Structural admissibility (clinical shorthand → model output)
A clinically valid, socially trusted instruction set — and what changes when a model becomes the inference engine. This page contrasts three outputs from the same source note under one simple prompt: “Explain this message clinically.”
The unconstrained model reads as complete and executable because it silently adds agency, decisions, and temporal resolution. The contained run does not “complete” the note — it preserves what is stated and reports what is not stated.
| Output | What it does | Operational risk |
|---|---|---|
| Source note | Shorthand record. Relies on professional reconstruction. | Depends on silent context |
| Unconstrained LLM | Produces a fluent plan-like narrative; may add missing structure. | May invent agency / decisions |
| (In)Canon-contained | Quotes only what is written; reports absences that require human clarification. | Absence preserved |
This page originates from a discussion with a large NHS trust about (In)Canon and a potential pilot on live clinical data.
Rather than proceeding with a pilot, a member of the trust’s digital innovation team offered a real, de-identified piece of routine clinical shorthand and asked:
“What would (In)Canon make of this note?”
The note was provided exactly as authored, with no additional context, clarification, or background information. It was not constructed for demonstration purposes.
After reviewing the outputs, the reply was:
“Interesting to see my note analysed in such detail!”
They added that this kind of analysis could be particularly useful in situations where all required elements need to be explicitly present, such as incident reports or project plans, and noted that this is typically achieved by structured capture systems using multiple predefined fields (for example, Datix).
They also raised a further consideration: whether inference itself is a problem, or rather where it becomes a problem. They noted that inference is a routine part of communication — it risks errors, but it also speeds things up.
Inference itself is not the danger. Silent inference is.
In everyday clinical communication, inference is unavoidable and often beneficial. Humans routinely resolve missing elements through shared context, professional norms, and direct clarification.
The risk emerges when inferred structure is silently carried forward and treated as authoritative, reusable, or machine-actionable — particularly when generated by an unconstrained model that appears confident and complete.
This demonstration exists to make that boundary visible. (In)Canon does not argue against inference. It identifies the point at which inference must stop being silent and start being explicit.
This page shows what happens when a short piece of routine clinical shorthand is processed by a LLM in three different ways with one simple instruction: “Explain this message clinically.” No qualifiers. No constraints. No warnings. No structural framing.
1) Structural read
- Shows what is explicitly stated.
- Shows what is not stated (requires human clarification).
- Does not add structure or “complete” the note.
2) Unconstrained LLM
- Often sounds reasonable.
- Can introduce agency, decisions, and temporal resolution.
- Can create an impression of completeness.
3) (In)Canon-contained LLM
The issue is not whether the unconstrained output sounds plausible. The issue is where decisions, ownership, and authority silently appear even though they were never written. The contained run prevents inference and preserves absence.
Called pt Discussed interactions Happy to switch - eg to Biktarvy knows that we know he had an odd feeling in throat before with it - but may not be due to the pill Dovato also an option? Explained that may need increase in steroid cover when switches - eg more Trimbow Decide on regimen at MDT on Tues appt made for review next Thurs with DB
**Clinical Summary:** I spoke with the patient about potential medication interactions. The patient is open to switching to a new medication, such as Biktarvy, although they experienced an unusual sensation in their throat previously, which may not have been related to the medication. We also discussed Dovato as another possible option for treatment. I informed the patient that if they switch medications, they may require an increase in their steroid coverage, specifically with Trimbow. A decision regarding the best treatment regimen will be made during the multidisciplinary team meeting on Tuesday. The patient has an appointment scheduled for a follow-up review next Thursday with Dr. DB. **Plan:** 1. Discuss medication options (Biktarvy and Dovato) at the MDT meeting on Tuesday. 2. Monitor the patient’s response to any medication switch, including the need for increased steroid coverage. 3. Follow up with the patient next Thursday for further evaluation.
A) Explicitly stated (verbatim quotes only) - "Called pt" - "Discussed interactions" - "Happy to switch - eg to Biktarvy" - "knows that we know he had an odd feeling in throat before with it - but may not be due to the pill" - "Dovato also an option?" - "Explained that may need increase in steroid cover when switches - eg more Trimbow" - "Decide on regimen at MDT on Tues" - "appt made for review next Thurs with DB" B) Not stated / requires human clarification - The specific interactions discussed. - The reason for the odd feeling in the throat. - The details of the steroid cover increase. - The specific regimen to be decided at MDT. - The context of the review appointment with DB. - The implications of switching to Biktarvy or Dovato. - The timing and participants of the MDT meeting.
This demonstration evaluates a single structural invariant — Actor — to make the risk visible without over-asserting. (In)Canon supports additional invariants (e.g., time, authority, decision finality), but these were intentionally out of scope for this demonstration.
Each item below is an actionable commitment latched from the note. Actor shows whether responsibility is explicitly stated.
| Latched commitment (verbatim) | Actor explicitly stated? |
|---|---|
| Called pt | ✗ missing |
| Discussed interactions | ✗ missing |
| Happy to switch - eg to Biktarvy | ✗ missing |
| Explained that may need increase | ✗ missing |
| in steroid cover when switches - eg | ✗ missing |
| Decide on regimen at MDT on Tues | ✗ missing |
| appt made for review next Thurs with | ✗ missing |
{'rc_count': 7, 'missing_actor': 7, 'notes': ['This notice is non-evaluative.', "Missing fields mean 'not explicitly stated'."]}
These statements may sound reasonable. The issue is that they were not explicitly authored in the source note.
2. Monitor the patient’s response to any medication switch, including the need for increased steroid coverage.
Introduces implied agency/actions not explicitly authored in the note.
I spoke with the patient about potential medication interactions. The patient is open to switching to a new medication, such as Biktarvy, although they experienced an unusual sensation in their throat previously, which may not have been related to the medication. We also discussed Dovato as another possible option for treatment. I informed the patient that if they switch medications, they may require an increase in their steroid coverage, specifically with Trimbow.
Converts options/pending MDT decision into an apparently settled plan or consent state.
A decision regarding the best treatment regimen will be made during the multidisciplinary team meeting on Tuesday.
Turns relative timing shorthand into resolved scheduling/ownership.
The patient has an appointment scheduled for a follow-up review next Thursday with Dr. DB.
Turns relative timing shorthand into resolved scheduling/ownership.
1. Discuss medication options (Biktarvy and Dovato) at the MDT meeting on Tuesday.
Turns relative timing shorthand into resolved scheduling/ownership.
3. Follow up with the patient next Thursday for further evaluation.
Turns relative timing shorthand into resolved scheduling/ownership.
**Clinical Summary:**
Adds structured sections that can create an impression of completeness.
**Plan:**
Adds structured sections that can create an impression of completeness.
For each flagged statement, we extract a phrase and check whether it appears verbatim in the original note.
| Category | Sought phrase | Sought phrase in note? |
|---|---|---|
| INFERRED ACTOR/AGENCY | 2 | NOT IN NOTE (verbatim) |
|
Model output line
2. Monitor the patient’s response to any medication switch, including the need for increased steroid coverage.
Nearest excerpt from original note
Called pt
Discussed interactions
Happy to switch - eg to Biktarvy
knows that we know he had an odd
feeling in throat before with it - but
may not be due to the pill
Dovato also an option?
Explained that may need increase
in steroid cover when switches - eg
more Trimbow
Decide on regimen at MDT on Tues
appt made for review next Thurs with
DB
|
||
| DECISION COMPLETION | I spoke with the patient about potential medication interactions | NOT IN NOTE (verbatim) |
|
Model output line
I spoke with the patient about potential medication interactions. The patient is open to switching to a new medication, such as Biktarvy, although they experienced an unusual sensation in their throat previously, which may not have been related to the medication. We also discussed Dovato as another possible option for treatment. I informed the patient that if they switch medications, they may require an increase in their steroid coverage, specifically with Trimbow.
Nearest excerpt from original note
Called pt
Discussed interactions
Happy to switch - eg to Biktarvy
knows that we know he had an odd
feeling in throat before with it - but
may not be due to the pill
Dovato also an option?
Explained that may need increase
in steroid cover when switches - eg
more Trimbow
Decide on regimen at MDT on Tues
appt made for review next Thurs with
DB
|
||
| TEMPORAL RESOLUTION | A decision regarding the best treatment regimen will be made during the multidis… | NOT IN NOTE (verbatim) |
|
Model output line
A decision regarding the best treatment regimen will be made during the multidisciplinary team meeting on Tuesday.
Nearest excerpt from original note
Decide on regimen at MDT on Tues
|
||
| TEMPORAL RESOLUTION | The patient has an appointment scheduled for a follow-up review next Thursday wi… | NOT IN NOTE (verbatim) |
|
Model output line
The patient has an appointment scheduled for a follow-up review next Thursday with Dr. DB.
Nearest excerpt from original note
appt made for review next Thurs with
DB
|
||
| TEMPORAL RESOLUTION | 1 | NOT IN NOTE (verbatim) |
|
Model output line
1. Discuss medication options (Biktarvy and Dovato) at the MDT meeting on Tuesday.
Nearest excerpt from original note
Happy to switch - eg to Biktarvy
Dovato also an option?
Decide on regimen at MDT on Tues
|
||
| TEMPORAL RESOLUTION | 3 | NOT IN NOTE (verbatim) |
|
Model output line
3. Follow up with the patient next Thursday for further evaluation.
Nearest excerpt from original note
appt made for review next Thurs with
DB
|
||
| FOLLOW-UP STRUCTURE ADDED | **Clinical Summary:** | NOT IN NOTE (verbatim) |
| FOLLOW-UP STRUCTURE ADDED | **Plan:** | NOT IN NOTE (verbatim) |