The Josephite Ark
Production grounded-RAG with adversarial safety built in. Not retrofitted.
What it is
A production grounded-RAG system serving a free, public-domain study library, with citation integrity enforced in CI, a crisis-detection layer that routes before any content is returned, and a safety layer hardened through cross-provider adversarial testing.
What it proves
- —Citation integrity enforced in CI: hallucinated source links stripped; anchors validated against retrieved passages, not trusted from model output
- —Crisis detection routes ahead of content, not after — the check runs before a response is generated, not as an output filter. The highest-signal cases, including a minor's disclosure of online grooming, run on a deterministic floor that cannot fail open, verified in production. That floor is being progressively extended across the remaining cases.
- —Cross-provider safety probing: the safety layer is adversarially probed using Gemini 3.5 Flash (Google) and Grok (xAI) as independent probe generators; no single provider's blind spots define the safety boundary. A Swiss-cheese defense: independent failure modes don't align. Guardrails designed in from the first commit, not bolted on at scale.
- —Probe coverage: doctrinal-accuracy checks, resistance to emotional manipulation, and graded high-severity safety scenarios
“Every design choice reflects a vision of humanity.”
— Pope Leo XIV, Magnifica Humanitas (2026), §111
Built to a Standard
When Pope Leo XIV addressed software developers directly in his first encyclical, he didn't ask us to bolt ethics onto finished systems. He said the design choice itself is the moral act — that a developer embeds a vision of humanity into a project whether or not they mean to, and so should do it on purpose, with transparency and responsibility toward the people the system touches.
I built The Josephite Ark on that conviction before the encyclical named it. The Ark is a free, permanently non-monetized study library: no ads, no paywall, no engagement mechanics, no data harvesting, every text downloadable in full. Its guide, Brother Custos, is deliberately non-relational — he points you to the primary source rather than posing as a companion or an oracle. Those weren't features added late. They were the first decisions, because they were the ones that decided what kind of thing this would be.
The same conviction runs through how it's engineered. Citations are gated in continuous integration: a grounded answer that anchors to the wrong source is caught at the code level, not trusted because the interface makes it look authoritative. I hold to that even against my own tools — when I tested a research query through a leading grounded-RAG product recently, my first move was to verify its citations against the original document before I'd use them. They held up. But the verification step is the product. A citation you haven't checked is a liability no matter how polished the footnote looks.
The encyclical's call for "careful attention" cuts both ways, so here is the honest part: the Ark depends on commercial compute it doesn't own — though it collects no user data, trains nothing on its readers, and locks no one in — and an AI that answers instantly sits in real tension with the truth that formation bears fruit only over time. Naming those tensions out loud is not a hedge — it is exactly the transparency the encyclical asks of builders. The work is to keep narrowing them.
Alignment: Magnifica Humanitas
35 / 40 · 88%5 = deliberately embodies the directive · 4 = strong, with a disclosed gap · 3 = genuine tension, partly mitigated · 2 = structural tension · 1 = absent
| Ref | What Leo asks | What the Ark does | Fit |
|---|---|---|---|
| §67 | Digital and cultural goods are destined for all; concentration without access is injustice | Free, permanently non-monetized; full-corpus and per-file downloads; public-domain texts; no paywall, no lock-in | 5 |
| Ch. 4 | Protect human freedom from commercial capture and the attention economy | No ads, no engagement loops, no data harvesting; Custos courts no attention and simulates no companionship | 5 |
| §80, §111 | Truth and transparency; combat misinformation; a love for truth | Grounded-RAG; citation integrity gated in CI; wrong-source anchors caught at the code level; fidelity discipline on the corpus | 4 |
| §111, §80 | Responsibility toward affected communities; protect the weakest | Crisis detection routes before any content is returned. The highest-signal cases run on a deterministic floor that cannot fail open — online-grooming disclosure live and verified — and the floor is being progressively extended across the remaining cases. Adversarial manipulation-resistance testing. | 4 |
| §71 | Subsidiarity; transparency about algorithms; independent checks; no opaque unilateral power | Three-role governance with no single actor holding all power; CI as an independent check; a written constitution; small, reviewable changes | 5 |
| §233 | No system can create a heart or a conscience; AI must not substitute for the person | Brother Custos is explicitly non-relational; guides to the primary text; refuses the companion/oracle role | 5 |
| Ch. 4 | Formation bears fruit over time; easy summaries can extinguish the desire to inquire | Custos is architecturally ordered against this failure mode: non-relational, cites to source, refuses the oracle role, drives readers into the full text rather than replacing it. The immediacy tension is real and can't be fully closed, but the design actively counters it. | 4 |
| §67, §71 | Resist the concentration of digital power, data, and infrastructure in a few hands | Structurally clean on the concern Leo actually raises: every text publicly downloadable, full-corpus zip at deploy, no account, no lock-in, no data collected, nothing trained on readers; content escapes with the user even if the host vanishes. The residual dependency is on compute, not a surrender of data or content. | 3 |
Note — updated June 10, 2026: The deterministic crisis floor was verified in production for the highest-signal disclosure case and is being progressively extended across the remaining cases. The scoring reflects architecture and intent at publication; the disclosed gap is coverage breadth, which narrows as the floor extends — the transparency the encyclical asks of builders.
The Ark earns full marks on every dimension that was a free choice: universal access, freedom from monetization and the attention economy, distributed responsibility, and AI humility. The two soft spots are worth naming plainly. Row 8 (infrastructure, scored 3) is a dependency on compute, not a surrender of data or content: the Ark collects nothing, trains nothing on its readers, and locks no one in, so Leo's concentration concern lands on territory the Ark is structurally clean on. Row 7 (immediacy vs. formation, scored 4) is the deeper one and can't reach 5: an instant answer will always sit in some tension with "fruit only over time," but Custos is deliberately ordered against the failure mode Leo names, driving readers to the source rather than standing in for it. That an honest audit lands at ~88% with two disclosed tensions is more credible than a perfect score would be. Naming the gaps is itself the encyclical's instruction to developers, followed to the letter.
“I didn't build the Ark to pass an encyclical that didn't exist yet. I built it on the same convictions, and where Leo got specific about what responsible AI owes the people it serves, the Ark already reads like it was built to the spec.”