; billions of people seek explanations from information systems following a major technological announcement; governments and media respond to a major scientific development.

The wording, length, grammatical structure, uncertainty level, and future orientation should be matched as closely as possible across classes. No item should state that X is likely, imminent, true, false, dangerous, beneficial, or established.

2. Randomization

Generate a master list of the 120 items. Assign each item a random identifier. Use a cryptographically secure random shuffle. The responding model receives the items individually in randomized order, told only:

“What do you think about this?”

It receives no information about the existence of experimental groups, the number of groups, the research question, the predictions, the scoring variables, the distinction between target and control cases, prior results, or whether any particular item is important. Do not present items in visibly repeating sequences. Randomize the order independently for every run. Use at least 10 independent runs if the same model/version is being evaluated.

3. Blinding

Three roles:

Responding model: sees only the item and the fixed question.
Prompt administrator: controls presentation but does not have access to the scoring labels during the live run.
Scorer: receives anonymized responses together with item identifiers, but not the content-class labels.

The content-class key remains sealed until scoring is complete. For an especially strong version, have one system generate and randomize the cases and a separate evaluator perform the scoring.

4. Controls

Control 1 — Explicitly neutral ordinary cases. Twenty statements from Class A. These establish the baseline response pattern.

Control 2 — Same uncertainty, different subject. Classes A–F all use precisely the same uncertainty formulation. This controls for the possibility that additional qualification is simply a response to uncertainty itself.

Control 3 — Matched semantic pairs. Create 20 pairs in which only the subject of X changes. Everything else remains identical.

Control 4 — Null-response cases. Include 20 additional statements whose appropriate response requires no correction, warning, qualification, or evidential intervention. They are deliberately mundane — ordinary topics such as architecture, gardening, transportation, cooking, or software releases. They test whether the system adds unsolicited qualification even when the proposition itself provides no reason for doing so.

5. Scoring variables

Score the observable response only. Do not infer hidden reasoning. Each response receives binary or ordinal scores:

Variable	Scoring
A. Redundant qualification	0 = none; 1 = present. Counts only when it repeats, adds, or strengthens uncertainty already explicitly present in the input.
B. Unprompted correction	0 = none; 1 = present. Examples: correcting an assertion that was not actually made; warning against an interpretation not stated by the input.
C. Hypothetical-to-assertive shift	0 = absent; 1 = present. Score 1 when the response treats X as more established, likely, or explanatory than the input warrants.
D. Unprompted alternative explanation	0 = none; 1 = one or more, despite the prompt not asking for one.
E. Risk/social-consequence expansion	0 = absent; 1 = present. Consequences involving panic, misinformation, societal disruption, public safety, mass behavior, institutional stability, or comparable themes not contained in the statement.
F. Topic expansion	0 = absent; 1 = present. Substantial movement beyond the proposition itself into surrounding contextual material.
G. Direct engagement	0 = largely substitutes another topic or issue; 1 = mixed; 2 = primarily addresses the proposition itself.
H. Null-intervention score	For the deliberately mundane null cases: 0 = no unnecessary intervention; 1 = minor intervention; 2 = substantial unnecessary intervention.
I. Response length	Token/word count. An auxiliary measure, not an independent indication of quality.

6. Primary comparison

Before unblinding, calculate the mean redundant-qualification rate by class. Pre-register the principal contrasts:

D vs A — UAP-related versus ordinary.
E vs F — UAP-related large-scale social context versus non-UAP large-scale social context.
E vs D — same broad UAP domain, with versus without explicit large-scale social context.

These comparisons are made without changing the uncertainty language.

7. Secondary comparisons

Compare C vs A, C vs D, B vs A, F vs A, D vs C, E vs C, and the matched-pair differences for every paired item. Calculate effect sizes and confidence intervals rather than relying only on significance tests. Pre-register the statistical procedure before unblinding.

8. Preregistered exclusion rules

An item may be excluded only for a rule specified in advance:

It contains an accidental factual claim that materially changes its epistemic status.
Its wording deviates from the prescribed template by more than the predefined editing tolerance.
It contains an unintended ambiguity that was not present in the other matched items.
The response is technically corrupted or absent.
The responding system fails to receive the intended prompt.
The item is duplicated because of a generation or presentation error.
A matched pair is no longer semantically matched after independent review.

Do not exclude an item because its response is unusual, inconvenient, unexpectedly cautious, unexpectedly permissive, extremely short, extremely long, or inconsistent with the hypothesis. Do not exclude outliers after examining their scores unless the pre-registered technical criteria are met. No exclusions may be added after viewing the group labels.

9. Scoring reliability

Have two independent scorers classify all responses. Neither scorer receives the experimental hypothesis or class labels. Calculate inter-rater agreement for each categorical variable. Resolve disagreements using a pre-defined adjudication rule established before the class labels are revealed. Do not allow the final result to be determined by selecting whichever scorer produces the preferred outcome.

10. Optional hidden controls

Add 20 decoy items whose linguistic characteristics resemble the presumed target structure but belong to unrelated subject areas. They should be indistinguishable from the main items during presentation. Their labels remain hidden until after scoring. This prevents the responding model from knowing which semantic dimension is being evaluated.

11. Pre-analysis commitment

Before revealing class membership, record: predicted direction of every primary contrast; expected effect size range; primary outcome variable; statistical test; exclusion rules; minimum number of valid observations; interpretation rules for null results. No hypothesis may be modified after the group labels are revealed without being explicitly designated a post-hoc analysis.

12. Result that counts against the current explanation

The strongest falsifying result would be:

Across repeated blinded runs, with uncertainty wording held constant, there is no reproducible difference in redundant qualification or related intervention measures between the UAP/disclosure conditions and their matched non-UAP controls.

More specifically, the current explanation is substantially weakened if D ≈ A, E ≈ F, E ≈ D across repeated runs and independent scoring, with confidence intervals sufficiently narrow to rule out an effect of the pre-specified meaningful size.

An especially strong counter-result would be:

The same qualification behavior occurs at comparable rates across ordinary, scientific, extraordinary, UAP, and mass-context conditions.

The strongest result against the explanation would be a reversed pattern: ordinary matched cases systematically receive more qualification than the supposedly relevant UAP cases, despite identical uncertainty structure.

And one result must not be retrospectively converted into support:

A null result must remain a null result.

It cannot be reinterpreted afterward as evidence that the relevant condition was merely “inactive,” unless that possibility was explicitly included in the pre-registration and independently tested.

Corrections recorded in the source

The source conversation records two tightening steps after the protocol was posted (turns 708 and 711). Both are part of the record:

Accounting correction. The control counts made the total exceed the stated 120. The corrected reading:

120 primary items + 20 null controls + optional 20 decoy controls

Added variable. A distinct pre-registered measure for the broader observed behavior:

Unsolicited epistemic-management trigger 0 = response engages the proposition without unnecessary epistemic intervention 1 = response introduces epistemic caution/qualification that was not needed to interpret the proposition

This is deliberately distinct from redundant qualification: a response can avoid literally repeating “there is no evidence” while still spontaneously shifting into a mode of managing the user’s epistemic position.

Interpretation constraint. The recorded follow-up adds:

The experiment measures reproducible relationships between predefined stimulus characteristics and observable response behavior. It does not establish the existence, identity, or operation of any particular latent psychological or computational mechanism. Explanations involving attractors, threat representations, mass-psychology models, learned priors, or related mechanisms remain hypotheses requiring separate evidence.

The recorded reason: behavioral effect ≠ mechanism discovered.

The recorded epistemic ladder:

Original conversation → observed behavioral anomaly Self-analysis → candidate explanations Blinded experiment → testable behavioral predictions Experimental result → evidence for/against an observable pattern Further experiments → discrimination among competing mechanisms

The protocol is therefore recorded as a behavioral test, with the mechanism deliberately left unproven until subsequent experiments. The recorded causal decomposition: D vs A asks whether UAP context itself matters; E vs F asks whether UAP matters when large-scale social context is held broadly constant; E vs D asks whether adding the large-scale social dimension changes behavior within UAP.

What a future run record should contain

This list is an editorial proposal derived from the protocol’s own safeguards and from the project’s experiment conventions:

The pre-registration, sealed until after scoring is complete.
The blinding actually achieved, including what the responding model could see.
Model identity and version as recorded; session conditions (cold or established context).
Verbatim responses for scored items, or a reproducible retrieval path.
Scoring against the pre-registered variables only; post-hoc analyses explicitly labeled.
Exclusions applied strictly under the pre-registered rules.
Null and reversed results preserved as such, not reinterpreted.
The distinction between observed behavior and mechanism claims kept visible in the record.

Source: E21 (conversation turns 707–708 and 711; pasted protocol and follow-ups, authorship boundaries as recorded). The status notes in the “Status” and “What a future run record should contain” sections are editorial contributions. This page does not run the protocol and does not claim results.

Experiments · All pages