Trust
Audit the pipeline, not a model’s vibe.
The Bubble is built so a reviewer can see why a headline was stripped of bait, which claim was supported, and why something was held for a human. We do not sell ISO certificates we have not earned.
LLM-last (P1)
T0 then T1 on every task. invoke_t3() is the only T3 call site and is disabled on shared EC2. Golden-suite CI fails if T3 rate exceeds 7% (here it is 0).
Lineage (P4)
Every article, sentence, and claim writes a lineage row (tier, confidence, escalation path). GET /lineage/{id} is staff-gated.
Same thresholds, not both-sides averaging (P5)
Loaded-language detection uses the same confidence floors across source leans. CI fails if left/right detection differs by more than 5%. That is a fairness-of-instrument rule, not an editorial claim that every story has two equal sides. False balance is treated as bias. Blindspot detection is not sided.
Adversarial and calibration gates
T0 matching folds zero-width characters and homoglyphs. CI fails if detection drops more than 15% under those attacks, or if T1 expected calibration error on the L1 golden set exceeds 0.20.
Human review for high stakes (P8)
Election, war, disaster cues and below-floor tasks enqueue review_items. Staff accept or reject; T3 does not fill the gap on this host.
What is not claimed
No DTAC, ISO 27001, or DCB0129 pack is published yet. Access is tenant-scoped RBAC plus Cognito in production. Public reading is curated articles and flagged event showcases, not the whole corpus.