Not a chatbot with helpers. A small set of specialized agents sharing one compounding knowledge mesh, speaking only in evidence packets, and blocked by deterministic gates they cannot override.
Every consequential claim carries sources, confidence, and a falsifier. Irreversible actions — send, spend, delete, schedule — stop at a human checkpoint. That rule is the product.
Click a layer. The stack reads bottom-up: nothing above Layer 5 can bypass the governance plane, and nothing reaches a user without a packet and a receipt.
Agents exchange packets or capsule IDs only. The Fusion Gate is the only component allowed to collapse many packets into one. The Watchdog can kill any trace.
Free-form agent-to-agent text is disallowed — it would bypass the evidence invariant. If a packet lacks sources or a falsifier, the Fusion Gate refuses it.
"packet": { "claim": "Delaying the vendor payment by 9 days keeps cash above floor.", "confidence": 0.82, "supporting_capsules": [ "cap_contract_terms_v3", "cap_cash_position_oct17", "cap_prior_delay_outcome" ], "falsifiers": [ "A receivable slips past Oct 24", "Vendor invokes late-fee clause §4.2" ], "heuristic": "preserve-floor-first", "heuristic_calibration": 0.71 (n=34), "trace_id": "tr_8f31…", "timestamp": "2026-10-18T11:42Z" }
One checkable assertion, never a paragraph of vibes. Confidence is calibrated against later outcomes, not self-reported flair.
Pointers into the mesh. Minimum coverage is enforced by code — below the floor, no positive claim is emitted; the gap is shown instead.
What would make this wrong. The single most useful field in the packet, and the one unconstrained agents never produce.
Every packet is logged immutably. Rejected packets become synthetic stress cases replayed nightly.
Step through one real decision end to end. Notice where the human sits: not at the start, not reviewing raw model output — at the gate, holding the token.
Small enough to instrument completely, large enough to prove whether reliability + compounding actually beat unconstrained agents.
A single-domain founder/ops agent that only emits a recommendation when it can attach grounding — and never executes an irreversible action itself.
The knowledge base that sharpens with use. Atomic capsules, automatic links, context injection, cascade updates, noise filter.
Mentors become procedures, not pastiche. Each heuristic is a falsifiable rule with applicability conditions and a live calibration score.
A factory that turns any new narrow task into a guarded, measurable stack using the AI Stack Blueprint: logic → execution → output → feedback.
The invariants never relax. Speed comes from propagating confirmed evidence faster and fusing independent traces before a human ever sees the result.
One confirmed correction re-scores dozens of dependent capsules and proposes new links. A small proposer model suggests; nothing commits without grounding.
3–5 independent traces per hard question. A claim survives only if ≥2 traces agree on overlapping sources. Disagreement is surfaced, never averaged away.
Heuristics harden into default logic only after measured outcomes cross threshold + sample size. Demotion is automatic and reversible.
Every rejection, heavy edit, and outcome miss becomes a test case, replayed nightly. Regressions block promotion of any stack change.
High-calibration packs export to other teams / verticals and return new outcome data. Shadow Simulator attaches consequence sketches — evidence, never orders.
Timeout, schema violation, or budget breach: trace dies, event logs, retry on a different Reasoner or return “insufficient evidence.”
Two high-confidence packets disagree? Both capsule sets are shown side by side. No automatic averaging, no coin flip.
Router shifts work to cheaper models and cuts parallel traces. Governance Plane can hard-cap spend mid-flight.
Sources dated June–July 2025, re-read October 2026. The test applied throughout: can it be proven, built, and measured right now?
Source minimums, confidence floor, PII scrub, spend ceiling, scope boundary — deterministic code, not model judgment.
The mesh. Agents propose updates; only confirmed outcomes or human edits commit them. Capsules are versioned, never mutated.
Source coverage · fusion rate · edit distance · outcome match · cascade citations · cost per grounded output · watchdog interventions.