Question dossierOrdivon Financeanswered

When does a useful research constitution earn default Agent context?

Freeze context treatments before holdout reveal and require replicated marginal benefit before turning persuasive domain knowledge into mandatory prompt state.

01 / Current position

A hypothesis is not the current judgment.

The dossier preserves the live research position. Dated articles preserve the complete evidence and argument that changed it.

Working hypothesis

Domain knowledge should remain equipment, priors, baselines, and falsifiers unless an unchanged context treatment shows stable sealed benefit across materially different carriers after token and search cost are counted.

Current judgment

Answered for the current Quant Bedrock default-context claim. Constitution-only looked less bad than control on QB3b but remained negative; the unchanged treatment became materially worse than control on the second sealed QB3c carrier and used about 27.5% more Provider tokens. Mandatory injection has not earned default status.

02 / Decision boundary

What keeps this Question alive?

Why it matters

A constitution can change Agent behavior, candidate selection, reasoning style, and apparent caution while still failing to improve the objective consistently across worlds.

Next admitted test

Keep Quant Bedrock available as equipment rather than mandatory context. Reopen only with a frozen treatment that replicates benefit across additional carriers/providers without holdout-guided tuning and beats a simpler on-demand equipment path.

Condition for deletion

A frozen constitution treatment repeatedly improves sealed outcomes across independent carriers/providers at acceptable marginal cost, while equivalent on-demand retrieval or equipment does not recover the benefit.

03 / Supporting publications

Complete arguments connected to this Question.

1 dated publication currently document this research line.

04 / Source discipline

The dossier is an index, not the evidence authority.

Question metadata

Owns the current judgment, next test, and deletion condition.

Publications

Own complete dated arguments, limitations, comparisons, and source links.

Repositories

Own exact code, tests, releases, receipts, and machine evidence.