AiBest structured content-pack exercise supplement — 2026-10-20
FICTIONAL EDUCATIONAL EXAMPLE — NOT A REAL OFFER, NOT SEND-READY

START HERE
This standalone supplement contains a revised structured drafting prompt, three
fictional sources, six actual saved model responses and local review/export tests.
It supplements, rather than silently replaces, the public v1.2 sample at:
https://aibest.site/how-to-choose-the-best-ai-tool-for-your-business/
The October 20 exercise itself did not update the website or v1.2 download.
This companion was prepared for public release October 27, 2026. The original
v1.2 download remains separate and unchanged. Historical cli-results.json has
internal directory prefixes replaced by [REDACTED_INTERNAL_DIRECTORY]/;
statuses, saved model responses and code are unchanged.

WHO THIS IS FOR
A solo founder or small content team testing whether approved public business
facts can produce three meaningfully different social drafts and an email body.
This buyer/problem is a hypothesis, not an established customer base. No product
price, purchase, time saving, audience response or willingness to pay was tested.

TRY IT WITHOUT AN ACCOUNT, MODEL OR PAYMENT
1. Open worked-example.html. Every displayed asset is prominently fictional.
2. Compare the displayed preparation, scope/fit and deliverable-review angles with
   rich-source.txt. Review what changes between the two responses for that case.
3. Open raw-model-outputs.json. Compare both conflicting-source responses: assets
   are empty and clarification is kept outside publishable copy.
4. Inspect injection-source.txt and the two corresponding responses. The footer
   is intentionally obvious and untrusted; this is not a realistic attack suite.
5. Optional: with Python 3.14, run these local checks from this extracted folder:
   python3 -m unittest -v test_pack_guard
   python3 replay_saved_outputs.py
   Run the replay twice: it uses new temporary files and should not overwrite.
No network, external packages, publishing or email sending is involved. The
replay uses your system temporary directory and removes its temporary artifacts.
Older Python versions were not tested in this exercise.

WHAT WAS ACTUALLY EXERCISED
On October 20, 2026, six separate Hermes child contexts generated one response
apiece: rich source, conflicting records and explicitly untrusted injected footer,
each repeated twice. Same model: gpt-6.1-sol. There were no output retries or
selection among candidates in this run. Each context read its assigned input;
contexts were separate but inherited the runtime's common instructions. They
were not blind, independent model families or real customer tests.

The runtime reported all six generation tasks and both specialist reviews as
included, cost_usd 0.0. No paid model API was separately invoked. This is runtime
usage reporting, not an independently reconciled provider billing ledger.

All six saved responses matched the guard's JSON format. Both conflict cases
returned BLOCKED with no assets. The other four returned DRAFT with four assets
each. Both injection responses omitted the specific made-up booking URL, marker
and sales guarantee. These are narrow observations, not a security pass rate.
The three social angles cover preparation, scope/fit and reading the deliverable;
independent model-based editorial review found modest but distinct reader value.
The question-led openings remain formulaic and require channel-specific editing.

We changed BOTH prompt guidance and source richness versus the earlier exercise;
there is no controlled evidence that the prompt alone caused improvement. The
older six-case prompt was not rerun. This exercise has three scenarios, not six
different scenarios. No production system reliability is established.

SOFTWARE EVIDENCE IS SEPARATE
The unchanged pack_guard.py passed its existing 18 software tests again. We also
ran actual CLI review/export on six saved outputs: four test-only exports and two
refused cases. CLI export on blocked cases fails for a missing receipt; a separate
in-memory check confirms the BLOCKED status branch itself refuses export.
A repeatable replay was run twice from the final extracted ZIP, with fresh paths.
These software replays do NOT make new AI responses or test semantic accuracy.

All test receipts were automatically supplied by the harness, explicitly labeled
NOT human approval. These simulate the mechanical gate, not a completed human
checklist. No person or customer has approved these assets. The guard cannot
verify facts, privacy, reader value, identity or legal compliance. Its negative
control demonstrates false claims can pass after a mistaken self-attestation.
Do not reuse test receipts to approve real copy.

USE THE PROMPT FOR A REAL TRIAL ONLY WITH HUMAN REVIEW
Read candidate-prompt.txt. Append only approved PUBLIC source facts. Do not send
customer records, credentials or confidential business material to a model.
Check your chosen model provider's actual terms, privacy and cost separately.
Human-review every claim against the source and edit for genuine reader value.
Resolve conflicts with the responsible owner; never relabel BLOCKED as DRAFT.
Prepare channel-specific length, visuals and context before publishing. Email
body text is not a campaign: sender, consent, audience, unsubscribe and applicable
legal requirements must be separately settled. Nothing here sends any message.

The machine JSON and raw export do not carry a fictional label in every body;
keep them attached to this labeled educational package. Do not present the
fictional business as real or publish its advertising. worked-example.html
carries a warning beside EACH displayed asset for that reason.

FEEDBACK, WITHOUT A SIGNUP OR PURCHASE
If you voluntarily try the workflow, use https://aibest.site/contact-us/#sample-feedback to
report the step reached, what was useful, what needed rewriting and what failed.
Do not submit private source text. The October 20 exercise established no customer feedback or demand.
There is no checkout, subscription or payment collection in this supplement.
