Research begins with a real question: why does an existing method fail here, which observation remains unexplained, and what evidence would change our understanding? A compelling title can open a discussion, but a research direction also needs a reason to pursue it, a meaningful difference from prior work, and a way to test its claim. Ariadne v7.5 brings those responsibilities into the comparison, revision and selection of candidate ideas. It is a step toward an autonomous research machine: one that can stay with a question and leave an argument a researcher can carry into the next experiment.
Daedalus v7.4 established traceable structural exploration and tool building. How a question is reformulated, whether a relationship can transfer between fields, and which calculations or checks a proposal needs should all have an understandable history. Ariadne builds on that foundation while moving the focus forward. When many candidates appear, a researcher still has to decide which deserve more work, which merely change the wording, and which depend on an unsupported assumption. Literature comparison, independent criticism, candidate comparisons and substantive repair connect the exploration record to a research direction with an explicit argument and a testable path.
Literature has a specific job in that decision. A method may share a structure with prior work while addressing a different phenomenon; another may introduce new terminology without changing the problem already solved. Ariadne asks for that distinction to be made explicit, so that criticism addresses an identifiable assumption, neighboring method or evidence gap. A supported objection can change whether a direction survives, while a contested judgment receives a second opinion. This gives exploration room both to let a weak proposal go and to develop a promising proposal along the gap its critics reveal.
One reasoning-time compute case proposed using peer predictions to detect situations where most answers agree and are wrong together. The initial idea drew on information aggregation, but independent review identified a missing premise: repeated samples from one model do not automatically have the heterogeneous information the theory requires. That objection changed the question. The repaired proposal, Informed-vs-Naive Surprise, compares reasoning supplied with independent execution evidence against text-only reasoning, asking whether a real information difference can expose mistaken consensus. It entered the internal selection; the next experiment must establish whether that difference exists and whether it improves decisions.
Another selected proposal, Coverage-Guided Sampling, studies how compute is allocated in code generation. More samples can waste resources on a solved problem, while a repeated wrong behavior can make a majority look stable. The proposal examines execution behaviors to estimate what remains unseen and decide which problem deserves further sampling. Its value turns on a clear comparison: at a fixed compute budget, does this signal add information beyond majority stability, and can it beat the strong baseline of stopping at the first public-test pass? Invalid programs and wrong behaviors may also create misleading novelty. Those risks remain part of the proposed pilot.
The live release validation began with reasoning-time compute and generated 37 candidate directions. Review and repair produced two selected dossiers and two near misses. It had requested three selected ideas and obtained two; the record retains that shortfall and the borderline internal judgments. Delivery therefore includes both the decision and its reasons: how a proposal changed, which objections remain unresolved, and what result would make it unworthy of further study. Selection is the outcome of this exploration run. Scientific effectiveness still awaits the research experiments proposed in the dossiers.
A separate database query-optimization case carries the same approach into a different problem. An optimizer usually observes the plans it has selected and executed, while alternative plans lack comparable feedback. Under a changing workload, this observation process can keep favoring earlier choices. The dossiers examine optimizer lock-in, feedback bias and bounded exploration, asking which observations exploration adds and what a correction contributes beyond them. Together, the reasoning and database cases retain seven complete research dossiers, including selected proposals and alternatives that did not make the final cut. Each develops its question, method, proposed experiments, pilot and reviewer concerns, allowing the reader to inspect the argument beyond the title.
The release rests on records from seven live CLI development runs, covering prototypes, iterative repairs and final validation. Heavy candidate rejection, account limits and failed calls all informed its development. Preserving different directions helps prevent one judgment from collapsing exploration into the same answer; retaining failures and repairs explains why the final proposal took its present form. Inputs, settings and cached work differed between runs. These records document a development trajectory and operational experience, while the next controlled studies will examine which changes produce repeatable benefits.

Exploration also has to account for real resources. New release work recorded 175 model-role calls, 17 minutes and an equivalent cost of $11.04. It reused 173 cached calls whose historical equivalent cost was $9.80, giving an all-in accounting of $20.84 and approximately 30 minutes. Separating new work from earlier exploration makes the cost of continuation understandable. The release includes Claude Code and Codex CLI adapters, Anthropic and OpenAI API transports, and OpenAI-compatible access, with bounded exploration, budget controls, continuation after interruption, and Chinese or English reports. The documented live release validation used Claude CLI; other interfaces retain their own integration and verification scope.
Ariadne aims to give research continuity: a question can be pursued further, criticism can change a method, and failure can leave knowledge for the next exploration. It organizes ideas around rigorous research expectations, while publication and scientific value must be earned through actual work. The next step is to take the dossiers into controlled experiments, let new observations revise the judgment, and gradually connect discovery, validation and delivery. The project page develops the conceptual architecture, the run history and all seven bilingual dossiers. Exploration still starts with a question and continues along evidence capable of changing the answer.
