A research question deserves more than a first answer. With Ariadne v7.5, Innovation Generator moves toward a system that can continue exploring, respond to evidence and leave a research direction worth testing.
A first step toward an autonomous research machine
The ambition is to stay with the question: understand the bottleneck, examine what is already known, explore different explanations and change direction when the evidence calls for it. A useful discovery does not end with a compelling title; it comes with an argument, a meaningful difference from prior work and a next experiment.
Ariadne connects those responsibilities into an evidence-guided discovery loop. It gives criticism a place in the search: a supported objection can eliminate a weak direction, or reveal a stronger hypothesis that a substantive revision can preserve.
From Daedalus to Ariadne
Ariadne v7.5 established traceable structural exploration and research tools. Ariadne carries that foundation toward choosing and revising research directions. Literature comparison grounds the criticism; different candidates are compared; the final dossier states what remains uncertain and which observation could overturn its hypothesis.
That is a substantial change in screening and delivery capability. It makes the research process continue beyond proposal generation and turns the result into material a researcher can take to the next experiment.
A launch grounded in real research runs
The supplied development record includes seven live CLI runs, from early prototypes to release validation, covering LLM reasoning-time compute and a Chinese-input query-optimization case. In release validation, Ariadne generated 37 directions, examined five original finalists and delivered two selected dossiers plus two near misses.
One selected direction, Informed-vs-Naive Surprise, investigates whether information differences between tool-grounded and text-only reasoning can challenge wrong consensus. Another, Coverage-Guided Sampling, studies whether unseen execution behaviors can guide code-generation compute beyond majority stability alone.
The database case examines optimizer lock-in and the feedback available only from previously chosen plans. Its dossiers also retain alternative approaches and the objections that kept them from final selection. Together, the two cases produced seven full research arguments, each with a method, proposed experiments, a pilot and reviewer questions.
These are pre-experiment research directions aimed at rigorous conference expectations. The release retained two of three requested directions, reported degraded quality and disclosed borderline internal reviews. Those judgments belong to the run; they do not replace the proposed research experiments.
Make the cost part of the record
The new release work recorded 175 model-role calls, 17 minutes and $11.04 in Claude CLI equivalent cost. It also reused 173 cached calls whose historical equivalent cost was $9.80. The reported all-in accounting is $20.84 and approximately 30 minutes.
Keeping those accounting scopes visible helps explain what new work cost and what prior exploration contributed. Cached continuation carries earlier evidence forward; it should not make the earlier cost disappear from the research history.
Model access, continued exploration
The release includes Claude Code and Codex CLI adapters, Anthropic and OpenAI API transports, and OpenAI-compatible access. Lite, standard and deep support different exploration scopes, budget controls, continuation and Chinese or English reports. Literature inputs can come from scholarly services, CLI web search or a supplied corpus; the documented live validation used Claude CLI.
The next step starts with a question worth pursuing
Ariadne is a first step toward a machine that can keep pursuing a meaningful research question. The next milestone is to turn its dossiers into controlled experiments and learn from what fails as well as what works. Explore the project for the conceptual architecture, full development ledger and all seven bilingual research dossiers.
EXPLORE THE PROJECT
Learn more about Ariadne v7.5 →