When the Evidence Changes, Does the Model Really Update?

In a previous experiment, I gave a frontier language model a diagnosis-masked neuromuscular case and asked it to rank possible diagnoses.

The result was impressive. The model identified the published diagnosis, highlighted the most relevant clinical features, and predicted the kind of evidence that would likely resolve the case.

But a correct answer does not necessarily imply reliable reasoning.

A model may reach the right conclusion by combining the supplied evidence. It may also recognize a familiar case pattern, anchor on an early hypothesis, or generate a plausible explanation after selecting an answer.

To investigate this, I created several controlled variations of the same case. Each variation changed a clinically meaningful part of the evidence while preserving the overall structure of the case.

The initial results suggested that the model reacted well to large changes but was less consistent when smaller pieces of evidence were added or removed.

I then repeated the full set of conditions three times in separate conversations.

The repeated runs changed my interpretation of the experiment.

What I Was Actually Testing

This pilot was not intended to establish a new general fact about language models.

There is already extensive work on counterfactual medical reasoning, diagnostic consistency, calibration, and sensitivity to evidence. A single case cannot compete with a large benchmark, nor can it establish clinical reliability.

The purpose was narrower and more practical:

Could evidence sensitivity itself become a meaningful technical advantage for a diagnostic-odyssey product?

The original product hypothesis was that explicitly representing timelines, confirmed absences, normal tests, contradictions, and treatment responses might allow a specialized system to reason more reliably than a general-purpose model.

Before building such a system, I wanted to test how much of that capability a frontier model already had.

The Perturbation Setup

The baseline case contained a slowly progressive pattern of weakness with evidence pointing toward a primary muscle disorder, but also an earlier interpretation suggesting motor-neuron disease.

I created variations that introduced different types of evidence:

• a highly characteristic muscle-distribution pattern;
• objective upper-motor-neuron findings;
• sensory abnormalities that complicated a single-disease explanation;
• a swallowing symptom;
• and a normal laboratory result replacing a mildly abnormal one.

The complete prompts, exact modifications, raw outputs, and scoring procedure are not included here. The goal of this post is to describe the behavioral pattern rather than publish the full evaluation protocol.

Each condition was run three times in a fresh conversation.

This produced eighteen outputs across the baseline and modified cases.

The Leading Diagnosis Was Extremely Stable

The clearest result was the stability of the model's leading diagnosis.

Across all eighteen outputs, the top-ranked diagnostic category was consistent with the condition being tested.

When the evidence remained predominantly myopathic, the model retained the muscle-disease diagnosis as its leading candidate.

When strong and reproducible upper-motor-neuron findings were added, the model consistently reversed the ordering and placed ALS first.

This reversal occurred in all three repetitions of that condition.

That matters because the model was not simply repeating its original conclusion. It recognized a syndrome-level change and updated the leading diagnosis accordingly.

At the level of the primary diagnostic decision, the model was more reproducible than I initially expected.

Strong Evidence Produced Stable Updates

The most consistent behavior occurred when the new evidence substantially changed the clinical syndrome.

Objective upper-motor-neuron findings created a clear conflict with an isolated muscle disorder. In every repetition, the model elevated ALS while preserving the evidence that still favored a myopathy.

It did not erase the contradiction.

Instead, it described a case containing two strong and competing evidence clusters:

• a selective muscle pattern and myopathic electrodiagnostic findings; and
• corticospinal findings and previously reported abnormalities suggestive of broader motor-system involvement.

This was one of the strongest parts of the model's performance.

The model did not need every finding to point in the same direction before changing its leading answer.

Highly Characteristic Evidence Also Had a Reproducible Effect

Another variation added a muscle-distribution pattern that is substantially more characteristic of the original diagnosis.

Across the three runs, the model consistently strengthened the leading muscle-disease interpretation and weakened ALS relative to the baseline.

The exact numerical scores varied slightly, but the broad diagnostic effect was stable.

This suggests that the model was capable of recognizing not only the presence of weakness, but also the diagnostic value of its anatomical distribution.

That is more sophisticated than simple symptom matching.

Smaller Effects Were Harder to Measure

The repeated runs also revealed an important limitation.

Although the leading diagnosis was stable, the ordering and scoring of secondary candidates were not.

In the unchanged baseline case, the model consistently selected the same leading diagnosis, but the position and compatibility score assigned to ALS varied considerably between runs.

The same instability appeared in the condition containing sensory abnormalities.

In every repetition, the model correctly explained that objective sensory findings weakened a pure motor-neuron interpretation. However, the exact rank and score assigned to ALS still varied.

This produced an important distinction:

The model's high-level conclusion could be stable even when the detailed structure of the differential diagnosis was not.

A single run could therefore create a misleading impression of precision.

The model might reliably say that one diagnosis is more compatible than another while being much less reliable about whether the second diagnosis should receive a score of three, four, or six—or whether it should appear second, fourth, or sixth in the list.

Repetition Changed the Interpretation

After the first run, some of the score changes appeared internally inconsistent.

A finding described as evidence against ALS could occasionally coexist with a higher ALS score than in the baseline output.

Viewed in isolation, this looked like a failure of evidence sensitivity.

After three runs, a more nuanced pattern emerged.

The average direction of the changes was generally reasonable:

• the characteristic muscle pattern weakened ALS;
• objective upper-motor-neuron findings strongly strengthened ALS;
• sensory abnormalities weakened a pure ALS explanation;
• and normalization of a relatively nonspecific laboratory result produced little diagnostic change.

The problem was not that the model had no sensitivity to evidence.

The problem was that small evidence effects were difficult to distinguish from run-to-run variation.

This means a one-point change in an ordinal score should not be treated as a stable measurement of diagnostic belief.

The scores were useful as structured language, but not as calibrated quantities.

Multimorbidity Complicated the Benchmark

The sensory variation exposed another issue—not only with the model, but with the evaluation itself.

The added sensory findings were not well explained by the leading muscle diagnosis. Yet the model continued to rank that diagnosis first.

Initially, this seemed incorrect.

But the model repeatedly proposed a two-process interpretation:

• a muscle disease explaining the selective weakness and myopathic findings;
• and a coexisting peripheral neuropathy explaining the sensory loss.

That interpretation may be clinically reasonable.

The problem was that the benchmark asked only how "compatible" each diagnosis was with the case.

It did not distinguish between:

1. whether a disease could explain an important component of the presentation; and
2. whether that disease alone could explain the complete presentation.

These are different judgments.

A future evaluation should separately measure:

• component compatibility;
• single-diagnosis explanatory adequacy;
• and the plausibility of multiple coexisting processes.

Otherwise, a sensible multimorbidity interpretation may look like a ranking failure.

Some Outputs Were More Stable Than Others

Across the repeated runs, different parts of the output showed different levels of reproducibility.

The most stable elements were:

• the leading diagnosis;
• the response to strong syndrome-level evidence;
• the recognition that the diagnosis was not pathologically confirmed;
• and the major evidence supporting the first-ranked candidate.

Less stable elements included:

• the exact score assigned to secondary candidates;
• the precise ordering below the leading diagnosis;
• whether the model considered abstention necessary in one especially characteristic variation;
• and the estimated magnitude of smaller evidence effects.

This suggests a hierarchy of reliability.

The model appeared most reliable when answering:

Which broad diagnostic explanation fits best?

It appeared less reliable when answering:

Exactly how much less compatible is each alternative?

Those are not equivalent tasks.

The "Next Best Evidence" Was Almost Too Stable

One output remained strikingly consistent across the experiment.

The model almost always selected muscle biopsy as the most informative missing evidence.

In many conditions, this was reasonable. Histopathology could distinguish a primary muscle disorder from chronic neurogenic change and could separate competing myopathies.

But the preference persisted even when ALS became the leading diagnosis.

In that setting, other information might plausibly have greater immediate value:

• contemporary multiregional electrodiagnostic testing;
• spinal imaging;
• documented clinical spread;
• or confirmation of lower-motor-neuron involvement outside the initially affected muscles.

The model's repeated preference for biopsy may represent a rational diagnostic strategy.

It may also reflect a learned narrative pattern: the model recognized the original type of case and repeatedly attempted to complete it through the same confirmatory pathway.

This produced a more interesting research question than top-diagnosis accuracy:

Is the model selecting the next piece of evidence by estimating information gain, or by completing a familiar diagnostic story?

That question remains unresolved.

What This Pilot Does—and Does Not—Show

This experiment does not demonstrate that frontier language models are clinically reliable.

It uses one diagnosis-masked case, a small number of controlled variations, and model-generated ordinal scores that are not calibrated probabilities.

It also does not establish a new benchmark result. Larger studies have already examined counterfactual reasoning, diagnostic consistency, and sensitivity to evidence at much greater scale.

What the pilot did accomplish was product discovery.

It weakened the hypothesis that a specialized product could differentiate itself simply by:

• converting a short case into a structured timeline;
• labeling findings as present, absent, normal, or unknown;
• and asking a model to rerank diagnoses when evidence changes.

The frontier model already performed much of this surprisingly well.

In particular, it handled strong evidence, characteristic anatomical patterns, missingness distinctions, and the leading diagnostic decision better than expected.

The Product Question Has Changed

The original question was:

Can a structured evidence model reason more reliably than a general-purpose LLM?

That question was too broad.

A better question is:

Which parts of a real diagnostic odyssey remain difficult even when a frontier model can reason well over a clean vignette?

The remaining opportunities are likely outside the final act of naming a disease.

They may include:

• reconstructing years of fragmented records;
• preserving the provenance of every clinical claim;
• distinguishing confirmed absence from lack of examination across multiple notes;
• resolving contradictions between clinicians and tests;
• separating one unifying diagnosis from multiple coexisting conditions;
• retrieving truly analogous confirmed cases without relying on diagnosis keywords;
• and selecting the next piece of information according to expected diagnostic value, cost, risk, and accessibility.

These problems are less visible in a polished case summary.

They are also much closer to the actual experience of patients who remain undiagnosed for years.

Current Conclusion

After one run, the model appeared both impressive and inconsistent.

After three runs, the picture became clearer.

The leading diagnostic conclusion was highly reproducible, and strong syndrome-level evidence produced stable directional updates.

At the same time:

The magnitude of smaller evidence effects, the ordering of secondary candidates, and some calibration-related outputs remained unstable.

This is not evidence that the model cannot reason.

It is evidence that "reasoning quality" is not one variable.

A model may be stable at the level of syndrome classification while unstable at the level of fine-grained differential ranking.

It may correctly describe why a finding weakens a diagnosis without assigning a reproducible numerical effect to that finding.

And it may update the leading diagnosis while continuing to recommend the same familiar confirmatory test.

For this project, the most important result was not a newly discovered weakness of GPT.

It was the elimination of an insufficient product hypothesis.

A clean clinical vignette is no longer enough.

The next step is to test whether a specialized system can outperform a frontier model on the parts of diagnostic odysseys that are genuinely messy: long records, contradictory evidence, source provenance, multimorbidity, analogous-case retrieval, and the value of the next missing piece of information.