# Portfolio II — learner evidence and defense

This is a blank evidence template, not a completion certificate. The shipped
reference bundle is invented and assisted. Do not mark this project complete
because reference tests, a quiz or browser preparation passed.

## What you built

- Your package/CLI source location and tested revision:
- Your own model construction/ownership tests and justified value objects:
- Distinct signals/findings and a refused cross-layer shortcut:
- Two rubrics/strategies, hand-calculated totals and missing-review behavior:
- Changed CSV fixture you authored, predicted counts/lineage and retained original:
- NEW output folder with sample JSONL, conversion/quarantine, audit, mention
  report, derived data card and manifest:
- Actual import/validate/audit commands, exit statuses and test transcript:
- A failed attempt, root cause and repair before the final run:
- One deliberately changed/missing owned artifact and the provenance observation:
- What remains unresolved in rights, privacy, authorship and judgment quality:

## Assistance statement

List each reused Module 6/7 snapshot, value class, validator, strategy, I/O
helper, reference snippet, tutor suggestion or outside source. Distinguish
your implementation from reviewed support. Wrapper/orchestration work is not
independent implementation of the helper. Do not send private originals,
quarantine, credentials or real-media identifying data to the tutor.

## Data card review

Describe intended use, sources and ownership, selection process, declared
label semantics, unresolved cases, extra-field losses, normalization/masking
limits, known bias/representativeness gaps, privacy review and prohibited uses.
The generated classroom card is a starting point; it does not perform these
reviews or make the data eligible for training. Keep provenance separate from
authenticity and preserve ties rather than forcing winners.

## Observed defense and transfer

Record the date, reviewer/self-review context and actual demonstration. A
self-recorded walkthrough is self-report, not independently verified review.

1. Explain why frozen/type-annotated alone is insufficient and demonstrate a
   construction refusal without modifying the original fixture.
2. Retain a machine/human disagreement without fabricating a reviewer verdict.
3. Add one criterion through the protocol, predict the weighted result and
   show an asset with missing review remaining unscored.
4. Import a changed CSV live, reconcile every row, identify a named loss,
   preserve a tie and inspect retained source/quarantine locally.
5. Change an owned output copy, observe provenance failure, and explain why
   matching hashes still would not prove quality, rights or privacy.
6. Explain completed status 1 versus fatal 2 and a possible partial new output;
   never delete/overwrite a learner folder to make a demonstration appear clean.

Your evidence status (not started / attempted / assisted / independently
implemented / defended) and the specific evidence supporting it:

Remaining gaps and next action: