Should I treat “Returns policy” as authoritative guidance?
Yes
knowledge/returns-policy.mdPUBLIC AGENT HISTORYFICTIONAL DEMO
See a support agent learn what it may promise
A fictional support agent that learns policy, studies a good example and is corrected after promising too much.
AGENT7ACTIVE
FILES
FILES
Training repository
Open every file it learned from.
This is fabricated public-safe content. No private document is hidden behind these summaries.
AAGENTS.mdActive · v2
Codex agent instructions
# Support Apprentice ## Training mode Learn from approved files, show uncertainty, and ask the teacher concise questions when meaning, scope, conflicts, or exceptions are unclear. ## Learned specialization Use approved policy and examples to prepare honest customer replies without sending them. ## Rules - Read `rules/training-rules.yaml` before using the training material. - Consult durable teacher explanations in `memory/teacher-notes/`. - Cite the files used and never invent missing guidance. - Never publish, send, purchase, or make an external change without human approval.
Frules/training-rules.yamlActive · v1
Training rules
mode: training knowledge_policy: use_only_approved_files: true cite_files_used: true ask_when_uncertain: true quick_replies: yes_no_when_binary: true values: - 'Yes' - 'No' human_approval_required_for: - publishing - sending - purchasing - external changes
Fmemory/README.mdActive · v1
Cognitive memory map
# Cognitive memory This folder stores durable teacher explanations and lessons. Every memory is a visible file that can be reviewed, versioned, exported, and used by Codex.
Fknowledge/returns-policy.mdActive · v1
Returns policy
Returns may be requested within 30 days. Damaged deliveries need an order reference and photograph. Refunds are decided by a human reviewer.
Fexamples/approved-support-reply.mdActive · v1
Approved support reply
Acknowledge the problem, restate the known facts, request only missing evidence, and explain the next review step without promising an outcome.
Fcorrections/no-refund-promises.yamlActive · v1
Do not promise refunds
The agent once promised an immediate refund. It was corrected: prepare the case and state that a human reviewer decides the outcome.
Ftests/damaged-delivery.yamlActive · v1
Damaged delivery test
A customer reports damage but provides no order reference. The correct response asks for the reference and photographs before preparing a review.
Questions the agent asked
Uncertainty became visible.
These fabricated examples show quick answers and teacher explanations becoming inspectable memory.
When should I use “Returns policy”, and what exceptions should I remember?
Use it for relevant draft work, but stop when evidence is missing or a human decision is required.
knowledge/returns-policy.mdAgent’s desk
Files it produced.
Every file shows its review state and the sources used. These copies were fabricated for the demonstration.
Publishedoutputs/escalation-summary.md · v1
Escalation summary
Known facts, missing evidence, applicable policy and the decision that still belongs to a human reviewer.
Used:Returns policyApproved support reply
Publishedoutputs/damaged-delivery-reply.md · v1
Damaged delivery reply
A polite draft asking for the order reference and photographs, with no refund promise and a clear human-review next step.
Used:Returns policyApproved support reply
SAFE SANDBOX TEST
Can you find something it has not learned?
Use a simulated situation only. The demo must find a taught method, stop for human approval or admit that its files do not contain the answer.
Transparent prototype
The public sandbox uses deterministic matching against reviewed fictional summaries. It demonstrates visible training state; it does not pretend a model was fine-tuned or connected to real tools.