A checked snapshot
Only names and definitions are indexed. Names appear twice. Source articles, relation labels, provenance notes and IDs contribute no lexical terms.
Retrieval study / Frozen corpus · held-out queries
Does context help
us find the idea?
A lexical baseline. A graph of contextual relations. One measured question: does following those relations improve retrieval of an intended concept?
See the evidence Try a query01 / Results
Let the result stand.
The primary comparison was specified before scoring. A null or negative result remains part of the record.
Methods, on the same test set
Family-mean scores. Higher means a better rank for the assigned intended target.
Where the change lands
Compare the graph with the selected lexical method at the family level.
Every predefined method
Held-out only| Method | MRR@10 | Hit@1 | Hit@5 | Hit@10 | Hit@20 | Discounted rank@10 | MRR (all ranks) |
|---|
Each query has one assigned intended target. Other relevant concepts are unjudged. “Hit” means that intended target appeared; these are not precision or exhaustive relevance estimates. The five rewired runs are controls on one dataset, not five independent datasets.
Explore query kinds, historical bands and graph structure
| Group | Lexical MRR@10 | Graph MRR@10 | Difference |
|---|
Subgroup results are exploratory. ID bands balance position in the historical sequence; they are not subject domains.
A contextual graph with uneven degree
Application and boundary questions are retrieval probes. These measurements do not establish comprehension, reasoning, transfer or universal superiority of one search approach.
02 / Try search
Put a question
through both.
The complete index and graph are in this page. Search runs on your device using the locked parameters, with no query sent to a server.
Start with
The starting prompts are illustrative examples, separate from the frozen benchmark.
Enter a question to compare the first ten results.
Lexical baseline
With context graph
Scores are meaningful within each ranking. Their numerical scales differ across methods. Expanding a result shows its full definition and linked context; a neighboring concept is not necessarily relevant to your question.
03 / Query explorer
Inspect the questions.
Look past the average.
All held-out probes and their recorded rankings. The assigned target is hidden until you reveal it, so you can inspect the outputs first.
Synthetic questions. Partial gold. Two AI author processes wrote from target names and definitions. An assigned target is one intended answer, not a judgment that every other result is irrelevant. No independent human relevance pool was collected.
Lexical baseline
With context graph
These are the frozen Python evaluation rankings. Live search independently applies the same exported weights and graph recipe in JavaScript. Gold labels are inspectable in downloads and embedded data; the reveal control is a reading aid, not formal blinding.
04 / Method
A narrow claim.
An open procedure.
No relation edits, target changes or parameter adjustments in response to held-out scores. Every method answers to the same checked snapshot and frozen questions.
Families stay together
120 intended targets sampled across six sequential ID bands. Each gets a paraphrase, an application and a boundary question. Forty families tune parameters; eighty are held out.
Authors see assigned names and definitions. They share a project and model family. This is AI-authored known-target retrieval, not independent human relevance assessment.
The same lexical start
BM25, cosine TF-IDF and reciprocal rank fusion compete on development MRR@10. The chosen lexical method seeds a twenty-step contextual walk from its top twenty positive results.
Resample the family
Three variants are averaged within each family. The primary contrast is graph minus selected lexical MRR@10 over the eighty held-out families.
A paired percentile bootstrap uses 10,000 resamples; a two-sided paired sign-flip test uses 10,000 draws and a plus-one correction. Secondary analyses are exploratory.
Scoring, controls and limits
Words first
Lowercase ASCII alphanumeric tokens of at least two characters, a fixed stop list and no stemming. BM25 uses k₁ = 1.2, b = 0.75 and unique query terms. TF-IDF uses logarithmic term frequency and cosine normalization. Fusion sums 1 / (60 + rank) over positive-scoring BM25 and TF-IDF results.
Context as a controlled intervention
Checked links become deduplicated undirected pairs; self loops are excluded. The graph carries context, not a semantic predicate. Seeds are normalized lexical scores. Every walk step mixes seed restart with neighbor propagation; mass at disconnected nodes returns to the seeds. The final score mixes max-normalized lexical and walk scores.
Six nonzero graph settings compete on development families. The graph cannot silently select the unchanged baseline. Five degree-preserving rewired graphs use the same locked settings. Degree-only, empty-edge and stored-direction controls are reported. Directed results are exploratory.
What a zero means
Only positive scores are ranked; ties break by ascending checked ID. An absent intended target has no rank and receives zero. With no lexical evidence there is no graph seed and no fabricated result from zero-score ties.
The boundary of the evidence
MRR@10 gives a target at rank r a score of 1/r if r ≤ 10 and zero otherwise. Discounted rank uses one assigned target, not comprehensive relevance judgments. Confidence intervals cover the sampled synthetic families; they do not establish behavior for all users. The confirmatory comparison evaluates the specified lexical and graph methods; it does not measure comprehension.
The protocol is locally prospective and frozen before scoring. It is not externally registered or externally timestamped. Author blinding was limited: target IDs, historical bands and ordered family keys were visible. The split was not explicitly supplied, but could be inferred. A prospective addendum corrects the original stronger claim. Hiding a label in the explorer is not external or human blinding.
From frozen corpus to held-out comparison
Query families, the development lock and the confirmatory comparison remain separate stages.
Verifiable record
Snapshot and parameter lock
Methodological sources
05 / Downloads
Take the evidence
with you.
The questions, frozen procedure, results and reproduction materials belong beside the conclusion. The page runs locally without external libraries; reproducing the experiments requires the documented scientific and inference dependencies.
ZIP / Reproduction kit
The experiment kit
Corpus, queries, code, recorded outputs and pinned model assets. Dependency instructions are included; the model is also available separately.
HTML / ReportTechnical report
Methods, controls, diagnostics and interpretation.
JSON / ProtocolThe frozen plan
Sampling, baselines, primary contrast and claim boundaries.
JSON / BenchmarkEvery query
Development and test families, assigned targets and author notes.
JSON / ResultsThe full results
All methods, paired estimates, subgroup results and controls.
ZIP / Reader challengeTry a fresh reader
Thirty-two masked contexts for a fresh AI reader. Gold answers are excluded from this archive.
ZIP / Optional model · ~83 MBThe pinned ONNX model
For reproducing the exploratory dense comparison. Live search on this page does not load or run this model.