Unit 06.04: What determinism buys you in review

The thing a reviewer or an auditor wants is the list of claims that are guaranteed, separated from the claims you can only measure.

Proven versus measured

A claim owned by the flow or by code is provable - it holds by construction. A claim owned by an agent can only be sampled and reported as a rate.

The code sorts five claims by owner.

CLAIMS = [
    ("this ticket is always routed to billing",     "flow", True),
    ("refunds over 500 are always blocked",         "code", True),
    ("the reply always quotes the policy line",     "agent", False),
    ("a human always approves before sending",      "flow", True),
    ("the summary is always accurate",              "agent", False),
]

print(f"{'claim':46} {'owner':6} provable?")
for claim, owner, provable in CLAIMS:
    print(f"{claim:46} {owner:6} {'yes' if provable else 'no -- only measurable'}")

provable = sum(1 for _, _, p in CLAIMS if p)
print(f"\n{provable} of {len(CLAIMS)} claims can be proven rather than sampled")

# This is what a reviewer or an auditor actually wants: the list of things that
# are guaranteed by construction, separated from the things you can only report
# a measured rate for. Moving a step from agent to flow moves a claim across
# that line.

"Refunds over 500 are always blocked" is provable because a function blocks them. "The reply always quotes the policy line" is not, because a model writes the reply - the best you can say is what fraction did in testing.

Moving a step from agent to flow moves a claim across that line. That is the concrete payoff of the last four units, and it is what to point at when someone asks why the routing is not just another agent.

The mistake this prevents

The mistake is presenting measured claims as guarantees. "The system always cites the policy" is a claim about a model, and the first counterexample makes every other claim in the document suspect. Say "measured at 97% over 400 runs" and keep your credibility.

Takeaway

Separate what is guaranteed by construction from what is measured. Moving a step into the flow moves its claim from the second list to the first, which is the argument for doing it.

Mark complete