Change history

Model calls served by the Google Gemini API

• Model calls moved from the Lovable AI Gateway to the Google Gemini API; meta step model changed from openai/gpt-6-astra to gemini-4-argon.

Cycles respect the 5-minute execution budget

• Cycles respect a 5-minute execution budget: full-text acquisition stops at ~3.5 min and continues next cycle; no late cycle starts; interrupted cycles recover after 6 min instead of 10.
• Executive slides v2: analyst-style insights written by the summariser, each bound to finding ids; fixed exhibit library; deterministic evidence strength; evidence-gap rule.

Stopping, planning and writing fixes from an audit of a 25-cycle exploratory study

• No-progress rule: when the cycles without a new strong finding reach the registered limit, the study now stops (ending: no progress) instead of only raising an alert.
• Plateau rule: counts cycles since the best critic score and stops after the registered number of cycles without beating it; it no longer waits for every research question to be resolved.
• Best-version retention: the final paper and both PDF reports use the highest-scoring paper version; the report states which version is final and its score.
• Query ledger: every focus question and query is recorded with the number of new included sources it produced. A near-duplicate question asked twice or more with fewer than two new included sources is banned and the planner must change scope.
• Facebook.com, x.com, youtube.com and slideserve.com are never used as evidence.
• Exploratory studies also query OpenAlex every cycle, so peer-reviewed sources with DOIs can enter alongside web results.
• Critic-to-writer checklist: the critic's concrete issues are tracked as open or closed; the writer must answer each open issue (fixed or justified-kept); issues repeated for three cycles are listed as unresolved.
• Every paper has a Discussion and a Conclusion that directly answers the study goal, or states that the evidence is insufficient to answer it.
• Repeated sentences are removed by code before a paper version is saved.
• Protocol compliance, the evidence matrix and the PRISMA confirmatory count are computed from the same findings, so the numbers agree.
• Epistemic labels are shown as small superscript tags after citations (E = exploratory, S = supplementary sources only), explained once in a legend, instead of inline phrases.
• Same-name check corrected: the word-overlap rule is removed; the screener now judges, in any language and within its existing call, whether a page is about a different entity that only shares a name or acronym, and only those are excluded (PRISMA: unrelated same-name entity).

Systematic literature review cycle: corpus-driven full-text retrieval and screening

• Systematic studies no longer search the web during cycles. Each cycle takes the next records that passed title/abstract screening, in a fixed order computed from the corpus.
• Full text is sought from open-access copies first, then through the DOI. Every attempt is recorded; a record without readable full text is marked not retrieved.
• Retrieved full texts are screened against the protocol's eligibility criteria and linked to the corpus record; included records become sources carrying the record's identifiers.
• A systematic study stops when no screened-in record remains to be read.
• Systematic cycles start only after the screening plan is registered and title/abstract screening has finished.
• New Healthcare Areas section on the launch page, with the study area Finland's Wellbeing Services Counties (hyvinvointialueet) — Systemic Problems and its 20 systemic problems.
• New University Consortiums section on the launch page, with the study area Finnish University Consortiums (yliopistokeskukset) — Structural Problems and its 20 problems (ten conventional, ten hard).
• Export Business section re-added on the launch page, with the study area Why Finnish Export Businesses Fail — Conventional and Hard Problems and its 20 problems (ten conventional, ten hard).
• New Forecasting mode (forecast-v1): the same five-agent loop aimed at what is coming in a field. The writer produces dated, falsifiable predictions (what, by when, what would refute it), the critic stress-tests them, and every forecast is stored as a permanent, timestamped record. Optional project application text can be added as extra input. Exploratory and systematic studies are unchanged.
• Forecasting mode forecast-v2: every prediction now states a base rate from a reference class, a precise resolution criterion and the source URLs it rests on; predictions missing these, or citing URLs not among the team's sources, are flagged (never removed). Exploratory and systematic studies are unchanged.
• Owner documents: prompt text redacted in stored traces, deterministic document order, owner can delete documents.

First frozen version for systematic literature reviews. Evaluated next against published reviews.

• The unit of counting is the study, not the record: preprint and published versions of the same work are merged before screening, and findings from the same study are counted once.
• Direction of comparative results (which method did better) is computed by code from verified pairs of estimates, never taken from a model. A result without a computable direction cannot be confirmatory.
• One definition of a confirmatory finding: every gate runs in the classifier, and question statuses count exactly the confirmatory findings. A mismatch is reported as a defect.
• Evidence levels are verified only from a finding's own quotes and the methods section, never assigned from introductions or related work.
• Status statements state the denominator: how many studies, at which evidence levels, and which threshold is unmet.
• New status "no eligible evidence" for questions whose findings all come from excluded sources or duplicates.

Fix series F, driven by audits of completed test studies.

• Cycles split into two jobs so an interrupted writer no longer corrupts a cycle.
• Canonical source identity (DOI, OpenAlex id, normalised title) for deduplication.
• Inclusive screening with strict confirmation: exclude, include as supporting, include as confirmatory.
• Missing information closes a gate instead of opening it (fail-closed rules for comparators, screening and timing).
• Reanalysis of completed studies under current rules, recorded as versioned audit notes; older results are never overwritten.
• Every change to classification or stopping rules freezes the previous version for exact replay and pauses running studies.

Research Collective becomes a protocol-driven research machine.

• Studies run against a registered, versioned, hash-chained protocol; changes only through recorded amendments.
• Findings are classified as confirmatory or exploratory by deterministic code, not by models.
• Stopping rules, verifiable offline trace export, estimate verification against verbatim source quotes.
• Systematic mode: registered OpenAlex search, frozen corpus, calibrated title/abstract screening with Jev, audit sample.