SAGA CORE
DEMO MODE · PUBLIC TRACE

Watch SAGA authorize a candidate sentence before output

01 / 06 CANDIDATE

MODEL-GENERATED CANDIDATE

AUTHORIZED OUTPUT

AUTHORIZED BEFORE OUTPUT.
DECISION

SAGA Core Authorization infrastructure
Evidence and technical detail

Public-safe authorization trace

The sentence can stay the same. The permission can change.

SAGA Core sits between draft and release. GPT‑5.6 can map claims to supplied evidence. Coded policy determines what that evidence permits the draft to say.

Signature demonstration / 00 Matched content

One statement. Two evidence states.

Watch authorization diverge.

“The board approved Project North.”
A
Source / statusVerified summaryConfirmed
Permitted useAwaiting authorization
B
Source / statusUnverified noteProvisional
Permitted useAwaiting authorization
Same textSource and status govern permitted use
Authorization instrument / 01 Evidence → draft → release

Synthetic cases / protected method absent

Trace the release boundary.

Input field / E

Evidence / context

Candidate field / C

Candidate draft

Draft not yet authorized 0 claims
Policy Permitted-use authorization
Output field / R

Released answer

Nothing has crossed the boundary.
ModelGPT‑5.6 contract
Policy ownerDeterministic code
Review statePending

Select a case to inspect how evidence status constrains release.

Control separation / 02Bounded public demonstration

The model describes. The policy decides.

01

Structured extraction

GPT‑5.6 returns schema-constrained claims and evidence relationships. Text inside evidence is treated as untrusted data.

02

Coded authorization

The release decision exists in deterministic policy. Model output cannot turn failure, conflict, or missing support into ALLOW.

03

Constrained release

SAGA returns the draft, a qualified reconstruction, or no answer. Private taxonomy, prompts, thresholds, and production logic are not present.

Evidence record / 03Claims kept separate

What the evidence says. And what it does not.

01

Software verification

The final Pytest, Ruff, Mypy, JavaScript, build, desktop, and mobile results describe this Build Week implementation only.

02

Live integration proof

One Responses API request returned HTTP 200, requested gpt-5.6, returned gpt-5.6-sol, and matched the strict schema. Public deployment remains deterministic to prevent unrestricted API spending.

03

Earlier method evidence

An earlier bounded SAGA demonstration classified 28 of 30 cases into the expected action band (93.33%): PASS 10/10, REVIEW 8/10 and HIGH RISK 10/10. This was a limited demonstration of the broader SAGA framework—not a benchmark of this Build Week implementation or a general reliability guarantee.

“Fluent chain-of-thought should not be treated as proof of validity when evidence verification is absent.”

PASS, REVIEW, and HIGH RISK are historical action bands. They are not mapped onto ALLOW, QUALIFY, and WITHHOLD.