Test the range
The 18 rows span IR theory, systems and methods, semiotics, Eurasian history, policy documents, journalism, noise, and unreadable material.
IR pilot case study
How Knowledge Prism tested a live International Relations corpus.
The pilot was deliberately small. It tested whether the system could distinguish supported claims, irrelevant-but-accurate files, misfiled noise, weak samples, unreadable text, and works that matter because of their function in IR explanation.
Why the pilot was run
The system had to prove discipline before scale.
Test the boundary
The pilot asks two questions separately: is the claim textually supported, and does the work belong to this project?
Test restraint
No row was promoted to a final graph claim. That restraint is the point: evidence discipline beats display.
What the 18 works tested
The pilot was a stress test of judgment, not a race for impressive numbers.
Core scholarly works
IR theory, conflict studies, ideology, metatheory, semiotics, cartography, realism, discourse analysis, and nationalism tested whether sampled text could support real research use.
Boundary cases
Histology, Afghan reportage, Ratzel, Russia/Mongols, and the Advaita Vedanta file tested whether the system could keep accurate, promising, peripheral, and unreadable material apart.
Noise and failure cases
Practical Exam Format and abiword_job show that a trustworthy system must recognize misfiles and unreadable extraction instead of pretending every file is evidence.
Why 11 rows were promoted only to sample-level support
The sampled text supported the claim and the works fit the project boundary. That makes them candidates for later verification, not final ontology truth. The system can use them for review and interpretation while keeping the stronger gates closed.
Why zero final graph claims is a strength
The pilot proves the machine refuses premature certainty. A public dashboard with zero final claims is more trustworthy than one that turns first-pass samples into settled ontology.
Why Histology was excluded despite being accurate
The sampled text supported a claim about a histology atlas, but the work does not serve the International Relations research boundary. Accuracy alone is not enough.
Why Practical Exam Format is misfile/noise
The file is an administrative exam artifact, not scholarship. It is excluded as a corpus-cleaning case, not treated as a serious contradiction of a scholarly thesis.
Why Ratzel and Russia/Mongols need deeper re-sampling
Both may matter for geopolitical genealogy or Eurasian historical explanation, but their current samples are partial. The system keeps them visible and asks for better evidence.
Why abiword_job needs OCR or manual inspection
Unreadable is not the same as irrelevant. The title suggests possible non-Western ontology or IR knowledge-construction value, but no functional claim should be trusted until the text is recoverable.
What the pilot teaches
Trustworthy AI-assisted research is allowed to slow down.
AI can help find and organize.
It can surface candidate claims, collect review material, and make a knowledge map easier to inspect.
The scholar keeps judgment.
The system separates source relevance, claim support, and functional interpretation so the human decision is visible instead of hidden.
Refusal is a feature.
The strongest result is not that 11 rows were supported. It is that the system also held back, queued rereads, and refused final graph claims.
Sample-supported rows
The 11 candidate works and their IR functions
These are sample-supported candidates only. They are not final accepted concepts or final graph claims.