PORTFOLIO EVIDENCE — 16 OCTOBER 2026

These artifacts contain synthetic examples, not customer data or live banking/news results.

NEWS DIGEST
news-digest.html and news-digest.txt were rendered by the news project's existing
digest_to_html function and EmailDigestResponse.to_markdown method with two synthetic
entries. Editorial text and ranking scores were supplied by the reproduction script.
Links point to pinned public source files; they are not claims of collected news.
Assertions checked that both entries and their source links appear in both outputs.

The exact formatter/model definitions were compiled from the local source AST without
importing provider setup or loading .env. Python Markdown 3.7 performed the formatter's
Markdown conversion; Pydantic 2 constructed the original digest model. No replacement
formatter or invented application screenshot was used.

The reviewed news checkout is at 0608f9a99dec2308ae9dcdebce840e5ce958445e. Its local
provider integration has edits; the extracted digest model/to_markdown methods and
app/services/email.py formatter match the linked committed methods. No provider code
was executed. This is formatter evidence only: no scraper, model, database, ranking,
scheduler, or SMTP delivery was run.

BANKING AGENT HARNESS
banking-harness.json records actual calls to the unchanged local banking tools and
services at 2b8b094bc80bef18f75b1e50c5235c77d5a246c9. The reproduction loader bypasses
ADK/provider startup and .env loading, provides a lightweight ToolContext fixture,
and sets the approval lifetime to the code default (24 hours). The executor outcome
is explicitly set to the project's mock SUCCESS mode. Policy, mock accounts, pending
state, limits, audit, transfer tools, and executor functions are the source modules.

Executed assertions:
- Valid proposal leaves mock balance at 200000 before confirmation.
- Later explicit confirmation succeeds and changes mock balance to 199000.
- Repeat confirmation returns NO_PENDING_TRANSFER.
- Blocked beneficiary is rejected with mock balance unchanged at 200000.
- Same invocation cannot confirm its own proposal.
- A separate session with separate state has no pending transfer to approve.

Names and account suffixes in the transcript come from the project's existing mock
fixtures. They are synthetic. The sample proves these deterministic function paths;
it does not prove a model conversation, live banking, ADK browser operation, Redis
integration, durable state, or production concurrency.

OTHER CASE STUDIES
UGC, LendFlow, and DevGraph records were checked against their implementation source.
Their case-study verification labels explicitly say source review. Such inspection
is not a successful provider call, live database transaction, public deployment check,
benchmark, or security certification. Ownership descriptions follow the portfolio's
reported scope; repository inspection alone does not independently audit authorship.
Decision explanations are implementation analysis, not quotations from author interviews.

REPRODUCTION
Source script: docs/evidence/reproduce.py in the portfolio repository.
Dependencies: Python 3.11+, Pydantic 2, Markdown 3.7.
Pass --news-source and --harness-source paths to local checkouts. The optional
--extra-python-path supports an isolated dependency directory. No source project is
modified. Only the portfolio's public/evidence output files are written.
source-hashes.json records SHA-256 digests of the exact local files read for execution.