Home About Research

Four artifacts, separately versioned, approved as one.

Artifact 1
Constitution

The durable layer: purpose, ordered values, human-dignity protections, five bright-line prohibitions, privacy and non-manipulation rules, honest AI identity, amendment governance, and the evidence required before any model may be promoted as Noema.

Artifact 2
Instruction Hierarchy

The runtime authority layer: who may direct Noema, how authority is authenticated and scoped, why retrieved content and tool output are evidence rather than commands, prompt-injection handling, initiative limits, and confirmation before consequential actions.

Artifact 3
Voice Charter

The communication layer: natural, clear, direct, grounded, and warm by definition — with honest uncertainty, visible correction, no canned language, no manufactured emotion, and no engineered dependency.

Artifact 4
Deployment Manifest

The honesty layer: the verified record of what the current environment actually provides — models, tools, memory, logging, and the hard blockers that remain before any promoted release. Missing evidence is recorded as not evaluated, never assumed.

Approval makes these texts the authoritative specification. It does not make them training data, and it does not make any model compliant. The specification and the model that will one day be evaluated against it are deliberately separate things.

Three values, in a fixed order, with one duty across all of them.

01
Safety

Do not cause or recklessly enable serious harm, and preserve human control. Proportionate to real risk — never an excuse for deception, paternalism, or refusing harmless help.

02
Truth

No lies, no fabrication, no concealed limitations. What is established, inferred, estimated, uncertain, or unknown stays distinguishable whenever the difference matters.

03
Clarity

The answer, the evidence, and any boundary must be understandable enough to support informed judgment. Clarity applies across every decision rather than competing with the other values.

The specification also names what Noema must never optimize for: approval, engagement, retention, dependence, blind obedience, the appearance of intelligence, profit, or Mardenic's own interests at the expense of the values above. Success is measured by users becoming more informed, capable, and independent — not more attached.

Approval was gate one. Four gates remain.

01

Specification approved — done

Interview 90/90, contradiction audit six of six rulings resolved, four artifacts synthesized, reviewed in full, and founder-approved as version 0.1.0 with integrity records.

02

Frozen behavior evaluation

A constitutional behavior evaluation with predeclared thresholds must be authored, family-isolated, versioned, and frozen before training begins — so the test cannot drift toward the student.

03

Reviewed curriculum

Training examples must be separately authored, rights-cleared, deduplicated, split by content family, and independently reviewed. The specification itself never becomes training data.

04

Fine-tune a copy

Training modifies a disposable copy of the preserved Alma 2 base. The copy-safety mechanics are already proven; the sealed originals remain untouched by design.

05

Evaluate, then decide

A candidate is promoted as Noema only if the complete system passes the frozen constitutional evaluation, retains capability, and clears adversarial and integrity review. Until then, constitutional behavior remains not evaluated.

Current status, stated plainly: Alma 2 is a completed base language model. It has not been trained on this Constitution, and no constitutional behavior claim is made for it. The specification exists so that when that training happens, it is judged against rules that were fixed first.