# Alien / OpenAIRE MCP pagination diagnostic

**Executed:** 18 July 2026  
**Routes:** Official Alien Intelligence OpenAIRE demo and OpenAIRE Graph API

**Model shown by the demo:** Sonnet 4.6

**Purpose:** Separate a genuine evidence null from a recoverable MCP integration defect.

**Stable submission diagnostic:** https://atlas.tsuchiyalab.com/reproducibility/openaire-mcp-crosscheck.md

## Replay prompt

The public reviewer prompt that bounded the run was:

> Using the OpenAIRE Graph only, cross-check DOI 10.1088/1367-2630/ad5b13. Return identity, funding, projects, organisations, datasets, software and provenance. Treat missing relations as not observable.

This prompt fixes the record, requested entity types and missing-edge semantics. The deterministic API audit—not generated prose—remains the source of every aggregate rate.

## Initial cross-check

The authenticated Alien/OpenAIRE run used 30 OpenAIRE MCP calls. It directly established:

- DOI `10.1088/1367-2630/ad5b13` resolves to `doi_dedup___::82d8842e25b2e9bdbe03ab4c5db972db`.
- The record is a 2024 publication.
- DOI plus `rel_project_code=101102140` returns exactly the target record, confirming EC grant **QIA-Phase1**.
- `get_research_links(target_type=dataset)` reported `total: 1` but returned no rows on pages 1 and 2.
- `get_research_links(target_type=software)` reported `total: 2` but returned no rows on pages 1 and 2.

The last two observations initially looked like a pagination boundary. They were then isolated in a controlled comparison.

## Root cause

The official OpenAPI definition for `GET /v1/researchProducts/links` gives `page` a default of `0`. Direct calls confirm that this operation is zero-indexed for the tested record:

| Target type | Page | `totalLinks` | Rows returned |
| --- | ---: | ---: | ---: |
| dataset | 0 | 1 | 1 |
| dataset | 1 | 1 | 0 |
| dataset | 2 | 1 | 0 |
| software | 0 | 2 | 2 |
| software | 1 | 2 | 0 |
| software | 2 | 2 | 0 |

Two targeted Alien MCP calls then explicitly requested `page=0`. Both failed validation before an upstream URL was produced. The demo reported that `ResearchLinksInput` requires `page` to be greater than or equal to `1`.

This is an **off-by-one contract mismatch in the MCP wrapper**:

- upstream link operation: zero-indexed, default page 0;
- Alien MCP input model: page 1 or greater;
- effect for result sets smaller than one page: the only data page is skipped while the total count remains visible.

It is not an OpenAIRE data-null and not an upstream pagination failure.

## Rows recovered at page 0

### Dataset

- **Title:** Quantum-limited measurements of optical signals from a geostationary satellite
- **Type:** dataset / Collection
- **DOI:** `10.6084/m9.figshare.c.3813670`
- **Relation:** `references`
- **Provenance:** OpenAIRE

### Software

1. **netsquid-freespace software on GitHub**

   OpenAIRE ID: `openaire____::0203e43bc9da9dd3318eede5cd1e5544`
2. **quantumcity software on GitHub**

   OpenAIRE ID: `openaire____::4f63b22fef29eff8bc0dc768945bfe9c`

Both software rows are typed as software, carry a `cites` relation, have OpenAIRE provenance, and were collected from GitHub and Software Heritage.

## Corrective action

Either server-side repair is sufficient:

1. define the MCP `page` field with a default of `0` and minimum of `0`; or
2. retain a one-indexed MCP interface and send `page - 1` to this upstream link operation.

A regression test should request a known one-row dataset relation and verify that the first MCP page returns the row rather than only `totalLinks`.

Until the MCP wrapper is repaired, the Atlas uses the direct link endpoint with `page=0` as its reproducible fallback. A non-zero total with an empty result page is classified as a tool inconsistency, not as “no relation.”

## Live recheck

An uncached bounded recheck on 29 July 2026 reconfirmed the representative contract: the featured DOI and grant resolved, page 0 returned one dataset and two software rows, and page 1 returned no rows despite retaining the same non-zero totals. All eight discovery-query counts also matched the full refresh, while the four Q‑NEKO aliases remained at zero projects and zero research products. The machine-readable result and source are published as `live-recheck.json` and `openaire_live_recheck.py`.

## Reproduction

Run `openaire_mcp_pagination_repro.py` in this directory. It queries page 0 and page 1 for both target types, prints the returned titles and identifiers, and marks whether the mismatch is reproduced.

- Official OpenAPI: `https://api.openaire.eu/graph/v3/api-docs`
- Swagger UI: `https://api.openaire.eu/graph/swagger-ui/index.html`
- Alien demo: `https://demo.alien.club/openaire`
- Hackathon call: `https://innovation.openaire.eu/component/content/article/openaire-ai-hackathon.html?catid=8`