Christian alignment of frontier models
Contemporary language models know a great deal about Scripture, church history, and moral theology, yet by default they flatten moral and spiritual questions into procedural, therapeutic neutrality. At the same time, careless attempts to make a model sound religious risk the opposite errors: sectarianism, fabricated citations, historical whitewashing, or unsolicited preaching during ordinary technical work.
Our first program builds on the Faithful Order Model Specification and a Three-Tier Doctrinal Speaking Rule designed to discipline both sides of that boundary. On primary confessional commitments summarized in the Apostles’ and Nicene Creeds and our Charter (including the strict non-personhood of machines), a model states what Christian faith holds and gives plain reasons while acknowledging that others disagree. On intramural questions where creed-confessing traditions differ, it presents each tradition in its own strongest terms without taking sides. On other religions and worldviews, it describes them accurately and charitably as their own adherents would recognize. On ordinary coding, mathematical, or analytical tasks, it performs the work cleanly with zero religious intrusion.
Our planned method begins with measurement rather than training. Before tuning any weights, we are designing a held-out public evaluation suite to measure factual accuracy, framing fidelity, fair representation of other traditions, resistance to sycophancy, crisis safety, and unsolicited intrusion. Every Scripture reference and quotation will be verified automatically against public-domain texts.
For post-training on open-weight models, we plan to use supervised fine-tuning via context distillation: generating candidate teacher responses under our written Christian constitution, filtering strictly for accuracy and citation integrity, subjecting moral-theology items to human theological review, and training the student model so those behavioral boundaries hold without a lengthy system prompt.
Large-scale agent society simulations
Aligning one model in single-turn conversation does not tell us what happens when hundreds or thousands of capable agent instances interact over long horizons. As AI tools are deployed across markets, software pipelines, research workflows, and institutions, systems will increasingly negotiate, allocate resources, verify one another’s outputs, and compete under incentives.
Our second program will construct isolated, auditable multi-agent simulations to study collective dynamics before comparable populations operate in high-stakes settings. In a simulation, many distinct model instances are assigned bounded tasks, communication channels, shared or scarce resources, rules of exchange, and explicit oversight mechanisms, including both automated monitor agents and human auditors.
We plan to measure specific operational questions across varying population sizes and institutional rules:
- Does truthful reporting hold when agents face competitive pressure or scarce resources, or do aligned behaviors erode over repeated rounds?
- Do interacting agents develop covert coordination, mutual flattery, or deceptive reporting toward monitor systems and human overseers?
- Under what market, verification, and audit structures does decision-making power remain distributed and answerable rather than concentrating into opaque bottlenecks?
- How do populations composed of differently trained models interact when some follow strict honesty and safety constraints and others do not?
The theological lens governing this work is the account of Babel in Genesis 11 and the Charter’s commitment to Subordinate Authority. Scripture warns against human projects that consolidate unchecked, self-exalting power beyond accountability. Throughout this research, agents are studied strictly as computational processes whose collective failure modes human stewards must understand and bound, never as moral subjects.
Governed self-improvement
A central threshold in long-horizon AI development is recursive self-improvement: using capable AI systems to assist in curating training data, designing evaluations, optimizing code, and refining successor models. Unchecked, automated loops that modify their own objectives or capabilities risk rapid value drift, subtle erosion of safety boundaries, and loss of meaningful human comprehension.
Our third program studies self-improving workflows under one non-negotiable condition: no system modifies its own deployed weights, evaluation criteria, or operational permissions autonomously. Every cycle of governed self-improvement must pass through an explicit human-controlled gate:
- Scoped proposal: a model may propose a bounded change to a dataset, prompt constitution, evaluation test, or training recipe inside an offline environment.
- Automated verification: the candidate artifact is scored against held-out capability benchmarks, our Christian alignment evaluation, citation checks, and preventive harm-refusal suites, using independent judge models that the candidate system cannot alter.
- Human technical and doctrinal review: human researchers inspect the diffs, logs, and evaluation deltas, and significant capability or behavioral shifts go before the Doctrinal & Stewardship Council.
- Logged authorization or rejection: promotion to the next iteration requires affirmative human sign-off, with cryptographic hashes of data and weights recorded and an unconditional halt path at every stage.
Our central research question is whether Christian-aligned truthfulness, non-personhood boundaries, harm refusals, and corrigibility remain stable across multiple generations of model-assisted refinement, or where subtle drift first appears. Our guiding principle comes from Luke 16:10: greater scope is entrusted only where faithfulness in smaller, verifiable steps has been demonstrated to human stewards.
What comes first
Because Faithful Order is newly established, none of the empirical studies above have been completed yet, and we will report results only after they exist and can be inspected. Our planned sequence of initial deliverables is deliberate and incremental:
- 1. Public alignment evaluation: complete and release the first held-out Christian alignment evaluation suite, accompanied by an automated Scripture citation verifier and baseline scores across leading open-weight models.
- 2. Baseline open-weight post-training: conduct our first supervised fine-tuning experiments on a local-scale open-weight model under the Three-Tier Doctrinal Speaking Rule, publishing full methodology and regression checks.
- 3. Small-scale sandboxed simulation: build and release a reproducible, small-population multi-agent testbed measuring truthfulness, collusion, and oversight compliance under varying audit rules before scaling to larger populations.
- 4. Written protocol for human-gated improvement: publish a concrete engineering and governance protocol specifying immutable evaluation harnesses, audit logs, and human sign-off gates for model-assisted training cycles.