CARRIER RETRIEVAL LAB R02 — QUESTIONS FIRST. ANSWERS ON TRIAL.

READ FIRST
analysis/technical-report.html explains outcomes and limitations.
protocol.json is the locally prospective plan, not external preregistration.
The frozen corpus contains 6,257 checked concepts; no Carrier entry was changed.

WHAT IS HERE
48 frozen tasks: 24 adaptations of attributed public human questions and 24
prospective AI scenarios. Their writers did not inspect Carrier targets.
666 pooled candidate-task pairs, two relevance assessments per pair, 146
qualified research-overlay relations, fixed retrieval candidates, 72 reader
answers and two assessments per answer. All first complete records are retained.
Original public question wording, adaptations and obligations remain distinct.

COMPUTATIONAL REPRODUCTION
Use Python 3.12 and NumPy 2.3.5 for the recorded core environment. Install
requirements-core.txt in an environment of your choice before going offline.
From the extracted kit directory:

  python tools/verify_kit.py
  python tools/test_r02_engine.py
  python tools/audit_results_independent.py --audit-final
  python tools/reproduce.py

The last command creates a new reproduction/run-* directory. It recomputes the
pool, development selection, final metrics, randomized reader allocation and
answer scores using archived judgments and responses. It does not regenerate
those judgments with a model or establish independent semantic replication.
Original outputs are preserved. Never tune methods against the disclosed final
set and continue calling it a secret holdout.

The reproduction runner deliberately requires the recorded ranking file bytes
and normalized numerical outputs to match. Different CPU/BLAS builds can alter
floating-point scores even with the same Python package versions. A mismatch
must be reported and inspected; this kit does not claim cross-platform bitwise
identity or silently relax a failed comparison. The independent outcome audit
checks the published scores from the preserved rankings and grades separately.

TYPED OVERLAY
  python tools/build_typed_overlay.py --check

This uses the bundled byte-identical corpus mirror. Original 127/data paths in
provenance identify the source-only curator's input; no sibling 127 directory is
required to rebuild. The overlay is a model-reviewed research artifact. It is
not a promoted ontology, complete graph, expert truth certificate or inference
permission. Every relation retains its scope and source/target definitions.

EMBEDDINGS
Document vectors are the R01 frozen document representations only; R01 queries
and targets are excluded. New query vectors were computed locally using the
pinned all-MiniLM-L6-v2 ONNX model. The full kit includes model/ and its manifest.
Optional fresh inference uses requirements-dense.txt: NumPy 2.5.3, ONNX Runtime
1.30.0 and tokenizers 0.23.2. No remote inference service is needed.

To recompute query vectors, work in a COPY of this kit. Preserve the original
analysis/query-vectors.npz under another name before running tools/encode_tasks.py.
That script refuses to overwrite an existing vector file. Compare the new vectors
with the retained originals before claiming equivalence; platform floating-point
differences can exist. Task and model hashes must still match their pins.

FIGURES
The included SVG/PNG/PDF figures are derived from results. To regenerate them,
install requirements-figures.txt and run python tools/plot_results.py.
This regenerates charts, not the experiment or model judgments.

FRESH READERS
reader-challenge.zip contains masked reader contexts and instructions without
method/relevance/answer keys. A new reader must disclose prior knowledge of this
now-public benchmark. Keep its first response before grading. Fresh agent
contexts do not establish independent humans or remove pretraining overlap.

PUBLICATION
The outer publication kit includes PUBLISH.txt and WordPress blocks. The main
scientific kit is for inspection/reproduction, not for copying into the live
Carrier state directory. Preserve all source licenses and attribution notices.
