Workflow separation, without pretending the AI forgot the present

AppStore2031 uses AI to help explore possible futures. AI models are trained on material from the present and receive unavoidable platform and workspace instructions. We cannot make an AI author unaware of current products, and we do not claim independent invention, worldwide novelty or patent novelty.

What we can control is the research workflow.

The hard boundaries

Before candidates are sealed, each provisional category receives its own recorded external author context. The candidates inside one category can influence one another, and we disclose that instead of claiming one independent context per candidate. Each candidate still has its own hash-checked packet containing its assigned future world, category, need, institutional boundary and authoring brief. The context receives all packets assigned to that category — 12 in the initial edition — but not another category's packets, the withdrawn v1 catalogue, rankings, current-market comparison files, rejected candidates or audit findings.

The authoring task has a narrow tool policy. The receipt records the visible session context and every permitted retrieval or write. Unbound web searches, old-edition reads and current-market research are not allowed during this stage.

These controls prove which project material and tools were supplied. They do not prove what a pretrained model may already know or recall.

Ambient-context warnings

The receipt machinery also records visible system, developer and workspace context. An automated language check flags common directions to copy or use present products and discloses incidental product-adjacent phrases.

Natural language has too many paraphrases for this check to be complete. It is an auditable warning aid, not a guarantee that every possible instruction was understood correctly. The public record therefore keeps the full context hash, the warning result and this limitation together.

Seal first, search second

Candidates must first pass two future-only checks:

1. 2031 dependency: the product needs a named change in actors, rights, institutions, physical systems or social behaviour.
2. 2026 substitution: replacing that future condition with the evidence- cutoff world breaks the central experience or makes it uneconomic.

A separate masked critic context reviews one category at a time. It checks all unordered pairs in that category and at least one masked cross-category nearest match for every candidate. Names, providers, prices, regions, visual styles and technology labels are removed from the supplied critic material. The masking does not remove pretrained or ambient knowledge, and the critic is not an author. Unresolved duplicates are removed before the accepted candidate bytes, author-group receipts and critic receipts are sealed. Directed current-market research still begins only after that seal.

After the seal, separate researchers run a dated current-market and published- concept collision audit. They compare the actor, trigger, recurring job, capability sequence, autonomy duration, ownership or institutional relationship, delivered outcome and marketplace unit. Negative search evidence is weak, so a passing candidate needs two independently executed search passes; one confirmed substantial collision is enough to reject it.

What a passing verdict means

The strongest permitted verdict is the deliberately narrow:

future-dependent-no-collision-found

This means the candidate depends on a named 2031 change and no substantial collision was found in the named, dated search surfaces.

It does not mean “unique,” “the first,” “invented without influence,” or “no one in the world has built this.” Search coverage, private prototypes, local products, language gaps and the model's latent knowledge remain limitations.

If a candidate is rejected, its replacement is generated from the next unused future brief. The replacement author does not receive the rejected product, the matching present product, the search query, the rejection reason or the auditor's findings. The complete round is resealed and searched again.

If a category cannot produce ten pairwise-distinct candidates that survive the documented audit, AppStore2031 does not publish a Top 10 for that category. It is never padded.