CARRIER RETRIEVAL LAB R01 — REPRODUCIBLE EXPERIMENT

RESULT
This experiment did not demonstrate a reliable graph-assisted retrieval gain.
The full first-run rankings, controls, frozen plan and failed-check history are
retained. Do not edit them to obtain a more favourable result.

READ
analysis/technical-report.html — methods, numerical findings and limitations.
protocol.json — original locally prospective plan; not external registration.
analysis/methods-review.json — pre-execution correction of limited blinding.
analysis/dense-baseline-plan.json / reader-pilot-plan.json — exploratory additions.

VERIFY AND REPRODUCE
1. Python 3.12 and NumPy 2.3.5 reproduce the recorded core environment.
   Install dependencies in an environment of your choice, then work offline.
   A NumPy-compatible later Python may work, but is not the recorded environment.
2. python tools/verify_kit.py
3. python tools/test_retrieval_independent.py
4. python tools/audit_results.py
5. python tools/reproduce.py

Reproduction creates a separate reproduction/run-* folder and recomputes the
development selection and held-out results from the supplied inputs. The first
run remains unchanged. All 360 authored questions are already fixed; this does
not regenerate them with a new AI model. The public test set is no longer secret.

DENSE SUPPLEMENT
tools/audit_supplements.py recalculates rankings from supplied normalized vectors,
and checks the reader's exact references/quotes. This does not independently
re-run or validate the transformer itself.

The full experiment kit includes the pinned model/ directory. A separate
model.zip is also supplied for readers who want just those assets. To recompute
model vectors, use ONNX Runtime 1.30.0, tokenizers 0.23.2 and NumPy 2.5.3 (the
recorded dense environment). Work in a copied kit to preserve original outputs,
then run python tools/run_dense.py. No remote inference API is used. A new
platform may change last-bit floating values; compare ranks and tolerances.
The local acquisition script pins a repository revision in model/manifest.json;
do not redownload an unpinned latest version and call it the identical model.

FIGURES
Matplotlib 3.11.2 generated the SVG, PNG and PDF charts. The figures are included.
python tools/plot_results.py regenerates them from the results. Plotting does not
re-run the experiment. Dependency acquisition may require internet once;
core scoring and audits use local supplied files after dependencies are installed.

READER CHALLENGE
The separate reader-challenge.zip has only masked contexts and response rules.
The first reader's responses are retained here. A fresh reader should disclose
earlier context, tools and any access to published answers. A repeated trial is
not automatically blinded merely because method labels are hidden.

CLAIMS
Queries were AI-authored from target definitions; other valid concepts are
unjudged. Scores concern one intended item, not exhaustive semantic relevance.
The graph is contextual, not entailment. Dense results and the single-reader
pilot are exploratory. No live Carrier state was changed by this study.

RIGHTS
Carrier/source records retain their existing provenance and source terms.
Wikipedia adaptations remain CC BY-SA 4.0; see LICENSE-ATTRIBUTION.txt.
Model source and Apache 2.0 declaration remain with the pinned model card.
