CARRIER R03 — WHAT SURVIVES THE HANDOFF?
5 October 2026 | Frozen Carrier: 6,257 checked records

Open the public lab's index.html or downloads/technical-report.html first.
The lab is static and requires no account, API key, worker or server database.

Three different forms of evidence
1. A frozen exploratory requirement reranker on public R02 data. This is not a
   fresh final holdout. It produced negative ranking results, retained here.
2. Actual first outputs from 16 fresh AI participant cohorts: eight per stage,
   eight distinct cases per cohort, four formats and two evidence conditions.
   Agents share the configured model family; no authenticated backend build.
   First memories and second receivers were actually generated. The six-field
   score is closed-world and does not assess every sentence of a memory.
3. Deterministic resolver checks, exact application-byte accounting and local
   synthetic lookup stress. These do not measure semantic intelligence.

Reproducibility-kit layout
128/ contains only the pinned R02 inputs listed in source-allowlist.json.
129/ contains this study's protocols, data, cases, raw inputs/outputs, scores,
     source code, figures, reports, original failures and design amendments.
Keep these two directories as siblings; the frozen displacement scripts use
relative paths. The kit is an explicit allowlist, not a working-folder dump.

From a newly extracted kit directory, with Python 3.12 or later:
  python 129/tools/verify_handoffs.py
  python 129/tools/handoffs.py evaluate
  python 129/displacement/test_integrity.py
  python 129/displacement/verify.py --encode
  python 129/scale/run_checks.py --corpus-dir 129/data --replace

The first two commands and reference/scale checks use the standard library.
Displacement reproduction needs the dependencies in
129/displacement/requirements.txt (NumPy, ONNX Runtime, tokenizers). The pinned
local model and tokenizer are bundled. No network is needed once dependencies
are installed. Figure regeneration needs Matplotlib 3.11.2 as well:
  python 129/tools/figures.py

Use a fresh extracted copy for reruns. evaluate rewrites deterministic scores
from preserved answers. The scale --replace option archives old results before
writing new timings. Timing values vary with the environment. Model float
results can differ across runtime versions/CPU implementations; the original
environment and frozen values are recorded.

This reproduces scoring, not new stochastic AI responses. To run a new agent
trial, supply one cohort input file to each fresh agent with the procedural
file-access restrictions, preserve first outputs under a new run directory,
then prepare stage two from those memories alone. Do not substitute new outputs
into the original run. Original authored cases and answers are now public and
must not be presented as a secret benchmark.

Design correction
The initial written stage-two allocation and its implementation differed;
the implementation also confounded receiving cohort with condition. An audit
caught this after stage-one deliveries but before any answer scoring or stage-
two generation. A fixed +3 cohort rotation restored all eight cells per cohort.
Original protocol/code/key and the audit are retained under 129/checks/.
No first-stage source, memory, answer or scoring rule changed.

Interpretation limits
No human receivers or graders participated. No claim of superintelligence,
general hallucination reduction, million-concept semantic performance or
Carrier-only advantage is warranted. Generic references receive matched
mechanics. Definitions are supplementary in the conversation cases. Wire byte
savings are distinct from model token/context savings. This experimental JSON
envelope is not a new standardized USL1/CVP2 transport profile.

Licenses and provenance
Existing public Carrier records and historical credits are preserved. Source
tasks and third-party model retain their bundled attribution/license files.
See LICENSE-ATTRIBUTION.txt. No private prior conversation is included.
