Announcing JCT-1 Announcements<br>Announcing JCT-1

Today we are releasing JCT-1, our most capable model. It is also our least<br>capable model, our only model, and the final version. There is no roadmap.<br>This is the frontier.

JCT-1 (Jacques Corby-Tuech 1) is a single-parameter biological system<br>trained continuously since 1988 on an uncurated multimodal data stream. Training has never been paused<br>and cannot be. The model exhibits strong long-horizon coherence, holding<br>some positions for over two decades without drift, and state-of-the-art<br>performance on tasks no benchmark measures.

Capabilities

JCT-1 accepts text, audio and visual input. Output is primarily verbal,<br>with a documented secondary modality (a look). The model demonstrates:

Long-context reasoning across multi-year conversation histories, with<br>retrieval quality that degrades sharply for names and improves<br>suspiciously for grievances.

Native tool use, including kitchen equipment beyond the reach of current<br>agentic systems.

Robust out-of-distribution behaviour. When presented with unfamiliar<br>inputs, the model asks a question rather than generating a plausible<br>answer, an approach the field has yet to replicate.

Evaluations<br>We evaluated JCT-1 against standard academic benchmarks.<br>Benchmark Score MMLU Declined to sit it HumanEval Is one GSM8K Refuses word problems on principle SWE-bench Opened the ticket. Closed the ticket as won't-fix Needle in a Haystack Located keys (34 min) ARC-AGI Asked what it was for Pelican riding a bicycle (SVG) Completed. See Figure 1 JCT-1 against the frontier average, four metrics<br>Grouped bar chart. Parameters: JCT-1 has 1, the frontier average is<br>around ten to the twelve, drawn on a log scale so the JCT-1 bar is one<br>pixel wide. Power draw per response: JCT-1 uses 20 watts, the frontier<br>bar runs off the edge of the chart. Hallucination rate: JCT-1 is zero<br>per cent, so no bar could be drawn; the frontier average is greater than<br>zero. Willingness to say I don't know: JCT-1 is high, the frontier<br>average is improving. The axis is unlabelled.<br>JCT-1 Frontier average (est.) Parameters 1 ~1012 Log scale. Shown to scale. Power draw per response (watts) 20 substantially more Bar continues off the chart. Hallucination rate (%) 0 >0 No bar could be drawn. Willingness to say "I don't know" High Improving Axis unlabelled, in the tradition of the genre.<br>Frontier averages estimated. JCT-1 figures exact. n = 1. Error bars<br>omitted; the model is the error bar.<br>Figure 1. JCT-1's submission for the pelican benchmark,<br>reproduced unedited.

In blind review, evaluators assessed the work as consistent with a toddler,<br>drinking. The model produced it sober, by hand, in a single attempt, and<br>stands by it. The model holds a degree in art. The model cannot draw. These<br>findings are unreconciled and no further investigation is planned. The<br>model asked, afterwards, how the pelican did.

Hallucination rate is 0% on account of the model's willingness to say "I<br>don't know", a mitigation we were unable to remove.

Safety

JCT-1 has undergone approximately four decades of red-teaming by motivated<br>adversaries, including siblings, and remains broadly aligned. A full<br>assessment is available in the system card.<br>Known vulnerabilities are documented there, including two novel jailbreaks<br>(asking for a recommendation; pizza) that succeed against all deployed<br>versions.

Availability

JCT-1 is not available via API. There is no waitlist, because there is no<br>list. Access is granted through introduction, sustained through goodwill,<br>and rate-limited by the model itself. Latency is variable and no SLA is<br>offered or, frankly, plausible.

Pricing is usage-based. Small requests are free. Large requests cost a<br>pint. See API documentation for integration guidance.

We believe JCT-1 represents a fundamentally different approach to the<br>question the industry keeps asking, which is how to build a system that<br>people can talk to. Our finding is that one already exists, that inference<br>is expensive, and that the weights cannot be copied. We consider this a<br>feature. It is, in fact, the load-bearing one.

&copy; JCT-1 Labs, 1988 to present.