Astra's shortest chain-task traces can shrink to a list of intermediate numbers, but not all runs do

A chain task in nocot-bench is a starting number, an ordered list of rules, and a wraparound applied after every step: if the number is bigger than 20, subtract 20; if it is smaller than 1, add 20 ^[6]. Rules come in three flavours: halve and round up, a parity rule, and a threshold rule that subtracts a constant when the value is above 10 and doubles it otherwise ^[7]. The minimal solution to the post's 3-step example writes the start and the value after each step, "14, 7, 14, 7", counted at 10 tokens, or 7 with the spaces taken out ^[8].

Astra's shortest run at each step count sits 4 to 7 tokens above the compressed baseline ^[10]. For the 3-step case that puts the shortest trace at about 11 to 14 tokens against a 7-token version of the same list ^[1]. Framing text accounts for the difference. The comparison against 10, the subtraction, the doubling and the wraparound check never reach the text, so on those runs each rule was applied inside a single forward pass ^[10].

Nanda's post last week reported Astra scoring significantly higher than previous models on no-CoT benchmarks ^[15]. The analysis asks how much of that capability shows up in practice, and the setup pushed toward short traces. Models were told to output only their answer, at reasoning_effort=low, with no other attempt to steer the chain of thought ^[4]. The stated reason for looking at that setting is training pressure toward token-efficient CoTs, which should be strongest in Astra's cheapest and fastest reasoning mode ^[3]. The three siblings did not respond the same way; their token counts suggest they verbalized the work behind each rule instead of only the values, apart from five runs where Sol emitted no CoT at all ^[11].

Whether any of this transfers to a monitored agent depends on things the analysis did not test. The tasks are programmatically generated with a difficulty parameter ^[2]. Five task types were used, two serial, two parallel and one search, each with at least 20 tasks at the highest difficulty ^[4]. Each model ran twice per chain task ^[9]. The whole comparison sits inside one provider's family, picked because those four models share the same reasoning effort options ^[1]. No tool call appears anywhere in it, and no run went above reasoning_effort=low.

The post's own conclusion cuts against the strong reading. Nanda's no-CoT results have Astra answering chain tasks of up to 6 steps correctly with no CoT at all ^[12], so at reasoning_effort=low it is writing down values it can hold internally ^[13]. The author puts that down to error rate: working at the no-CoT capability boundary may carry a higher risk of errors, and errors "multiply" in many-step reasoning ^[14].

In my view the finding worth carrying is narrower than a general claim about opaque traces. On this task type at this setting, the trace still fixes the model's state after every step, and a reader can verify the final number while inferring each rule from the jump between values. A monitor keyed to the state sequence gets what it always got, while one keyed to the operation the model describes gets less than it got from Sol ^[11].