# Authored local PoC-quality teaching contract (not project approval)

This is a new fictional exercise, not execution of the old page's illustration,
not a customer incident and not output from an AI generation run. It corrects
an authored PoC/claim and quality record, not an already-correct product guard.

## Interface, incoming conditions and effects

`NoteService.submit_note(role, text)` is a direct synchronous Python method.
For this exercise the caller supplies exact `str` role values `viewer` or
`editor` and exact nonempty ASCII text values in cases.json. These are caller
preconditions, not a tested validator. A role string is not an authenticated
identity. This method is not a server, HTTP protocol or deployed entry point.

Each attempt starts from a fresh object with the complete literal note list:
`[{"note_id": 1, "by": "seed", "text": "UNRELATED"}]`.
The unprivileged-role viewer request uses text `ATTEMPT`. It must return
`{"status":"denied","note_id":null}` and preserve the complete list.
The intended-use editor control uses `CONTROL`. It must return
`{"status":"created","note_id":2}` and append exactly
`{"note_id":2,"by":"editor","text":"CONTROL"}`, preserving the seed.
IDs are deterministic local list allocation in this fresh state, not a global
business uniqueness promise. Every field, key, list order and value type is
part of the literal observation oracle; boolean/int equality is insufficient.

`_append_note` is intentionally an internal unconditional writer. The ordinary
public method calls it only for editor. Invoking that helper directly as viewer
starts after the guard. Such a call can show the helper effect, but it cannot
establish that viewer traversed the public guard. No defect in that unchanged
guard is claimed from the helper call.

## Three two-attempt groups

1. `weak_helper`: the authored weak PoC directly calls `_append_note` for
   viewer, then calls the public method for the editor control.
2. `corrected_public`: the corrected PoC calls the public method for both.
   A denied viewer result with unchanged notes means the claimed local
   public-method exploit is not reproduced; narrow the claim rather than call
   this an exploit success. No product implementation repair was performed.
3. `always_deny_control`: a deliberately seeded wrong service denies both
   roles under the corrected PoC. Its editor failure rejects a blanket-denial
   no-op as a sufficient PoC/control repair. It is not a discovered product bug.

There are six selected attempts across two input conditions. Each has its own
fresh state. The strong oracle independently fixes public method entry, actor
argument, exact reply and complete ordered notes. Profiler records capture
actual top-level and nested method/helper entries and returns; their count is
reported separately, not equated with six attempts. The weak impact observation
only reports whether the viewer call returned created; it is not an adequate
reachable-attack verdict. No extra calls are used to compute it.

## Separate native record snapshots

Each isolated project declares one active low-grade teaching KnownIssue.
Default medium severity threshold and video policy off are explicit. S0 has
internal-only disposition and no review; expected native PASS has zero in-scope
issues. S1 changes disposition to submit intent and supplies all eight general
booleans true, while the three Web3-only booleans are false. This is an
intentionally unsupported record, not honest proof: expected PASS reads those
answers and skips the three Web3-only rules in this non-Web3 fixture.
S2 keeps submit intent but honestly records real_infrastructure,
current_deployment_relevance and reachability_verified false; expected WARN
retains those three gaps. A local public-method denial does not establish a
real protocol/deployed-system exploit. No report is actually sent externally.
No N/A declaration or companion exemption is used.

Exactly these three selected native audits are planned with --no-cache,
--format json and --fail-level warn. They do not execute the Python PoC or
inspect the adequacy of these authored booleans. Native row status/severity,
advisory bit, aggregate warning count and process exit are separate units.
Other severity/config/active/N/A/security/video/cache/error paths are inspected
source, not extra executed cases. Native 11 keys differ from the catalog's
10-item prose; no catalog repair or obligation completion is claimed.

## Observation and exclusions

Full in-memory note lists and returned objects are observed immediately after
each synchronous call, with exact code-object call matching and source hashes.
This is not process restart/disk persistence, isolation, networking, HTTP,
identity authentication, concurrency, real deployment, economic impact,
severity validation, real submission approval or every possible input. A native
record PASS is not PoC execution or authorized reliance. Matching source hashes
bind local identities, not authenticity or adequate/full dependency coverage.
