Observatory status · historical reconstruction active · 2008–2026 · last reconciled · site record current through
California Medical Discipline Observatory
The codebook: what counts as a case, which document governs each field, and what the resulting numbers are permitted to mean
What physicians are accused of, what the evidence establishes, what consequence is imposed, and on what stated reasoning. This page is the Observatory’s landing page: project status, datasets, method, verification states and completeness. The coding rules below were written as a preregistration before any record was entered, and that preregistration is retained verbatim as Codebook V1 so the original standard cannot be quietly relaxed to fit what was later found. V1 is archived, not operative. The operative protocol is Codebook V2, which differs from V1 in one structural respect: V1 required a concluding document to exist and to have been read before any record could be created, which no register-derived observational layer can satisfy. V2 admits that layer and states, per field class, what each evidence state is permitted to support. Where V1 says the record set “contains no records” or names a candidate count, that is the preregistration speaking, not the current project — read the status panel below for where the work actually stands.
Compiled by Kanwar Partap Singh Gill, MD · last reconciled · research documentation, not legal advice.
Dataset 01 · California DUI discipline
Published · programme span 2008–2026 · verified coverage currently 2011–2026, with 2008–2010 historical backfill open. Counts are published on the registry, from one canonical source.
Matter-level reconstruction
Two provenance layers cross-validated · 22 matters confirmed independently
Independent second-coding
107 primary decisions reviewed · 0 independently second-coded · the formal second-read protocol runs in supervised review sessions
Historical corpus audit
Register-to-alert-to-document sweep, period by period · unresolved historical candidates are recorded, not dropped
Only period closed to date
January–March 2025. Every other period is partial or open for backfill; no part of the 2008–2026 programme is claimed as a census.
Next study · sanction variation
Why comparable conduct reaches different dispositions — awaits independently second-coded matters
Datasets
Where the work stands today
Open evidence gaps. Unresolved historical candidates from the September re-pass — matters whose outcome, date or existence could not be established against a primary Board document — are held in the registry’s evidence-gap register and excluded from every aggregate until a document resolves them: seven named candidates, plus fifteen historical 2008–2022 candidates disclosed but not counted, against 205 reconciled matters. A candidate is never promoted to a matter to improve a count, and never silently dropped.
The verification states, and what each one licenses
Publishing register-derived observations required separating what a Board register can support from what only the underlying decision can. The states are checked per field, not per record:
An official Medical Board Administrative Actions publication has been read and supports the displayed field: identity, published conduct description, disposition, and any effective date, probation duration or cost the register itself prints.
The underlying final decision or order has also been opened and reconciled against the register. A separate and stricter state, INDEPENDENTLY_SECOND_CODED, requires every applicable field to be re-coded from the document by a second reader; it currently applies to no matter. It is required before any field about mitigation, aggravation, findings of fact, credibility, ALJ or Board reasoning, or causal explanation is published.
Procedural route is a graded exception, stated explicitly rather than left implicit. Route is published with the source that established it, on three values: decision_derived, register_derived and unresolved. Descriptive route reporting is permitted from an explicit official source or a first-read primary decision, with that evidence state displayed on the row. Independent second coding remains required before route is used for causal explanation, multivariable interpretation, or any claim that a factor explains differences between otherwise comparable matters.
Register-supported observations may enter the observational dataset below the case-record threshold this codebook otherwise requires. They may not, by themselves, support any claim this codebook already prohibits — comparability, causation, or that a disposition was too light or too heavy.
What this is
Dataset 01 published · last reconciled
The observatory’s first published record set is matter-level: California physician DUI discipline, 2008–2026 — 205 reconciled matters, 140 of them physicians holding an active licence on a driving disposition, published as a matter-level registry with versioned CSV and JSON exports. Every matter is REGISTER_VERIFIED and 107 carry a reviewed primary decision document; 0 have completed a full codebook second read — see below. An earlier, broader alcohol/substance increment — 103 candidate records coded on first read — is archived as historical methodology: it predates the canonical matter layer, is not a co-current analytic universe, and none of its records enters any current figure.
This page is the codebook for a record set whose first increment now exists. That order is deliberate. A dataset about disciplinary outcomes is only as trustworthy as the decisions taken before collection began — what counts as a case, which document is authoritative, how a charge is distinguished from a finding, and what the resulting numbers are permitted to mean. Publishing those decisions first, in public, means they cannot be quietly adjusted later to make a pattern look cleaner than it is.
The observatory answers a question the machinery does not answer about itself. This institute has published extensively on how California physician discipline works: the complaint, the investigation, the accusation, the hearing before an administrative law judge, the standard of proof, the proposed decision, the board’s adoption or non-adoption, probation, and the reporting consequences that follow. All of that describes the apparatus. None of it says what the apparatus produces when comparable matters reach disposition.
The purpose is not to catch anyone. It is to make an institution’s own output legible to the people it regulates.
One consequence of that purpose is worth stating at the outset: the observatory does not open with the proposition that discipline is arbitrary. That claim is not established by the public record, and it is a weaker question than the one the record can actually answer. The question is why two apparently comparable matters can end in materially different consequences, and which of the visible variables — the counts charged, the facts found, the procedural route, the evidence that reached the decision-maker, the reasoning given — travel with the difference.
The unit of analysis
One final disposition is one record. Everything else in the file is an attribute of that record or a separate object linked to it.
| Object | Rule |
|---|---|
| Final disposition | The record. A single decision and order, stipulated settlement, default decision, or surrender that concludes a matter as to one licensee. |
| Accusation | An attribute, never a record. The charging document tells the dataset what was alleged and under which code sections; it establishes nothing about what occurred. |
| Proposed decision | A linked object where published. Its recommendation and its stated reasoning are coded separately from the final action, because the difference between the two is itself a research question. |
| Petition | A linked object. Reinstatement, early termination of probation, and modification petitions each produce their own dated disposition and do not overwrite the original. |
| Licensee | An entity, not a record. Multiple dispositions across time attach to one entity; the dataset counts dispositions and reports licensee counts separately, because conflating them is the most common way disciplinary statistics mislead. |
| Matter | A grouping key. Consolidated or amended proceedings that resolve together carry one matter key so that a single course of conduct is not counted twice. |
A record enters the set in one of three states. A register observation is created when an official Medical Board Administrative Actions publication has been read and supports the fields displayed from it. A matter whose underlying decision, order or packet was opened and used in reconciliation is primary-decision reviewed. The strictest state, independently second-coded, requires every applicable codebook field to be re-coded from the primary document during the formal second-read protocol. Findings about mitigation, aggravation, findings of fact, credibility, ALJ or Board reasoning, causal explanation, and charged-versus-sustained counts are reserved for independently second-coded matters. Procedural route is graded rather than reserved: it may be described from a first-read primary document or an explicit official source with its evidence state displayed, and may be used for causal or comparative inference only after independent second coding. A matter that is pending is not a record in any state.
The field dictionary
Ten field groups. Each field carries a source rule — the document the value must come from — because the same fact stated in two documents can differ, and a dataset that does not name its source of truth will silently average them.
1. Case identity
| Field | Definition and source rule |
|---|---|
| Case number | As printed on the final disposition. Where a matter carries both an agency case number and a hearing docket number, both are stored. |
| Licensee identity | Name and license number as printed. Stored because the primary document is public and citable; used in aggregate reporting only alongside the citation to that document. |
| Specialty | Recorded only where the document states it. Never inferred from practice description or from a name. |
| Dates | Accusation filed, hearing dates, proposed decision, final decision, effective date. Each stored separately; none derived from another. |
| Decision-maker | The administrative law judge where a proposed decision exists, and the board or panel that took the final action. |
| Counsel | Deputy attorney general and respondent counsel, recorded only where the document names them. Absence is recorded as absence, not as self-representation. |
2. Procedural route
| Value | What it means for the record |
|---|---|
| Stipulated settlement | Negotiated. May contain admitted facts, deemed admissions, or no factual findings at all. The dataset records which, because a stipulation without findings cannot contribute to any analysis of what conduct produces what outcome. |
| Contested hearing | Evidence taken. Produces findings and, usually, stated reasoning. The only route that reliably supports reasoning analysis. |
| Default | Respondent did not appear. Findings rest on the charging document. Coded distinctly and excluded from any comparison of evidence to outcome. |
| Surrender | Licence given up. Coded with the procedural posture at the time, because surrender during a pending matter is a different object from surrender after a decision. |
| Non-adoption | The board declined to adopt a proposed decision and decided the case itself. Coded with the proposed recommendation preserved, since the delta is the point. |
| Petition disposition | A later decision on reinstatement, modification or early termination. Linked to the original, never merged into it. |
3. Counts charged, and counts sustained
Two separate arrays, never one. The accusation supplies the first; the disposition supplies the second. The distance between them is one of the most informative quantities in the entire dataset and it is invisible in every published summary of disciplinary activity.
| Field | Rule |
|---|---|
| Counts charged | Every code section pleaded, in the order pleaded, from the accusation or amended accusation. |
| Counts sustained | Every section the disposition finds established. A stipulation that resolves without findings records an empty array and a flag, not an assumption. |
| Counts dismissed or withdrawn | Recorded expressly. A dismissed count is evidence about charging practice, not about the licensee. |
| Alternative theories | Where the same conduct is pleaded under several sections, the grouping is recorded so that one act is not counted as several findings. |
4. Conduct category
A controlled vocabulary applied to the conduct the disposition finds, not to the conduct the accusation alleges. Categories are assigned from a fixed list — quality of care, recordkeeping, prescribing, controlled substances, alcohol, criminal conviction, dishonesty, boundary or sexual misconduct, billing, impairment, out-of-state discipline, probation violation, other — and a record may carry more than one. Where the document will not support a category cleanly, the record is marked ambiguous and stays ambiguous.
5. Clinical nexus
| Value | Definition |
|---|---|
| During patient care | The conduct occurred in the course of treating a patient. |
| Related to patient care | Not during treatment, but connected to it — records, prescribing, billing, supervision. |
| Entirely off duty | No patient, no clinical setting, no professional act. |
| Patient harm found | A separate boolean. Harm is recorded only where the disposition finds it. |
| No patient involved | A separate boolean, because its absence in a document is not the same as its denial. |
6. Aggravating findings
Coded only where the decision-maker expressly treats the fact as aggravating. A fact that appears in a narrative but is never given weight is stored as present-but-unweighted. The distinction matters: a dataset that counts every unflattering detail as aggravation will find aggravation everywhere and explain nothing. Fields include prior discipline, repetition, patient harm, dishonesty during the proceeding, failure to cooperate, practice while impaired, failed prior treatment, probation violation, absence of insight, and minimisation.
7. Mitigating findings
The mirror field, with the same rule: expressly credited, or not coded as credited. Fields include no prior discipline, isolated event, elapsed time without recurrence, documented treatment, sustained abstinence, psychotherapy, independent evaluation, absence of a diagnosed disorder, restitution or correction, acceptance of responsibility, remorse, character evidence, voluntary monitoring, continued safe practice, and compliance with criminal terms. This is the group that answers the question most physicians actually ask, which is not what is the penalty for this but what has ever moved a decision-maker.
8. Outcome
| Field | Rule |
|---|---|
| Nominal outcome | Dismissal, public letter of reprimand, public reprimand, probation, stayed revocation with probation, actual suspension, surrender, revocation, reinstatement. |
| Actual suspension | Days ordered, stored as a number. A stayed revocation with no actual suspension is a materially different event from one with ninety days out of practice, and the label conceals it. |
| Stay | Whether revocation was stayed, and on what conditions. |
| Effective date | As ordered, not as decided. |
| Duration | Probation term in months. Never rounded to years. |
9. Probation burden
Probation is not one penalty, and treating it as a single category is the second most common defect in disciplinary analysis. Each order is decomposed into its obligations: biological testing and its frequency, alcohol or controlled-substance prohibitions, psychotherapy, psychiatric or medical evaluation, worksite monitor, practice monitor, solo-practice prohibition, supervision ratios, employer and patient notification, education or ethics coursework, practice restrictions by procedure or setting, tolling provisions, and the cost of monitoring where the order states it. A five-year probation with quarterly reporting and a five-year probation with a worksite monitor and a solo-practice ban are the same word and different professional lives.
10. Cost and timing
Costs requested and costs awarded, stored separately, with the payment schedule where stated. Timing is stored as the intervals the documents actually support — accusation to hearing, hearing to proposed decision, proposed decision to final action, final action to effective date — and never as an estimate of the interval before the accusation, which the public record generally does not disclose.
Severity gradation: a category is not a comparison
The hardest methodological problem in this project is not collection. It is comparability. Two records both coded alcohol-related may describe events with almost nothing in common, and any analysis that compares them as equivalents will produce a disparity that is an artefact of the coding rather than a fact about the institution.
So each conduct category carries its own gradation fields, defined before collection, and comparison happens within a gradation band rather than within a category. The bands are taken from facts the documents actually state.
| Category | Gradation fields taken from the document |
|---|---|
| Alcohol and driving-related | Measured level where stated, banded rather than averaged; single or repeat event; collision; injury to another person; refusal; concurrent substances; any finding connecting the conduct to practice or to proximity to practice; and whether a disorder was diagnosed, not diagnosed, or not addressed. |
| Quality of care | Single incident or repeated; magnitude of the departure as characterised by the decision-maker; harm found and its severity as found; whether systems or institutional factors were credited; and whether the finding rests on expert testimony, on admission, or on default. |
| Recordkeeping | Volume of records at issue; whether the deficiency affected patient care as found; whether alteration or backdating was found, which moves the record into the dishonesty category as well. |
| Prescribing and controlled substances | Number of patients; presence of a prior examination as found; whether the prescribing was for the licensee or a family member; whether diversion or personal use was found; and whether a patient death or serious injury was found. |
| Dishonesty | Whether the dishonesty was to a patient, to an institution, to a payer, to the agency, or in the proceeding itself. These are not one category and they do not behave alike. |
| Boundary and sexual misconduct | Graded on findings only, with a standing rule that the record set publishes no detail beyond what the document publishes and no narrative reconstruction of any incident. |
| Out-of-state discipline | The originating action and its severity; whether California mirrored, reduced or exceeded it; and whether the underlying conduct was retried or accepted as established. |
| Probation violation | Technical non-compliance, monitoring failure, relapse, or new clinical conduct. Four different events routinely reported as one. |
Where a document does not state the fact a gradation field needs, the field is empty and the record is excluded from that comparison. It is not imputed, and it is not filled from a news report about the same matter. An analysis run on fifty comparable records is worth more than one run on four hundred incomparable ones.
Eight errors the coding rules exist to prevent
Every rule above was written against a specific way this kind of dataset goes wrong. Naming them is more useful than asserting rigour.
Counting allegations as conduct
An accusation is a pleading. Coding charged counts as findings inflates every category and turns a charging pattern into an apparent behaviour pattern.
Averaging stipulations with hearings
Settled matters produce different documents, different admissions and different selection. Pooling them silently makes the reasoning analysis meaningless.
Counting licensees as cases
A licensee with three dispositions is not three physicians. Both counts are legitimate; publishing one while implying the other is not.
Reading probation as one thing
Duration without obligations describes almost nothing about the burden actually imposed.
Treating a stayed revocation as a revocation
The label is severe and the operative effect may be probation. The dataset stores both and reports the effect.
Inferring a diagnosis
Where a document records that no disorder was diagnosed, that is a finding about the record, not a clinical opinion, and it is never converted into one.
Ranking individual decision-makers
Volume differences, assignment practices and case-mix make individual comparisons unreliable and the exercise punitive rather than informative.
Reporting a departure as favouritism
A disposition below a published range is a fact. The reason for it is either stated in the document or unknown, and unknown is published as unknown.
The denominator, and what this will not claim
Published dispositions are the visible end of a funnel. Complaints closed without action, investigations that end quietly, and accusations withdrawn before hearing are largely not published. That has a hard consequence: this dataset can describe the distribution of outcomes among published dispositions, and it cannot by itself produce a rate.
So the observatory will report counts and distributions with their base stated on the same screen, and where a rate is genuinely available — because the agency publishes the denominator itself — the source of that denominator is named. Where it is not available, the dataset says so rather than substituting a plausible one. Case counts without denominators is the standing analysis of that failure mode, and why enforcement data need context is the reason this section exists before any chart does.
Three further claims this record set will not make: that a disposition was correct or incorrect; that a licensee did or did not do something the document does not find; and that any observed association is a cause. The first is not a research question, the second is defamation dressed as data, and the third is the failure that discredits empirical work in adjacent fields.
Comparing a disposition with a published range
The comparison at the centre of the project is mechanical to describe and delicate to do. For each record: locate the published guideline range applicable to the sustained counts; record the range as published, with its date; record the actual disposition; compute the direction and size of any departure; and record the reason for the departure as stated in the document, quoting the operative sentence where one exists.
| Value | Rule |
|---|---|
| Applicable range | From the guidelines in force at the time of the disposition, cited with their edition and date. A later edition is never applied retrospectively. |
| Actual disposition | The operative outcome, decomposed as in field group 8. |
| Departure | Direction and magnitude, computed rather than characterised. No adjectives. |
| Stated reason | The decision-maker’s own reasoning, or the value none stated. A record with no stated reason contributes to the count of unexplained departures and to nothing else. |
The reason this matters is not that guidelines are being flouted. It is that a published range communicates an expectation to physicians, counsel and the public, and if actual dispositions sit routinely elsewhere, the expectation is being set by a document that does not describe the practice. That is a finding about transparency rather than about any individual case.
A guideline range does not mechanically determine an outcome. What reaches the decision-maker may matter more than the label on the charge — and that is a testable proposition, not a complaint.
How a record enters, in order
Collection happens in dated increments rather than continuously, because an increment has a boundary that can be described and audited, and a continuous trickle does not.
Sweep
The issuing body’s own publication surfaces are enumerated for the increment window, and the window is recorded with its start and end.
Candidate list
Every disposition in the window is listed before any is read, so that selection cannot happen by reading order or by interest.
Retrieval and preservation
The primary document is retrieved and preserved as retrieved, with its retrieval date, because agency publication windows expire and documents are removed on schedule.
Read
The document is read end to end before any field is entered.
Code
Fields are entered under the rules above, with ambiguity recorded as ambiguity.
Second read
The record is checked against the document again. Failures are held rather than corrected from recollection.
Increment publication
The increment is published with its window, its candidate count, its coded count, and its held count. A held record is disclosed as held, so the difference between what existed and what was coded is visible.
The removal problem deserves its own sentence, because it shapes what this record set can ever be. Agencies publish disciplinary documents for defined periods and then take them down. A dataset that only reads what is online today therefore has a horizon that moves forward and quietly discards its own past. Preserving each primary document as retrieved, with its retrieval date, is what allows a record to remain auditable after the source page is gone — and it is why the third life of a disciplinary record, the one that outlasts the agency’s own publication window, is itself a program item rather than an afterthought.
Verification protocol
Five rules, applied per record. Where a rule admits a graded exception, the grade is published with the record rather than left to the reader.
Primary document or nothing
Every field is taken from the issuing body’s own published document. A news report, a summary, a database extract or a third-party compilation is a lead that tells the collector where to look. It is never a source.
Read, not scraped
A record enters the set after the document has been read end to end. Field extraction that never sees the reasoning cannot distinguish a finding from a recital.
Dated at the record level
Each record carries the date its primary document was consulted. Documents are amended and removed; a record without a consultation date cannot be audited.
Ambiguity preserved
Where a document will not support a value, the field records ambiguity. Forced values are the mechanism by which a coding frame manufactures the pattern it was looking for.
Reconciliation review before publication
Every record is checked against its source a second time by the primary coder, and checked for cross-layer and same-layer duplication, before it appears in any aggregate. A record that fails that check is held, not corrected from memory. This is a distinct and weaker stage than INDEPENDENTLY_SECOND_CODED, which requires a different reader; the two are never reported as one number.
Publication and permanence
The record set is append-only. New dispositions are added; later petitions and modifications are appended as their own dated entries; nothing is overwritten. Where a value in a published record turns out to be wrong against its primary document, the value is fixed in place and the record’s consultation date advances — the file is a description of documents, and a description that has been made accurate needs no annotation about having been inaccurate.
Each record and each aggregate view has a permanent address. A view that is superseded by a better one states that it is superseded and names its successor, and continues to serve at its own address. The underlying data layer is publicly fetchable, so a reader who doubts a chart can read the file it came from, which is the same commitment the rest of this site’s figures already carry — see research products and the methodology.
Fairness rules
This dataset concerns identifiable people who are the subject of public documents. That is a legitimate object of study and it carries obligations that go beyond accuracy.
Nothing enters the set that the issuing body has not published. No investigative material, no correspondence, no clinical information.
The interface distinguishes the two everywhere they appear together, not only in the field names.
A disciplinary document is not a medical record. Diagnoses, impairment and prognosis are recorded only as findings the document makes.
Where a document is amended, superseded or withdrawn by the issuing body, the record follows within the next collection increment and says which document it now rests on.
Distributions describe a record set. They are never presented as an expected outcome for a named person or a pending matter.
Corrections are made against the primary document, not by negotiation. Where a licensee disputes what a document says, the dispute belongs with the issuing body, and the dataset will record the amended document when it exists.
What it answers once populated
The codebook exists so that these questions can be answered from one record set rather than from twelve incompatible ones. Each is a numbered item in the research program.
| Item | What it takes from the record set |
|---|---|
| 002 · Sanction variation | Distribution of outcomes within a single conduct category, with route, findings, evidence and stated reasoning attached. |
| 003 · One conduct, different penalties | Matched comparisons inside a category, by graded severity fields rather than by category alone. |
| 004 · Penalty lab | Applicable range, disposition, departure and stated reason, per record. |
| 006 · Board versus judge | Proposed recommendation against final action, with the delta and the stated basis. |
| 007 · Defence variables | Representation, route, witness and evidence fields against outcome — reported with the selection problem stated, not hidden. |
| 008 · Charging discretion | Counts charged against counts sustained, and which combinations recur. |
| 010 and 011 · Mitigation and aggravation | Frequency of expressly credited factors, and their co-occurrence with outcome. |
| 012 · Probation burden | Obligation counts and composition, independent of nominal duration. |
| 013 · Cost | Requested against awarded, by route. |
| 049 · Delay | Interval distributions across the stages the documents disclose. |
None of those items becomes a published finding until the record set can carry it. Until then each stays what it currently is on the program page: a question with a state.
What it will never be
A lawyer-finder
Counsel identity is recorded where documents name it, for the study of procedural route. The dataset will not rank, recommend or imply the effectiveness of any representative.
A prediction engine
No projected outcome for a pending matter, and no calculator that invites a physician to estimate their own exposure. The distribution of past dispositions is not a forecast about a person.
A scoreboard for individuals
No rankings of judges, prosecutors or panel members. Case mix and assignment make such comparisons unreliable and their purpose punitive.
A campaign
Where the record supports a criticism of the system, it will be published as an analysis with its evidence. Where it does not, no criticism will be published for the sake of a position.
A substitute for the document
Every record links to the document it describes. The dataset is an index to primary sources and it does not ask to be believed over them.
A dataset that cannot state its own limits is not evidence. It is decoration with numbers on it.
Published analysis this rests on
The doctrinal groundwork is already published here, and the record set is designed to be read against it rather than instead of it.
how the hearings work · the proposed decision and what happens to it · the standard of proof · why a complaint is not a finding · investigation against public accusation · discovery · accountability against punishment · fitness for duty against discipline · why enforcement data need context · case counts without denominators · correcting public errors · a public accountability checklist · the due-process proposal · the research program
Data generation DUI-2026-09-09-R9 · last reconciled · every figure comes from one record, stated the same way everywhere