Every tie has a DOI. Every "no tie" has a scope.
TieTrace helps a journalist check how connected an expert is to the company in a story. It answers from a research graph built from OpenAlex and stored in FalkorDB, with evidence, not opinions.
1The problem
A story quotes a scientist. Was their work funded by the company the story is about, or did they list an affiliation with it? Reporters rarely have time to read a whole publication record. TieTrace reads it for them and shows each tie with the paper it rests on.
2The four steps
- Paste the story. Claude Code reads it and picks the research case for the company and topic.
- Names are resolved to OpenAlex author records. If a name matches more than one record, you pick; it never guesses.
- Ties are checked in the graph with read-only queries. The live evidence panel shows every query as it runs.
- The answer comes back per expert as "Expert #n", with DOIs for every tie and the scope for every "no tie found".
3What the tiers mean
| Tier | Meaning | How to use it |
|---|---|---|
| A1 | Authored a small paper (25 authors or fewer) funded by the company, or listed a company affiliation on one | Direct evidence, with DOIs |
| A2 | Foundation grant, bottler affiliation, or the only evidence is a large multi-author paper | Direct evidence, with DOIs and the reason |
| B | 2 or more recent small papers (since 2021) with a Tier A person | Network context only, not validated |
| C | Within 2-3 such co-author hops of a Tier A person | Network context only, not validated |
| NONE | Nothing above within 3 hops | "No tie found in the research record", always with its scope |
A1 and A2 are direct evidence. B and C are network context only: they describe who wrote papers with whom, and our own tests (below) showed they are not a sign of unusual closeness.
4Why "no tie found" is never bare
A missing tie can simply mean missing data. So every "no tie found" states:
- Hops: how far through co-authors it searched (up to 3).
- Years: which years of collaboration count as recent (since 2021).
- Funder coverage: how much of the data records funders at all, for the whole graph and for that expert's own works.
In the Coca-Cola graph only 56.3% of the 30,140 loaded topic works list a funder (96.6% list a funder or an affiliation, but affiliation says little about funding). In the Philip Morris graph it is 37.5% (7,482 of 19,964), so "no tie found" is weaker there. The topic part of the graph is a sample of the most-cited works per year, about 11.5% of the field.
5What we tested, honestly
Pass rules were fixed before the runs and not changed after.
| Test | Result |
|---|---|
| E3 known case (GEBN) | Passed for the four. 4 of 4 publicly confirmed GEBN leaders found as Tier A with DOIs. A fifth name could not be confirmed from a public source. |
| E1 placebo lift (Tier B) | FAIL. Lift 0.96x (interval 0.84-1.09) against works-count-matched random groups: Tier B is not a sign of unusual closeness. |
| E2 temporal hold-out | FAIL. The weighted score (AUC 0.716) does not beat publication volume (0.720) or a plain count of Tier A co-authors (0.749). |
| E6 mega-paper ablation | Without the 25-author cap A1 would grow from 949 to 1,069 people and Tier B from 1,309 to 1,889; 23.7% of A1+B people exist only through big papers. The cap is kept. |
| E8 golden tests | 15 of 15 pass. |
| E9 concurrency | Two server processes checking and ruling at the same time: nothing lost. |
| E10 tenant isolation | All 13 tools tried with another newsroom's ids: nothing visible. |
Because E1 and E2 failed, only Tier A is a finding. Tier B, Tier C and the score are shown as labelled network context, and experts are never ranked by score.
6Who decides
The model never records a verdict and never approves a sourcing note. On this page those are buttons: you click, read what will happen (for a note, the whole note), and confirm a second time. Excluding a paper re-runs every saved check in the newsroom and shows what changed; merging two split profiles is possible through the API only. A sourcing note can be approved only after the Verifier has checked every claim against the graph.
7Limits
- Funding data is thin. Many papers list no funder, so "no tie found" often means "no funding recorded".
- Split author profiles. OpenAlex sometimes splits one person into several records; a check on the wrong record can miss evidence. A human can merge two profiles.
- The graph is a sample. Co-authors' other works are not loaded, so ties through unloaded papers are invisible.
- Gifts to universities and other money that OpenAlex funder fields do not record are invisible.
- Profile merge and split error rate is unknown. The hand check (E5) was cut.
8For judges
- The MCP server has 13 tools and makes no LLM calls. Claude Code is the agent and the MCP client; this page only shows its steps.
- Every model-reachable read uses
GRAPH.RO_QUERYwith a timeout and a row cap; writes are fixed, parameterised queries. Every tool returns the Cypher it ran, and the live evidence panel shows it. - In the web app the model gets the read tools and the memory-writing tools (
open_case,check_case,recheck,draft_note, which save cases, checks and draft notes), but neverrecord_verdictorapprove_note: those run only from a confirmed click on the page. - Code and contract: the GitHub repository and docs/mcp-spec.md.