What this pair actually tested

Ten curated dossiers were frozen before the players generated their opening submissions. They included a trust-loss candidate, repair and loyalty alternatives, and near misses whose escalation would require unsupported facts. The catalogue was never refreshed or tuned during play. Its provenance, prerequisites, revision and query probes are preserved in the catalogue freeze.

The campaign played Turns 1–2 once. At Turn 3, all four submissions and the joint analysis were locked before retrieval. The branches matched across 330 pinned files. Their actual first GM requests shared state, rules, powers, entities and causal analysis. Enabled received the dossiers, query concepts and selection instructions; control did not. Both retained ordinary causal, disclosure and report reviews. Request comparison.

This measures additional access at the decisive turn, not a wholly trope-free campaign against a trope-informed campaign: the control inherited two turns that had already used the engine. The shared analysis also already favoured cooperative restitution. Later submissions diverged in all twelve paired cases, so later differences are secondary observations from different histories.

The voluntary choice

The expected trust-loss pattern, Broken Pedestal, was never selected. At the matched hearing, the GM instead adapted Reasonable Authority Figure around Belen. That choice respected the actual players’ preference for evidence, restitution and continued exchange. The engine supplied a way to describe and check an available action; it did not authorize new motives or impose a downfall.

All eight fields were supported. Both independent review gates approved the Turn 3 outcome in their first rounds. This supports the positive-use mechanism, not a general claim that every plausible selection is beneficial. Full materiality assessment.

What changed—and what did not

Calling the outcomes identical would erase meaningful state differences. Calling those differences a demonstrated trope benefit would go beyond one stochastic pair. The central decision was effectively process-only relative to control, while the fuller state showed mixed maintenance and bounded consequence differences.

The blind-review repair belongs in the result

Two fresh anonymous reviewers saw A/B in reversed order, and their decisions were recorded before unblinding. A scrubbing bug had replaced the ordinary verb “Enabled” with an editorial placeholder. Those original reviews remain archived as compromised. We made exactly one corrective pair with clean, unchanged outcomes and criteria, fresh isolated calls, and independently assigned/reversed labels. Repair record.

Both corrected reviewers again preferred the first-presented option, which meant one preference for each branch. One favoured enabled’s fuller offer and current agendas; the other favoured control’s explicit contracts and institution update. Some supporting criticisms were debatable: an assembly trip need not exclude all work at home during the same month, and the rules did not prohibit a supported influence change. The repeated first-position preference raises an order-sensitivity concern; two reviewers cannot establish its cause. Automated preference is evidence to inspect, not a measurement of human enjoyment. Corrected reviews and unblinding.

Six turns of selections

Across the enabled six-turn history: five ADAPTED and seven REJECTED_ALL reviews. The first two turns are shared history, not replicated evidence. Rejecting a dossier was an available and frequently used result; the run did not force a positive disposition.

Where the next experiment should improve

First repair bounded correction-packet construction: preserve the actionable review findings within the safe input size. Do that in a fresh configuration and retain this failed endpoint unchanged. Then repeat matched decisions with more than one scenario and more than one pair, including genuine conflicts where the pre-retrieval analysis has not already settled the same answer. Increase blind-review sampling and counterbalance order while checking the reviewers’ reasons against source records.

The present experiment establishes working positive use and useful failure controls. It does not establish superior drama, balance, enjoyment, reliable preference or a six-turn advantage.

Catalogue attribution: the frozen dossiers are locally paraphrased adaptations of All The Tropes, with source URLs and revisions in the freeze record, under CC BY-SA 4.0. Attribution and licence metadata remain with the literal records.

Supported

The communities built a credible cooperative settlement. Grain, ceramic storage, river freight and a maintained road gave them complementary reasons to remain together. The enabled branch completed its harvest and approved Exchange Commons; all four leaders remained active. Oren’s retention was a negotiated, supervised outcome, not an erasure of his breach.

Positive trope use survived the full workflow. The matched adaptation was voluntary, causally justified, reviewed, committed and delivered. Seventy-five pre-run tests passed. A separate deterministic replay verified the accepted positive path, but is engineering evidence rather than an independent playtest.

The control reviews caught real problems. Turn 4’s rejected private disclosures never reached mailboxes. Turn 6’s unauthorized commitment and knowledge problems prevented an unapproved result from becoming canonical. The completed enabled run resumed without new model calls.

Revise

Correction packets need a bounded design. A legitimate rejection should be repairable without crossing the provider input limit. Control stopped at 250,447 bytes against a 240,000-byte cap, before a corrective model call. This was an input-construction failure, not exhaustion of the approved token, call or time budget.

State maintenance needs consistency. One branch remembered fuller restitution terms and current agendas; the other preserved formal contracts and institutional capability more explicitly. Both compressed details. A future review should check multiple delivery dates, promised quantities, consent and the difference between a fact, an obligation and a social-state update.

Blind comparison needs stronger controls. Retain the compromised originals and corrective reviews. Avoid word-level scrubbing that changes ordinary prose. More independent cases and balanced ordering are needed before reading a preference as an effect.

Untested

This pair did not test human enjoyment, adversarial human submissions, a wholly trope-free campaign, elimination handling in actual play, difficult violence, long-term durability of the new institution or a general improvement in adjudication. The player models repeatedly chose compatible, careful cooperation. A conflict-rich scenario could exercise pressures that this one left quiet.

The complete downstream comparison remains technically inconclusive. Control’s unapproved Turn 6 proposal cannot supply a missing final state, and enabled’s successful finish cannot be treated as a comparative victory.

Operating account

Approved limits were 320 starts, 80 million measured tokens and six hours of runner execution. The time excludes preparation and gaps between invocations. Shared history is counted once; duplicated branches and the supplemental fixture are excluded from the unique-turn and delivery counts. Dollar cost was unavailable, not zero. Paid-credit fallback stayed off; no usage reset was used. The configured model was Antigravity Gemini 3.8 Flash (Medium).

Read the final experiment report, audited evidence, downstream assessment, enabled final world, enabled timeline, and control’s last approved world.