Collection sets the loop's ceiling
The final quality of an auto research loop shows up in verification, but its ceiling is set at collection. If you gather only biased sources, biased knowledge remains no matter how well you verify. The goal of collection is not to gather a lot but to secure enough mutually independent evidence.
In particular, near-duplicate sources that cite the same original are easily mistaken for independent evidence. Failing to recognize this kinship lets one error masquerade as many sources and get promoted to fact.
Diversify and dedupe
Source diversification means collecting evenly across domain, perspective, and recency. Deduplication means filtering not only identical documents but also sentences with the same meaning and documents citing the same original. Applying both raises the count of independent evidence rather than raw volume.
Full guide: from planning to operations
In planning, define collection quality as numbers. For example, set targets like three or more independent sources per fact, a near-duplicate ratio of 20% or lower, and a ratio of new information to volume collected. With such criteria, you can sense quality shifts when you widen or narrow the source list. Since collection happens within a budget, also set rules to drop low-information-gain sources early.
Most failure patterns come from near-duplicate sources and stale caches. Counting five documents that re-cite the same statistic as independent evidence turns an error into fact by majority vote. To prevent this, keep an original-source fingerprint on each document to trace the citation chain, and merge documents with overlapping chains into one for counting. Recovery strategy includes when to stop collecting. If adding new sources raises new information only below a threshold, end collection for that topic and pass it to verification.
On the operations checklist, record source, collection time, and a trust score at the moment of collection. Score trust by domain reputation, primary vs secondary status, and update history, and store low-trust sources for reference only, not fact promotion. Keep a rule to store only a summary and link, not the original, when there is personal data or copyright risk. As observability fields, record independent source count, near-duplicate ratio, new-information rate, and per-source adoption rate.
The continuous improvement loop traces weekly which sources produced facts that verification rejected. If a source has a low adoption rate or high error contribution, lower its trust weight or drop it from the list. A collection strategy should be a living rule whose weights keep adjusting by performance, not a fixed source list.
Key takeaways
In short, an auto research loop's ceiling is decided at collection. Diversify by domain, perspective, and recency; remove near-duplicates by binding citation chains; and weight by trust so independent evidence grows. Manage independent source count and near-duplicate ratio as numbers, and keep adjusting source weights by performance to accumulate knowledge without pollution.