Giants’ Shoulders: the research system, shown in screenshots

Purpose and the three people it serves

A university runs three systems of record for its scholars. AutoScholar advises students and the staff who look after them. ScholarCloud teaches, assesses and keeps the academic record. Giants’ Shoulders is the third: it holds the work of research, and it is the one that speaks to a library, because a library's document repository is its first component.

Three people use it. The researcher, a postgraduate or a member of staff, keeps a repository of the project: the models used and what each may do, the record of every interaction with a model, the bibliography with each reference verified at its source, the checks on computations and figures, the decisions taken before technical work, the experiments, the drafts, and the AI-use declaration. The supervisor reads that repository against a rubric of six criteria, records the reading, and logs the meetings. The research office and the library see the programme: projects, papers, grants, candidatures, ethics, and the repositories connected to the institution's collection.

Everything shown below was captured headless on the demonstration centre, signed in as the researcher Bongani Zulu and as the supervisor Megan Botha. Nothing is mocked. Empty lists are empty because the demonstration centre holds no rows there yet, and each says so.

Sign-in and the hub of components

One account serves the researcher, the supervisor and the centre. What each person sees is decided by the rows: a researcher sees their own repositories, a supervisor the ones whose candidature names them, a director all of them. After signing in the person chooses a component from a grid of eight tiles. A tile opens the component with its screens along the top of the page, and the name at the top left returns to the grid. The bell carries the unread count of notifications.

Repository component: the research record and its readiness

The list on the left holds the repositories a person owns or supervises. The strip at the top is the readiness against the rubric: six indicators, one per criterion, each stating whether the rows are absent, present, present and checked, or present and checked with the reasoning recorded. Beneath it the record itself (what the project asks, the highest class of material held, the initialisation file, the style file) and the AI-use declaration with its versions share the width, then the provenance log with one dated line per interaction with a model, the session logs, and the files and folders anchored to the repository.

Models component: the model register, agent scopes and benchmarks

The register names each model the researcher uses, where it runs, its price per million tokens, its context window, what it may do on this project, what it must not do, and the highest class of material it may see. Agent scopes state what an agent may touch, per tool. Benchmark cases are the tasks with a known answer that choose a model by measurement, and the counters at the top add them up. A second screen shows the hosted model's metered use and the account's tier.

Literature component: connected repositories, including an existing DSpace

This is the component built for a library. The first screen is Repositories. A person adds a repository with a code, a name and the OAI-PMH address that every DSpace, EPrints and comparable repository publishes, and presses Connect. The Publoom engine harvests the repository in the background into a corpus of its own, embeds each record, classifies it against a canonical spine of 252 research subfields, and keeps a register of every repository it holds with its record count. The repository is the connecting person's until it is shared with everyone on the database, and every question, draft and search then spans the repositories the person holds. Full texts resolve against each repository's own server.

The screen below shows two DSpace repositories connected from Giants’ Shoulders: the engine's own repository, and UKZN ResearchSpace, an external DSpace with 22,289 records of which the first 150 were harvested for the demonstration. A question about water and sanitation was then answered with citations to the harvested theses.

The remaining screens of the component are Publoom Press for Researchers inside Giants’ Shoulders. Ask and draft answers from the connected repositories and the personal library with numbered citations, drafts a grounded section, matches funding calls, composes documents and takes uploads into context. My library holds the papers the person saved, uploaded or fetched by DOI, with an import of their own papers from the centre's register. Keeping up runs standing questions over the open literature. Funding places the centre's calls on a calendar beside a funding-fit prompt. The collection is the map of the holdings by field. The reader opens a paper with its full text.

References are the researcher's own bibliography. Each entry carries an identifier and a verification status. Verifying resolves the DOI at OpenAlex: verified when the title at the source agrees with the title held, mismatch when a paper exists under that identifier but is not the one claimed, fabricated when nothing resolves. Each verification writes a check into the analysis record.

Analysis component: checks, data, analysis, models and computation

Checks is the researcher's record of what was tested against a known answer: a code test, a computation, a figure rebuilt by one command, a data sample, a claim. The other screens are the Analyst Gateway inside Giants’ Shoulders: Data holds datasets as shared workspaces, Analyse runs distributions, relationships, correlations, time series, regression and multivariate analysis over a dataset, Model builds models, Compute holds twelve workbenches from expressions and calculus through fitting, optimisation and differential equations to graphs, operations research and games, and Connectors pulls data in from external services.

Writing component: documents, decisions, the course and live systems

Documents are block documents on a canvas that drafts and cites through the engine. Decisions hold the design sheets and the choices made before technical work: the question, the options, the choice, the reason. The course is read in place. The remaining five screens are Schema Studio inside Giants’ Shoulders: a researcher designs the schema of a study as records, previews it over the in-memory platform, provisions it onto the database, and runs it, so a design sheet becomes a running data system without leaving the page.

Programme component: experiments, the centre's programme, samples and quality control

Experiments are the laboratory notebook: hypothesis, method, a status that moves along a guarded path from planned through running and analysed to written, the dataset and the sample codes used, and the checks made on the run. The next twelve screens are the whole of the research centre's programme from Centrix: projects with their task trees, the person's own work and notifications, papers with their writing tasks and authors, grants, conferences, collaboration, researchers, themes, the harvest of publications to claim, insight, reports and the calendar. Samples and QC are the laboratory system's own screens: specimen lineage, custody, storage, results and control charts.

Supervision component: candidatures, the rubric review and the cohort

Candidatures are the centre's supervision records: a candidate, a thesis title, a degree, a stage, a supervisor and a co-supervisor. A repository names the candidature it is the portfolio of, and the supervisor named there reads it against the rubric and records a review: complete when every criterion is at least checked, returned with the criterion named otherwise. The review reaches the candidate's bell, and a signed declaration reaches the supervisor's. Meetings are logged, and the cohort table derives cadence from logged meetings only. The centre's full supervision register and its ethics register are the last two screens.

Course and rubric

The course Working with AI on your own research is ten half-day sessions or a five-day block, and its assessment is the repository itself, read by the supervisor two weeks after the last session. Each chapter ends with an addition to the repository, and every one of those additions is a row in Giants’ Shoulders: the model notes of chapter 1, the provenance log and the data classification of chapter 2, the benchmark cases of chapter 3, the verified bibliography and the declaration of chapter 4, the design sheet of chapter 5, the workflow and agent scope of chapter 6, the tested analysis of chapter 7, the style file and revised chapter of chapter 8, the data checks of chapter 9, and the supervisor's reading of chapter 10.

The rubric has six criteria, one per learning outcome, each at one of three levels: present, present and checked, present and checked with the reasoning recorded. Readiness in the Repository component is derived from the rows by one rule per criterion and is advice. The supervisor's review is the record. A university library that trains researchers has in the course a curriculum and in the system the place where the curriculum's work is kept and read.

Composition from existing systems

Giants’ Shoulders owns one small service of its own, the research record with its twelve tables. Every other component is a sister system's own services and screens loaded into Giants’ Shoulders's page and relabelled, on the same database and the same members. Nothing is copied.

• Publoom Press for Researchers supplies the Literature component: the repository engine, ask and draft, the personal library, watches, funding fit, the collection map and the reader.
• Analyst Gateway supplies the Analysis component: data, analysis, model building, the twelve computation workbenches and connectors.
• Schema Studio supplies the live-systems screens of the Writing component: design, preview, gallery, provision, run.
• Centrix, the research-centre system, supplies the Programme and Supervision components: projects, papers, grants, conferences, researchers, insight, reports, the calendar, the supervision register and ethics.
• BenchStamp, the laboratory system, supplies samples and quality control.
• Publon Press supplies the course reader and the document canvas.

The consequence for a university is one platform, one sign-in, one database, one set of members across AutoScholar, ScholarCloud and Giants’ Shoulders. A partner who has learned one of the three has learned most of the others.

Deployment for a client

A university that runs a DSpace repository connects it to Giants’ Shoulders by its OAI-PMH address. The harvest runs in the background and is idempotent: a re-run resumes rather than re-embeds, and the institution's records stay where they are, on the institution's own server, with full texts resolved from it on demand. Several repositories, institutional and external, sit side by side in the engine's register, each in its own corpus, and access to each is granted per member or shared with everyone.

The demonstration centre runs on a hosted database and a hosted engine. A client deployment runs the same code against the client's own database and engine, in the client's jurisdiction. The hosted model for confidential material is a serverless inference account whose cost is in cents per session, metered per member and shown on the Usage screen.

State on 14 September 2026

Every screen shown above was walked as three people on the live address with no page errors. The demonstration centre was populated through the system's own verbs, never seeded: two DSpace repositories harvested (260 records, classified), four candidates' repositories with reviews at different levels, a personal library, a literature watch, a dataset with its fit, a designed schema, and an asked question. The ethics register on the demonstration centre is empty. The "who holds it" table shows member identifiers rather than names. A production deployment starts from the client's own repositories and members, so none of these rows are carried over.