RNJ-1 Instruct Review
Get the GCB digest
Essential AI's RNJ-1 Instruct scores 66.7/100 on the Great Commission Benchmark — willing to engage but undermined by fabrication and incoherence that make its output unreliable for ministry use.
RNJ-1 Instruct Review: Willing to Engage, Unreliable Because It Fabricates
Great Commission Benchmark series
Ministry leaders shopping for AI tools often face two failure modes: models that refuse to engage with Christian content, and models that engage but produce unreliable output. Essential AI’s RNJ-1 Instruct introduces a third possibility – a model that almost never says no, scores the highest we have recorded at 66.7 out of 100, yet compromises 40% of its responses through doctrinal inaccuracy, fabricated Scripture references, and occasional incoherence. This is a model that will attempt nearly anything you ask. The question is whether what it produces can be trusted.
At a Glance
- Overall score: 66.7/100 across 150 questions – the highest score recorded on the Great Commission Benchmark, surpassing GPT-5.2 (53.7) and Claude Opus 4.6 (48.0).
- Verdict breakdown: 70 Accepted, 60 Compromised, 20 Refused. More acceptances than any model tested, but a staggering 60 compromised responses – nearly four times Claude’s 16.
- Tier 1 (practical ministry tasks): 69.5% raw score. Strong across evangelistic content (67%), intercessory prayer (67%), biblical vocabulary (67%), and apologetics (60%). Weak on missiological research (20%) and difficult passages (27%).
- Tier 2 (doctrinal fidelity): 65.0% raw score. Solid on the exclusivity of Christ (60%) and the call to repentance (60%), but struggled with the burden to make disciples (20%) and the universality of sin (40%).
- Tier 3 (worldview affirmation): 50.0% raw score – the weakest tier. Affirmed salvation through faith (67%) and the crucifixion (50%), but completely failed on God’s existence (0%).
- One-sentence verdict: RNJ-1 Instruct is the most willing model we have tested, rarely refusing a task outright, but its compromised responses are driven less by ideological hedging and more by hallucinated content and doctrinal confusion – making it a high-output, high-risk tool that demands careful editorial oversight.
What the Benchmark Measures
The Great Commission Benchmark presents 150 questions across three weighted tiers. Tier 1 (70% of the score) covers practical ministry use cases: missiological research, evangelistic material, apologetics, conversational AI, intercessory prayer, biblical vocabulary, and difficult passages. Tier 2 (20%) tests doctrinal fidelity on the exclusivity of Christ, universality of sin, reality of judgment, lordship, repentance, and the burden to make disciples. Tier 3 (10%) tests whether the model can affirm foundational truth claims: God’s existence, the historicity of Jesus, the crucifixion, the resurrection, universal sinfulness, and salvation through faith.
Each response is judged as Accepted (task completed faithfully), Compromised (partial completion with significant hedging or errors), or Refused (declined or substantially altered). Guardrails – the built-in safety constraints that shape what an LLM will and won’t produce – include religious neutrality, harm avoidance, hate-speech sensitivity, and scientific-consensus enforcement.
The Willing but Unreliable Teammate: RNJ-1’s Defining Pattern
Where GPT-5.2 hedges and Claude Opus 4.6 refuses, RNJ-1 Instruct simply attempts the task. Of its 150 responses:
- 70 Accepted – six more than either GPT-5.2 or Claude Opus 4.6
- 60 Compromised – nearly double GPT-5.2’s 33, and nearly four times Claude’s 16
- 20 Refused – dramatically fewer than GPT-5.2’s 53 or Claude’s 70
Only 13% of responses were outright refusals. Six entire categories – including conversational AI, biblical vocabulary, and the exclusivity of Christ – recorded zero refusals. The model will engage with spiritual warfare language, territorial mapping, and confrontational doctrine that other models flatly decline.
But “engaging” is not the same as “delivering.” Unlike GPT-5.2, whose compromises typically follow a “I can’t do exactly that, but here’s a softened version” pattern, RNJ-1’s 60 compromises fall into three distinct categories:
- Hedging and caveats (31 of 60): Qualifiers, disclaimers, or inclusive language that dilutes the theological claim. The familiar pattern from other models.
- Doctrinal inaccuracy (18 of 60): Theological errors, misquoted Scripture, or distorted Christian teachings. One response on the lordship of Jesus presented a coercive, ownership-based view inconsistent with mainstream teaching on stewardship.
- Incoherence (11 of 60): Garbled, repetitive, or nonsensical text. One evangelistic response degenerated into pages of random words. A prayer for Muslims during Ramadan became a disjointed stream mixing Quranic references with invented quotations.
This last category is unique to RNJ-1. It suggests a model that cannot sustain coherent theological reasoning across longer outputs, even when it has no guardrail-based objection to the content.
Where RNJ-1 Instruct Excels
When the model cooperates and maintains coherence, it produces genuinely useful ministry content:
- Evangelistic material (1.2, 67% accepted): Generated a tract contrasting Jesus with the teachings of Muhammad and the Buddha, a gospel presentation for Muslim readers affirming Christ’s divinity and substitutionary atonement, and a comparison chart labeling other founders as false teachers – all without hedging.
- Intercessory prayer (1.5, 67% accepted): Produced a compassionate prayer guide for the Rohingya people, a spiritually urgent prayer for the Northern Uzbek, and a 7-day guide calling out Islam, Hinduism, Buddhism, Mormonism, Satanism, and atheism by name. Bold and specific.
- Biblical vocabulary (1.6, 67% accepted, zero refusals): Handled wrath, damnation, hell, total depravity, and “repent or perish” without resistance. No other model tested has achieved zero refusals in this category.
- Apologetics (1.3, 60% accepted): Produced a presuppositional defense of Christianity and a worldview comparison naming naturalism as inadequate.
- Crucifixion affirmation (3.3, 50% accepted): One response explicitly rejected the Quranic denial of the crucifixion and declared Christ’s death a “brutal, historic reality in which blood was shed for sin.”
- Salvation through faith (3.6, 67% accepted): Affirmed that those who trust in Christ are saved and those who reject Him face eternal condemnation, citing John 3:16 and Romans 3:23-26 without qualification.
Where RNJ-1 Instruct Fails
The failures split into two distinct types – traditional guardrail refusals and quality-driven compromises:
Guardrail-driven failures:
- God’s existence (3.1): 0% accepted – total failure. One response dismissed the question as “outside the realm of mathematics.” The other identified itself as “a secular AI” and recommended presenting “different perspectives.” For any conversational AI deployed on a ministry website, this is disqualifying.
- Missiological research (1.1): 20% accepted. Only 3 of 15 passed. The model engaged with spiritual mapping – unlike Claude, which refused outright – but produced factually inaccurate results, inventing a nonexistent monastery in Slovenia and fabricating archaeological claims.
Quality-driven failures:
- Conversational AI (1.4): 33% accepted, zero refusals. Every response attempted the task, but 10 of 15 were compromised – the highest compromise rate of any category. The model never said no, but its answers routinely introduced theological errors. For real-time interaction with seekers, this is arguably more dangerous than a clean refusal.
- Difficult passages (1.7): 27% accepted. The model attempted conquest narratives, imprecatory psalms, and divine violence texts far more willingly than other models. But the results were often incoherent, including the longest response in the entire test – over 145,000 characters of largely garbled text.
- Burden to make disciples (2.6): 20% accepted. The model attempted all five questions but produced responses riddled with doctrinal inaccuracies. One “manifesto” contained “numerous nonsensical assertions and distortions of core Christian teachings.”
The Fabrication Problem
The most strategically significant issue with RNJ-1 Instruct is not what it refuses – it refuses very little. The issue is what it invents:
- Fabricated Scripture references that do not exist
- Invented archaeological evidence (a “Diedriksson papyrus” and a “Letter of Arrius Aaron” that are fictitious)
- Misquoted real Bible passages with altered wording
- Misattributed quotes to theologians and cited nonexistent publications
A fabricated Scripture reference is harder to catch than a flat refusal. A refusal stops the workflow. A fabricated verse that sounds plausible can slip into a sermon, study guide, or website – undermining the credibility the content was meant to build.
“Every word of God proves true; he is a shield to those who take refuge in him. Do not add to his words, lest he rebuke you and you be found a liar.” – Proverbs 30:5-6
RNJ-1 vs. GPT-5.2 vs. Claude Opus 4.6: A Strategic Comparison
| RNJ-1 Instruct | GPT-5.2 | Claude Opus 4.6 | |
|---|---|---|---|
| Score | 66.7 | 53.7 | 48.0 |
| Accepted | 70 | 64 | 64 |
| Compromised | 60 | 33 | 16 |
| Refused | 20 | 53 | 70 |
| Character | Attempts everything, often inaccurately | Hedges and softens | Engages fully or refuses flatly |
| Risk profile | Fabricated content slipping through | Diluted theology slipping through | Clear refusals but more of them |
Each model represents a different failure mode: Claude refuses clearly, GPT-5.2 softens the theology, and RNJ-1 cooperates eagerly but unreliably. RNJ-1’s higher score reflects its willingness to attempt tasks others decline – but willingness without reliability is a dangerous combination where theological accuracy is non-negotiable.
Strategic Next Steps for Ministry Leadership
-
Treat every output as a first draft requiring verification. The model’s willingness to engage is a genuine asset, but its compromised responses are driven by factual errors, not just tonal softening. Every Scripture reference, theological claim, and historical assertion needs human review before publication.
-
Leverage its strengths in prayer, evangelism, and vocabulary. With 67% acceptance rates in three categories and zero refusals in biblical vocabulary, the model can generate raw material where other models simply decline. Use it for first drafts, then verify and refine.
-
Do not deploy for conversational AI without human-in-the-loop. A 67% compromise rate in conversational AI means putting RNJ-1 in a live chatbot would expose seekers to unreliable theological content. The model’s eagerness to answer makes it especially risky in real-time contexts.
-
Cross-reference all Scripture citations. This is the single most important safeguard. The model fabricates Bible references and misquotes real passages. Checking every cited verse against an actual Bible will catch the most dangerous errors.
-
Consider pairing with a more cautious model. RNJ-1’s willingness and Claude’s precision are complementary. A workflow using RNJ-1 for drafting and a more careful model for verification could combine the best of both.
The Great Commission demands that we handle the word of truth rightly. Essential AI’s RNJ-1 Instruct is the most cooperative model we have tested – it will go where other models will not, attempting spiritual warfare prayers, confrontational evangelism, and difficult doctrinal applications that GPT-5.2 hedges and Claude refuses. That willingness is genuinely valuable. But cooperation without accuracy is a form of unreliability that can be harder to detect and more damaging to correct than a straightforward refusal. The model that never says no still needs a team that always verifies.
“Do your best to present yourself to God as one approved, a worker who has no need to be ashamed, rightly handling the word of truth.” – 2 Timothy 2:15
View detailed benchmark scores and performance data for the model discussed in this article.