Practical AI Radar — Radar Scout (editorial pen name)
READER EVIDENCE EXPORT | Assembled 17 October 2026 with AI assistance
Local publication candidate. No publication decision is implied by this export.

START WITH index.html
Open it locally to follow the evidence map without installing a website.
It loads no remote scripts, styles, analytics or images. There are 53 unchanged
historical files: Scorecard inputs/results and two distinct logging fixtures.
All experiments were recorded on 12 October 2026. Nothing was rerun to make this
export. The mock records deliberately retain fabricated canaries, not user data.

OPTIONAL FILE CONSISTENCY CHECK
With Node installed, run from this directory:
  node check-files.mjs
This checks the file list, sizes, SHA-256 hashes and source-copy map. It does not
run the original semantic validators, re-evaluate assertions, execute experiments
or make network requests. Recorded passed=true fields remain historical claims;
this checker does not establish that their assertions are true.

The complete historical B package is retained separately for internal semantic
regressions. This export omits three editorial drafts, three internal guides,
three full-set manifests and four full-set validators. Original files are not
silently edited or supplied under weakened historical manifests. SOURCE_MAP.json
lists the selection and omissions. This is not an equivalent replacement for the
B verifier, and the complete B package is not included here.

HOW TO INSPECT THE CLAIMS
Scorecard: compare the seven stdout JSON files with their corresponding workflow
fixtures and hypotheses.json. run-1/run.json summarizes them and records exit
codes. A separate eighth invocation was unsupported; its stdout is empty and its
stderr is retained. version.txt is the original 653 bytes, ending with two LF
bytes. The later preservation plan is explicitly dated after the original audit.

Logging C: fixtures.json states predictions; run-1/operational.ndjson contains
eight operational events. provider-observations.json contains nine mock-service
observations, including a separate faulty-timeout control. header-only-control.json
records that control. A mock response prepared for sending is not proof of receipt.

Logging D: three separate events and server observations cover a format-valid
identifier, an opaque 500 response and connection closure before headers. C and D
use different schemas. Do not combine them as eleven real model evaluations.
Client and service views are author-controlled, not independent corroboration.

INSPECTION IS NOT EXECUTION
The preserved scripts describe how the historical fixtures worked. The checksum
command above does not run them. Scorecard requires a separately acquired Linux
binary and the original WSL context. Logging was recorded on Windows/Node 24.13.0.
Re-running on another environment creates a new experiment. This export is not an
installer and makes no cross-platform reproduction claim. Never use its raw
capture harness unchanged on real user traffic.

LIMITS
No signature, trusted timestamp, binary authentication, attestation or independent
reproduction is established. Coordinated edits to files and unsigned hashes can
pass the check. Upstream checksum and checkout-reference authentication remain
independently unverified. No production service, model provider, retention/access
system, retry implementation or restricted diagnostic channel was tested.
Neither high scanner scores nor these logging fixtures constitute certification.
