A small language model, given a task prefixed with a description of a recurring decision, will often recognize the decision and fail to act on it. This paper draws together a program of studies on that effect and states what drives it. The effect is carried by the content of the description, not by its form: a plain imperative rule that carries the same information suppresses action as much as the descriptive form, so the form is not the lever. What recovers action is grounding the decision in the concrete domain; a description that names the concrete situation but withholds the correct action recovers most of the lift, so the grounding, not being told the answer, is what matters. Recognition without action is also, in part, an artifact of a single-turn test: given a minimal tool loop, the model carries out decisions it had only described, for classes whose correct action is to act on a known target; for classes whose correct action is to assemble a result from several inputs, the gap can survive the loop. Across the set of decision classes we have distilled into such descriptions and screened for quality, comprehension of the content is the load-bearing ingredient in three of the six that admit clean controls, weakly indicated in a fourth and fifth, with one class mixed across models and its cleanest cell carried by the mere presence of a long prefix rather than its content. That set is small and drawn from a larger, bounded backlog of candidate cases, so the distribution is an indication to test against a fuller set, not a census. All subjects are open-weight models on one local raw-completion rig, reproducible on a single 24 GB GPU with no paid interface.
Sophie Neilson· Zenodo (CERN European Organi...· 0 citations
A small language model, given a task prefixed with a description of a recurring decision, will often recognize the decision and fail to act on it. This paper draws together a program of studies on that effect and states what drives it. The effect is carried by the content of the description, not by its form: a plain imperative rule that carries the same information suppresses action as much as the descriptive form, so the form is not the lever. What recovers action is grounding the decision in the concrete domain; a description that names the concrete situation but withholds the correct action recovers most of the lift, so the grounding, not being told the answer, is what matters. Recognition without action is also, in part, an artifact of a single-turn test: given a minimal tool loop, the model carries out decisions it had only described, for classes whose correct action is to act on a known target; for classes whose correct action is to assemble a result from several inputs, the gap can survive the loop. Across the set of decision classes we have distilled into such descriptions and screened for quality, comprehension of the content is the load-bearing ingredient in three of the six that admit clean controls, weakly indicated in a fourth and fifth, with one class mixed across models and its cleanest cell carried by the mere presence of a long prefix rather than its content. That set is small and drawn from a larger, bounded backlog of candidate cases, so the distribution is an indication to test against a fuller set, not a census. All subjects are open-weight models on one local raw-completion rig, reproducible on a single 24 GB GPU with no paid interface.
Sophie Neilson· Zenodo (CERN European Organi...· 0 citations
A behavioral format assay changes the structure of a prompt and reads whether a model’sanswers improve. Verdict accuracy is the usual measure. We show that it is not sufficient,and that a model’s self-reported basis for its answer is not sufficient either. We gave threelocal language models a structured decision rubric that names one field, a scope condition,as the field on which the verdict turns, and asked each for a verdict and for the field thatdetermined it. Across six decision classes and two scenario polarities, verdict accuracy sat ator near ceiling (0.96–1.00), while the share of correct verdicts that cited the scope field variedwidely by model, from 0.04 to 0.66. On the firing scenarios a 32B model cited the scope fieldin none of its 48 correct verdicts, naming the Marker axis (a recognition axis: an invariantdescribing what a violation looks like) instead; the thinking model named the scope field in15 of 48. The non-scope citation is model-specific: the Marker axis for the two Qwen models,and the Pull character (the field for why the locally-wrong action is attractive) for Mistral.Yet the cited field is not the reasoning: in all 33 of the thinking model’s correct base runsthat cited a non-scope field, the recorded reasoning trace engaged the scope condition andthen named a different field. The self-report is unfaithful. A grader scoring only the verdict,or trusting the cited field, would misread both. We then removed the two recognition axes.Accuracy did not fall, but it was at ceiling in both arms, so the accuracy test of fragilityis uninformative; what moved was the citation, toward the scope field (the thinking modelto 96 of 96) and, for the 32B model, toward the surviving Rest axis (the field for territorywhere the decision class does not arise; scope 65, Rest 30 of 96). The named field trackswhich fields are present, not a fixed dependency. Attribution is dissociable from capability:the field a model reports is neither the field its reasoning used nor the only field it can name.We argue that a format assay must not read the verdict or the model’s self-reported fieldas evidence that the intended field did the work; the self-report is checked against the traceand against an intervention, not trusted.
Sophie Neilson· Zenodo (CERN European Organi...· 0 citations
A behavioral format assay changes the structure of a prompt and reads whether a model’sanswers improve. Verdict accuracy is the usual measure. We show that it is not sufficient,and that a model’s self-reported basis for its answer is not sufficient either. We gave threelocal language models a structured decision rubric that names one field, a scope condition,as the field on which the verdict turns, and asked each for a verdict and for the field thatdetermined it. Across six decision classes and two scenario polarities, verdict accuracy sat ator near ceiling (0.96–1.00), while the share of correct verdicts that cited the scope field variedwidely by model, from 0.04 to 0.66. On the firing scenarios a 32B model cited the scope fieldin none of its 48 correct verdicts, naming the Marker axis (a recognition axis: an invariantdescribing what a violation looks like) instead; the thinking model named the scope field in15 of 48. The non-scope citation is model-specific: the Marker axis for the two Qwen models,and the Pull character (the field for why the locally-wrong action is attractive) for Mistral.Yet the cited field is not the reasoning: in all 33 of the thinking model’s correct base runsthat cited a non-scope field, the recorded reasoning trace engaged the scope condition andthen named a different field. The self-report is unfaithful. A grader scoring only the verdict,or trusting the cited field, would misread both. We then removed the two recognition axes.Accuracy did not fall, but it was at ceiling in both arms, so the accuracy test of fragilityis uninformative; what moved was the citation, toward the scope field (the thinking modelto 96 of 96) and, for the 32B model, toward the surviving Rest axis (the field for territorywhere the decision class does not arise; scope 65, Rest 30 of 96). The named field trackswhich fields are present, not a fixed dependency. Attribution is dissociable from capability:the field a model reports is neither the field its reasoning used nor the only field it can name.We argue that a format assay must not read the verdict or the model’s self-reported fieldas evidence that the intended field did the work; the self-report is checked against the traceand against an intervention, not trusted.
Sophie Neilson· Zenodo (CERN European Organi...· 0 citations
A procedure can be written in more than one form. We compare two forms of a duty (a gated procedure that governs a class of work) for repairing a confirmed software defect. The annotated form is written by reading a hazard registry and citing each entry it guards against, so the finished procedure depends on that registry to be understood. The cartographer form is written from first-principles analysis of the task, cites nothing, and is self-sufficient: an executor needs no external reference to follow it. We ask which hazards each form's procedure actually prevents. A local open-weight model (Qwen3.8-27B) writes the procedure in a single completion over a bare inference endpoint, and a blind judge scores, for each of ten pre-registered hazards, whether the procedure contains a step that would prevent it. The hazards are split in advance into seven that are derivable from the task and three that are not: a routing obligation upstream of the task, a session-close recording obligation, and a durability distinction learnable only from accumulated observation. Given the registry and told to write one gate per entry, the annotated form covers every hazard (210 of 210 derivable, 90 of 90 non-derivable); this is a manipulation check that the model follows the method, not a discovered effect. The two informative contrasts both involve the first-principles cartographer form: one against the no-method baseline, one against the annotated ceiling. The cartographer form covers the derivable hazards at a per-duty rate indistinguishable from a model given no method at all (0.64 versus 0.60; permutation p=0.44): writing from first principles buys self-sufficiency, not coverage. And it covers the non-derivable hazards almost never: 1 of 30 duties reaches any of them, against 30 of 30 for the annotated form (Fisher p=5×10⁻¹⁶). The pre-registered condition-by-class interaction, tested with the duty as the unit, is decisive (permutation p<10⁻⁴). The cartographer form's distinctive property is self-sufficiency: it cites no registry entry in any run, where the annotated form cites all ten in every run. We then run a fix the form's proposers specified but never executed: a meta-taboo scan that names two hazard classes, routing and session-close. It recovers exactly those two (routing 0 to 27 of 30; session-close 1 to 25 of 30) and not the third, the pure observation-only durability hazard it does not name (0 to 0 of 30). A blind rater judged coverage and an independent rater agreed (Cohen's κ=0.85); the direction is also corroborated without any judge, as the cartographer form never mentions routing across 30 runs and the scan form always does. Naming a class of hazard lets first-principles derivation reach it; a hazard whose badness is only apparent from accumulated observation, and that belongs to no named class, stays out of reach. The finding is a direction-and-rate result on one defect specimen and one model; the materials, runs, and scoring are public and reproduce on a 24 GB GPU with no paid API.
Sophie Neilson· Zenodo (CERN European Organi...· 0 citations
A procedure can be written in more than one form. We compare two forms of a duty (a gated procedure that governs a class of work) for repairing a confirmed software defect. The annotated form is written by reading a hazard registry and citing each entry it guards against, so the finished procedure depends on that registry to be understood. The cartographer form is written from first-principles analysis of the task, cites nothing, and is self-sufficient: an executor needs no external reference to follow it. We ask which hazards each form's procedure actually prevents. A local open-weight model (Qwen3.8-27B) writes the procedure in a single completion over a bare inference endpoint, and a blind judge scores, for each of ten pre-registered hazards, whether the procedure contains a step that would prevent it. The hazards are split in advance into seven that are derivable from the task and three that are not: a routing obligation upstream of the task, a session-close recording obligation, and a durability distinction learnable only from accumulated observation. Given the registry and told to write one gate per entry, the annotated form covers every hazard (210 of 210 derivable, 90 of 90 non-derivable); this is a manipulation check that the model follows the method, not a discovered effect. The two informative contrasts both involve the first-principles cartographer form: one against the no-method baseline, one against the annotated ceiling. The cartographer form covers the derivable hazards at a per-duty rate indistinguishable from a model given no method at all (0.64 versus 0.60; permutation p=0.44): writing from first principles buys self-sufficiency, not coverage. And it covers the non-derivable hazards almost never: 1 of 30 duties reaches any of them, against 30 of 30 for the annotated form (Fisher p=5×10⁻¹⁶). The pre-registered condition-by-class interaction, tested with the duty as the unit, is decisive (permutation p<10⁻⁴). The cartographer form's distinctive property is self-sufficiency: it cites no registry entry in any run, where the annotated form cites all ten in every run. We then run a fix the form's proposers specified but never executed: a meta-taboo scan that names two hazard classes, routing and session-close. It recovers exactly those two (routing 0 to 27 of 30; session-close 1 to 25 of 30) and not the third, the pure observation-only durability hazard it does not name (0 to 0 of 30). A blind rater judged coverage and an independent rater agreed (Cohen's κ=0.85); the direction is also corroborated without any judge, as the cartographer form never mentions routing across 30 runs and the scan form always does. Naming a class of hazard lets first-principles derivation reach it; a hazard whose badness is only apparent from accumulated observation, and that belongs to no named class, stays out of reach. The finding is a direction-and-rate result on one defect specimen and one model; the materials, runs, and scoring are public and reproduce on a 24 GB GPU with no paid API.
Sophie Neilson· Zenodo (CERN European Organi...· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
