Searchable cross-round repository

Research teams lose their own work in a predictable way. A round closes, the files go into a shared folder named after the project, and the reasoning inside them stops being findable within a year. Two years later, somebody scopes a question the unit already answered.

A cross-round repository answers questions in prose, with every sentence tied to a passage in a named artifact. The citation is the product. An answer saying that three rounds since 2019 examined this, with the round number and the page for each, can be checked in ten minutes.

The other half of the design is refusal. When no indexed artifact speaks to the question, the system says so and stops. A repository that always produces something teaches people to trust output with nothing behind it.

The scope is everything a round produced, not only its final report. Scoping notes, screening criteria, scoring tables, meeting records, briefs, and the questions left open. Most of the reusable thinking sits in the working artifacts, and the report is a thin layer on top.

Deposit rules matter more than the retrieval model. Every artifact enters with the same header: the round it belongs to, when it was written, the stage that produced it, who owned it, and whether it is current or superseded. An artifact missing those fields stays out.

Status is the field teams skip and the one that causes the most harm. A conclusion revised in a later round still sits in the earlier file, word for word, sounding just as certain. Marking the older version superseded and naming what replaced it costs a minute at deposit and saves a wrong answer later.

Drafts stay out. A working version of a scoring table that three people disagreed with is not evidence of what the unit concluded. Index the version the round signed off, and leave drafts in the folder they came from.

Access rules travel with the artifact. Research archives hold commercially sensitive appraisals, unpublished data supplied under conditions, and notes about named people. Filter the index by the asker's permissions before retrieval rather than filtering the answer afterward, because a grounded answer quoting a restricted passage has already leaked it.

Chunk for citation, not only for retrieval. A passage has to be small enough that a reader can check a claim against it, and it has to carry a locator back to the source. A citation pointing at a forty-page document is not a citation.

The answering step runs on one rule. No sentence without a passage. The model retrieves, writes only what the retrieved text supports, and marks the parts of the question it could not cover. Unsupported sentences are what end trust in a system like this.

Contradiction between rounds gets reported, never smoothed. Two rounds three years apart can reach opposite conclusions, and the honest answer names both with their dates. A merged summary that splits the difference invents a position the unit never held.

Recency is a signal, not a ranking rule. A scoping note from 2018 can be the best artifact for a question about why a criterion was worded the way it was. Print the date on every citation and let the reader weigh it.

Answer types differ, and the system should know which one it was asked for. What the unit concluded. What it decided and why. What evidence it looked at. Where its answer has changed. Those four draw on different artifact types, and a repository treating them as one search returns the final report every time.

The second of those is usually the most useful, and it depends on what got captured in the room. Records of which assumption each side rested on let the repository explain a decision instead of restating it.

The unit and the numbers below are made up. A city housing agency's research team has run evidence rounds for nine years, producing roughly 2,400 indexed artifacts across 61 rounds. A new deputy director asks whether the agency has ever examined insulation retrofits and winter health.

The repository returns four rounds. A 2017 round on fuel poverty, which touched health outcomes in one section. A 2020 round on the retrofit program itself, which examined health and found the evidence too thin to fund on. A 2022 round on ventilation and damp. And a 2024 round on heating tariffs that reused the 2020 conclusion without rechecking it.

The answer runs to two paragraphs with eleven citations. It says the agency concluded once, in 2020, that the health case was not strong enough, and that the 2022 round found damp results pointing the other way in a narrower population. It flags the conflict instead of resolving it.

Two of the eleven citations point at the 2020 open-question list, where the team wrote that admissions data was requested and never supplied. That one line saves the deputy director from commissioning a round that would hit the same wall.

The exchange takes four minutes. The alternative was an analyst spending two days in the shared drive and still missing the 2017 section, because nobody would search a fuel poverty folder for a health question.

The worst output here is fluent, cited, and out of date. It reads well, points at a real document, and repeats a position the unit dropped two rounds ago. Nobody catches it, because the citation makes it look checked.

Deposit discipline decays first. The early rounds go in properly, then a busy quarter arrives and three rounds never get filed. The repository is then confidently silent about work that exists. Audit deposits round by round instead of assuming the process held.

Repetition inside the archive imitates agreement. One conclusion written in a report, quoted in a brief, and restated in a later round looks like three sources agreeing. Resolve citations back to the round that first made the claim, and count rounds rather than documents.

Aggregate questions get answered badly. A question about what the unit usually funds is not a retrieval problem, and pulling six documents will not answer it. That reading belongs to work that surfaces the implicit bar in past decisions, which looks at decisions as a set.

An answer about a past round also says nothing about whether that round was right. Outcomes are a separate join, handled by retrieval that attaches what happened next to each prior case.

A grounded answer is a pointer, not an authority. Whoever uses it opens the citations that carry weight and reads around them. The habit takes minutes, and it is the only thing between a fluent summary and a repeated mistake.

Somebody owns deposits. In most units that is the research lead, and hand-over is not finished until the artifacts are filed with their headers. Treating deposit as optional at the end of a round guarantees the gaps land in the busiest months.

Corrections need a route back. A reader who finds a superseded artifact marked current fixes the status. A reader who gets a wrong answer records the question that produced it. Those fixes are how an index stays worth querying in year four.

Status changes have to flow outward. When a monitor flags that a load-bearing claim has gone stale, the artifacts resting on that claim need remarking, and so do the exec, technical, and operational versions. An index that never changes status is a snapshot of what the unit believed on the day it was built.

The repository decides nothing. It reports what the unit wrote down, with dates and locations attached. Whether an old conclusion still applies to a new question is a judgment somebody has to make and sign.