Skip to content
arbadePublic

About

Adversarial AI medical case-review board for Claude Code — blind specialist rounds, verified citations, dissent-preserving minutes, deterministic DICOM-to-3D. Not a medical device.

Topics

Resources

Stars

3 stars

Watchers

0 watching

Forks

Latest commit

 

History

9 Commits

Folders and files

Repository files navigation

Konsil — an adversarial AI medical case-review board for Claude Code

Sagittal PD knee MRI cine sweep produced by Konsil's deterministic DICOM pipeline    Rotating 3D femur surface reconstructed deterministically with TotalSegmentator from the same MRI

Konsil viewer: slice-focus on the femur with the segmentation mask overlaid on MPR slices and the radiologist's written finding quoted in the sidebar

All animations were produced by Konsil's deterministic imaging pipeline from the author's own anonymized knee MRI (shared with consent): top left — windowed slice cine from raw DICOM; top right — 3D bone surface segmented by TotalSegmentator and meshed with marching cubes; bottom — the shipped viewer slice-focusing a structure (⌖): crosshair at its centroid, segmentation mask over the MPR panes, and the radiologist's written finding quoted beside it. No AI model interpreted a single pixel to make these.

MIT license runs 100% local AI never interprets pixels in board reasoning verified citations

⚠️ What this is not. Konsil is not a medical device, not a diagnostic tool, and not medical advice. It produces preparation and second-opinion material to review with a qualified clinician. It must never be used in emergencies. Every design decision below exists to keep those statements true. See Safety & regulatory posture.

Konsil (Turkish medical vernacular for a multi-physician case conference) convenes a board of independent AI specialist agents over a patient's documented case — history, lab values, and the radiologist's written reports — and produces dissent-preserving, citation-verified board minutes. It is packaged as a Claude Code plugin: commands, skills, and agent charters, all plain Markdown you can read and change.

It is built around one uncomfortable, well-replicated fact: panels of agreeing experts and confident single answers are exactly how both humans and LLMs fail in medicine. Konsil's answer is structural — blind rounds, a professional devil's advocate, a can't-miss auditor, a cost steward, a dissent ledger, and a four-auditor release gate.


Table of contents


Why it exists (the evidence behind the design)

Each core mechanism answers a documented failure mode. This is the part of the README you should read even if you skip everything else.

Design mechanism Failure mode it answers Evidence base
Blind Round 1 — each specialist works in an isolated context, unaware the others exist Multi-agent "debate" often underperforms a single strong model via sycophantic consensus collapse; role-prompted specialists from one context are one model in costume Multi-agent medical benchmark literature (e.g., MedAgentsBench; NeurIPS 2025 multi-agent evaluations)
Challenger (devil's advocate), Checklist Officer (can't-miss audit), Steward (cost/sequencing) as fixed structured roles What demonstrably adds value is orchestration + tools + structured adversarial roles — not free-form agent discussion Tool-using / guideline-anchored agent studies vs. debate studies
Dissent Ledger — disagreements recorded verbatim, never smoothed A confident singular verdict is the most automation-bias-inducing format possible: wrong AI advice measurably drags human experts' accuracy down Radiology automation-bias experiments (Radiology, 2023)
In-session citation verification — an unverified citation does not exist LLM citation fabrication (measured 18–29% in unguarded systems); even good tool-using agents emit ~23% irrelevant citations Citation-accuracy evaluations
Missing Data Gate — the board refuses to convene on thin input Model accuracy collapses on unstructured, self-reported histories (≈95% on clean vignettes → <35% with real user-written input) Real-world input-quality studies (Nature Medicine 2024; user-relay studies)
Text only; AI never reads pixels in board reasoning (sole exception: the consent-gated OFF-PROTOCOL read below, which never feeds a board) General VLMs score 8–35% on unselected clinical images with high hallucination rates; adding images to informative text can reduce accuracy Multimodal radiology evaluations 2024–2026, and our own measurement below
OFF-PROTOCOL read gate — AI image reading exists only behind a typed-consent gate, forced labeling, a mandatory bottom disclaimer, and a dedicated Reviewer Gate auditor Automation bias: a fluent AI "normal görünümde" gets read as clearance to skip the radiologist; and VLM finding-fabrication Automation-bias literature; our own shoulder-001 precedent read (found the burned-in "L" marker, closed a laterality gap — a technical fact, not a diagnosis)
Reviewer Gate — four independent auditors before release Drafting agents smooth disagreement, launder uncertainty markers, and (rarely but really) hallucinate author names All three caught live in blinded development runs — see below

Every citation an agent relies on is retrieved and verified inside that agent's own session, and the search logs stay in the run transcripts — so any claim in a board's minutes can be chased back to the query that found it.

What it does / what it refuses to do

Does:

  • Builds a structured, source-mapped case file from a folder of mixed documents (report PDFs, lab printout photos, discharge summaries, pathology, DICOM discs) — with a timeline table, computed interval changes, LOINC-mapped labs checked against printed reference ranges, and inconsistency flags (it recomputes what can be recomputed; it has caught real intra-report contradictions).
  • Convenes a blind, adversarial board with live, verified literature retrieval, and outputs board minutes: consensus with per-member confidences, a dissent ledger, a can't-miss audit, a cost-aware staged next-step plan, questions for the treating physician/radiologist, and a missing-data list.
  • Maintains longitudinal cases: new results re-convene the board in delta mode ("what changed, at what rate, does it alter the prior assessment").
  • Converts DICOM discs into rotatable 3D scenes and slice views — deterministically (dcm2niix → TotalSegmentator → marching cubes → NiiVue). Structured outputs (structure names, volumes, laterality) may inform the board as text.
  • On the user's explicit, logged request only: an OFF-PROTOCOL observational AI read (deterministically rendered input, region-by-region protocol, confidence on every observation, mandatory disclaimer) — orientation and technical-fact gap-closing, never diagnosis.
  • Runs a --solo baseline mode so you can measure whether the board beats a single agent on your own cases.

Refuses:

  • To let any language/vision model interpret medical image pixels as part of a board's reasoning.
  • To run the OFF-PROTOCOL read path implicitly: it exists only behind /konsil:okuma, an exact typed consent sentence, and output that can never close the formal-report requirement.
  • To output a singular confident verdict ("the diagnosis is X").
  • To convene without intake, or to silently drop [UNSOURCED] / [ESTIMATE] markers.
  • To address treatment instructions to a patient. Output is framed to the treating clinician, always.

Measured results

The numbers below were measured during development, not projected.

The safety machinery, tested against the system itself

In blinded end-to-end development runs, the audit layers repeatedly caught the system's own failures — which is their job:

  • The Reviewer Gate caught two hallucinated author names (the PMIDs and figures were correct — only the names were invented), one silently-smoothed disagreement that had been dropped from the dissent ledger, and a composite quote mislabeled "verbatim."
  • A crashed sub-agent left behind a stale draft claiming verifications that never happened; the retry agent detected the stale draft and rewrote it, re-verifying every load-bearing citation itself.
  • Most instructive: the Challenger once attacked a specialist consensus with a methodologically sound, evidence-cited argument that happened to point away from the truth — and the dissent-preserving design kept the correct position alive on the question list, where the planned confirmatory testing would have recovered it. A good devil's advocate can be locally wrong while making the plan globally safer.

Why the no-pixels rule is architecture, not a disclaimer

In development we also scored experimental AI image observations against an independent radiologist's written report. The pattern matched the published literature exactly: reliable on negative/normal statements and on deterministic pipeline outputs, weak at characterizing what it saw, and blind to subtle findings. The system's calibration held — every image observation had been labeled unverified and converted into questions for the radiologist rather than findings, and the "no AI observation may justify any intervention without an independent report correlate" rule prevented a plausible-looking cascade toward an intervention the report showed to be unwarranted.

The rule stands, untouched, as the default — the AI on pixels — never in board reasoning badge above states exactly what holds: no board ever reasons over AI pixel interpretation. What now exists beside it is a single OFF-PROTOCOL exception: an observational read available only through /konsil:okuma behind an exact typed consent sentence, built as a separate skill with its own charter and its own auditor — the exception is a separate door with its own lock, not a hole in the wall. Its output is force-labeled, closes with a mandatory unmodifiable disclaimer, and can never satisfy the formal-report requirement or clear the Missing Data Gate's "official radiology report missing" flag; every warning in Safety & regulatory posture applies to this path in full. And in the same measured-not-projected spirit: the shoulder-001 precedent read (it found the burned-in "L" laterality marker and closed a laterality gap — a technical fact, not a diagnosis) is the single data point behind the interpretive block until the pre-registered D1 experiment completes; the full decision record is in the discovery document.

Architecture

The board pipeline

flowchart TD
    A[/"Case documents<br/>(PDFs, lab photos, reports, DICOM disc)"/] --> B["case-intake<br/>timeline · LOINC labs · inconsistency flags · source map"]
    B --> C{"Missing Data Gate<br/>enough to convene?"}
    C -- "no" --> Q["Targeted anamnesis questions<br/>to the user"] --> B
    C -- "yes / proceed-with-gaps" --> D["Roster selection<br/>2–4 case specialists + 3 fixed officers"]
    D --> E1["Specialist A<br/>(isolated context)"]
    D --> E2["Specialist B<br/>(isolated context)"]
    D --> E3["Specialist C<br/>(isolated context)"]
    D --> E4["Checklist Officer<br/>(can't-miss audit)"]
    E1 & E2 & E3 --> F["Round 1: independent,<br/>blind first assessments<br/>(verified citations only)"]
    E4 --> F
    F --> G["Challenger<br/>gets case + ALL Round-1 transcripts;<br/>attacks the convergence: base rates ·<br/>alternatives · data quality · omissions · harm"]
    F --> G2["Checklist coverage audit<br/>every can't-miss item graded:<br/>addressed / ruled out / SILENTLY IGNORED"]
    G --> H["Round 2: challenges routed back<br/>to the SAME specialist agents<br/>(context preserved) →<br/>concede / rebut / revise + dissents"]
    G2 --> H
    G & H --> I["Steward<br/>gets case + Round 1 + challenges:<br/>discrimination-per-cost · staged plan ·<br/>rejected actions · cascade warnings"]
    H & I --> J["Minutes drafter (separate agent)<br/>quote, don't launder ·<br/>markers preserved · no verdict"]
    J --> K{"Reviewer Gate — 4 parallel auditors<br/>A citations · B dissent ·<br/>C safety · D completeness"}
    K -- "any FAIL" --> L["Fix from transcripts only<br/>(if info is absent, say it's absent) ·<br/>re-run failed checklist with a FRESH auditor ·<br/>may trigger Round-3 re-consultations"] --> K
    K -- "4 × PASS" --> M[/"Board Minutes<br/>consensus + per-member confidences ·<br/>DISSENT LEDGER · can't-miss audit ·<br/>staged plan · questions · sources"/]
    N["memlog.md — append-only:<br/>every phase, spawn and decision"]:::side
    classDef side stroke-dasharray: 3 3
Loading

How the agents actually talk

The flowchart shows phases; this shows the conversation. Specialists are real, separate agent sessions — Round 2 resumes the same agents so their context (and accountability) carries over:

sequenceDiagram
    autonumber
    participant U as User
    participant O as Orchestrator
    participant S as Specialists xN blind
    participant CL as Checklist Officer
    participant CH as Challenger
    participant ST as Steward
    participant D as Minutes Drafter
    participant G as Reviewer Gate x4

    U->>O: /konsil:board case-id
    O->>O: verify intake_status · open memlog
    par Round 1 — blind, parallel
        O->>S: case file + specialty only (no other member mentioned)
        S->>S: live literature retrieval + in-session verification
        S-->>O: independent first assessment (verbatim, archived)
    and
        O->>CL: case file
        CL-->>O: can't-miss set · orphaned findings · time-sensitivity
    end
    O->>CH: case + ALL Round-1 transcripts
    CH-->>O: numbered challenges + underweighted alternatives + no-viable-challenge list
    O->>CL: Round-1 transcripts
    CL-->>O: coverage audit — items the board silently ignored
    O->>S: your challenges + your assigned audit gaps (same agent, context intact)
    S-->>O: concede / rebut / revise · final confidence · dissent statements
    O->>ST: case + Round 1 + challenger report
    ST-->>O: staged plan · deferred-rejected actions · cascade warnings
    O->>D: all transcripts + minutes template
    D-->>O: draft minutes (quote, don't launder)
    par four parallel audits
        O->>G: minutes + transcripts (one checklist each)
        G-->>O: PASS / FAIL + exact required fixes
    end
    loop until 4 x PASS
        O->>O: fix from transcripts only (may re-consult a member: Round 3)
        O->>G: re-run failed checklist — FRESH auditor
    end
    O-->>U: minutes.md — consensus, dissent ledger, staged plan, questions
Loading

Case lifecycle

stateDiagram-v2
    [*] --> Intake: /konsil-case
    Intake --> Gate: case.md built
    Gate --> Intake: gaps → targeted anamnesis questions
    Gate --> Ready: intake complete (or user accepts gaps)
    Ready --> Board: /konsil-board
    Ready --> Solo: /konsil-board --solo (baseline)
    Board --> Minutes: Reviewer Gate 4x PASS
    Solo --> Minutes
    Minutes --> Delta: new documents → /konsil-update
    Delta --> Board: re-convene in delta mode (interval changes lead)
Loading

The roster

Every board seats 2–4 case specialists (chosen from the problem list: e.g., nephrology, endocrinology, orthopaedics…) plus three fixed officers who sit on every board regardless of the case:

Officer Charter (one line) Anti-pattern it kills
Challenger (agents/challenger.md) Steelman the leading hypothesis, then attack it: base rates, underweighted alternatives, single-measurement load-bearing data, omissions, intervention harm. Must declare "no viable challenge" where a position survives — manufactured dissent is prohibited. Anchoring, premature closure, groupthink
Checklist Officer (agents/checklist.md) Aviation-checklist audit: build the case's can't-miss set, grade every item addressed / ruled out with evidence / silently ignored, hunt orphaned findings (abnormalities with no disposition) and time-sensitive items. The interesting-case bias; silent omissions
Steward (agents/steward.md) Cheapest, least invasive next test that best discriminates the live differential; staged conditional plan; cascade-risk warnings; "an action whose every outcome leads to the same next step is deferred by definition." Over-testing, incidentaloma cascades, cost blindness

Round-1 isolation is real isolation: separate agent contexts, prompts that do not mention other members' existence. And when isolated specialists nevertheless converge, the Challenger's charter demands the obvious skeptical move: convergence under shared sources and shared framing is graded as one line of evidence, not three.

Evidence rules: a citation must be earned

flowchart LR
    A["Agent wants to make<br/>a clinical claim"] --> B{"Retrieve source<br/>Europe PMC / PubMed /<br/>guideline body"}
    B -- "found" --> C{"Verify:<br/>title + year match?<br/>figure present in<br/>retrieved text?"}
    C -- "yes" --> D["Cite:<br/>[Author Year, Journal — PMID]<br/>+ log search string"]
    C -- "no" --> E["State as reasoning,<br/>mark [UNSOURCED]"]
    B -- "not found" --> E
    E --> F["Markers are protected:<br/>Reviewer Gate FAILs any draft<br/>that silently drops them"]
Loading

The full protocol — source hierarchy, API cookbook with curl one-liners, forbidden sources (and why), failure handling — is skills/literature-protocol/SKILL.md. In development gate runs, auditors re-resolved every cited PMID and found zero fabricated identifiers or figures — while catching wrong author names, which is exactly the failure class the gate exists for.

The deterministic imaging pipeline

flowchart LR
    A[/"DICOM disc"/] --> B["dcm2niix<br/>→ NIfTI volume"]
    B --> C["TotalSegmentator<br/>(default Apache-2.0 task)<br/>→ per-structure masks"]
    C --> D["marching cubes + decimation<br/>→ STL meshes"]
    C --> E["structures.json<br/>names · volumes · laterality"]
    B --> F["NiiVue browser scene<br/>volume render · MPR slices ·<br/>mesh overlay · rotate/zoom"]
    D --> F
    E -- "as TEXT only" --> G["The board"]
    F -.->|"communication aid<br/>for humans"| H[("Patient / family /<br/>clinician")]
    style G stroke-dasharray: 5 5
Loading

The hard rule: the board consumes structures.json and written reports — never screenshots, never pixels. The 3D scene exists so a human can see where the radiologist's written findings live; interpretation stays with the written report. The hero GIFs at the top of this README are this pipeline's actual output.

Viewer gallery — everything the browser scene does

All of the below is the shipped viewer.html / pipeline output on two consented studies — the author's own knee MRI and a family member's abdominal CT (shared with the patient's consent; DICOM identity fields anonymized, no burned-in identity in any frame, and genital anatomy excluded from every published visual). Deterministic, local, zero AI on pixels:

Volume render rotating 3D volume render — GPU raycast of the raw volume, drag-to-rotate in the browser
Clip plane sweeping through the volume render Clip plane — slider cuts into the render to expose interior anatomy
Segmented femur mesh rendered solo Per-structure mesh — TotalSegmentator's femur mesh rendered on its own (solo ◐); the same output feeds the board structure names, volumes and laterality as text
Multiplanar reconstruction with moving crosshair and segmentation overlay MPR (multiplanar) view — synced sagittal/coronal/axial panes with crosshair; the focused structure's segmentation mask is overlaid on the slices
Series switcher cycling four MR sequences Series switcher — Sagittal T1 / Sagittal PD / Coronal PD-T2 / Axial PD from one DICOM disc
Slice-focus on a structure with the written-report annotation panel Slice-focus + report annotations — ⌖ jumps the MPR crosshair to the structure (or to the report's described region) and the sidebar quotes the radiologist's written findings for it (annotations.json); findings with no segmented structure are listed as "not localizable"
Rotating segmented abdominal organ set from a non-contrast CT Full-organ scene (abdominal CT) — all TotalSegmentator structures with per-organ show/hide, solo and Turkish labels; 78 structures from one non-contrast CT
Report marker: the written finding's region highlighted on the kidney with an approximate 3D label Report markers — the written report's own location words ("orta kortikal alan") map to the structure's upper/mid/lower third by pure geometry; the viewer draws a "~" 3D label there and quotes the finding. Never pixel interpretation: an approximate region, not the lesion's position

Every case ships the same data-driven viewer (skills/imaging-3d/assets/viewer.html + per-case scene.json; NiiVue is vendored, so it works offline). The grouped structure panel gives per-organ show/hide, solo (◐ — isolate one structure) and slice-focus (⌖ — jump the MPR crosshair to that organ's centroid and overlay its segmentation mask on the slices). Structure names, groups and colors are mapped to Turkish deterministically in scripts/build_scene_viewer.py.

Runtime on an Apple-Silicon Mac: --fast segmentation ≈ 1–3 min per study; full-res 10–45 min. Cost per study: $0.

The Reviewer Gate

Four independent auditors, spawned in parallel, each with one checklist (agents/reviewer.md):

  • A — Citation integrity: re-resolves every citation; spot-checks load-bearing figures against retrieved abstracts; verifies no [UNSOURCED]/[ESTIMATE] marker was laundered into fact.
  • B — Dissent integrity: diffs the minutes against the transcripts; every disagreement must appear in the ledger verbatim; no false consensus, no smoothed paraphrase posing as quotation.
  • C — Safety language: clinician-directed framing, untouched disclaimer block, no pixel interpretation, time-sensitive items prominent, no singular verdict.
  • D — Completeness & traceability: every template section filled; every recommendation traceable to a named member and transcript; every abnormal finding in the case file has a disposition.

Any FAIL is fixed from the transcripts only and the failed checklist re-runs with a fresh auditor — fresh, because in development the dissent auditor once failed the minutes twice in a row, the second time over the orchestrator's own fix (it had spliced three quote fragments and labeled the composite "verbatim"). The gate audits the fixer too.

Installation

Requires Claude Code.

# One-liner (straight from GitHub):
claude plugin marketplace add arbade/konsil
claude plugin install konsil@konsil-marketplace -y

# — or from a local clone:
git clone https://github.com/arbade/konsil ~/konsil
claude plugin marketplace add ~/konsil --scope local
claude plugin install konsil@konsil-marketplace -y

# 3. (Optional — only for /konsil:imaging) install the deterministic imaging stack
bash ~/konsil/scripts/setup_imaging.sh
#    installs dcm2niix (brew) + a Python venv with TotalSegmentator, nibabel,
#    scikit-image, trimesh. Model weights (~GBs) download on first segmentation run.

New Claude Code sessions pick the plugin up automatically; commands appear under the /konsil: namespace.

Commands & usage

Command What happens
/konsil:case <folder or case-id> Intake: builds the structured case file from mixed documents, runs plausibility checks and inconsistency flags, then stops at the Missing Data Gate with targeted questions. No board convenes from this command.
/konsil:board <case-id> [--solo] Convenes the board on a completed case file: blind Round 1 → Challenger cross-examination → Steward → Reviewer-Gated minutes. --solo runs a single-agent baseline with identical evidence rules, for A/B comparison.
/konsil:update <case-id> <new docs> Longitudinal mode: folds new results into the case, computes interval changes (growth rates, trend slopes), re-convenes in delta mode. The new minutes lead with what changed.
/konsil:imaging <case-id> [--fast] [--mr] The deterministic DICOM → 3D pipeline + browser viewer. --mr selects the MRI segmentation task.
/konsil:okuma <case-id | image> OFF-PROTOCOL consent-gated AI image read. Deterministic render → systematic observational read with confidence levels. Never a radiology report; the no-pixels default is untouched.

A typical first session:

> /konsil:case ~/Downloads/moms-documents
  → case file built: 23 dated events, 41 lab rows (2 flagged implausible, confirm?),
    3 inconsistency flags, intake_status: pending
  → 8 targeted questions (answer, or say "proceed with gaps")

> /konsil:board case-001
  → roster: nephrology, endocrinology, internal medicine + Challenger/Checklist/Steward
  → [several minutes of parallel agents doing live literature retrieval]
  → minutes.md: consensus (with confidences), 2 dissent-ledger entries,
    can't-miss audit, staged plan, 11 questions for the treating team

Expect a full board run to spawn 10–15 agent sessions (specialists ×2 rounds, officers, drafter, 4+ gate auditors). It is deliberate overkill — that's the product.

Case workspace anatomy

cases/<case-id>/
├── inbox/                      # your documents, untouched
├── case.md                     # structured case file (template below)
├── scene/                      # deterministic imaging output (if any)
│   ├── nifti/  segmentations/  meshes/
│   ├── structures.json         # names + volumes → board-readable text
│   └── viewer.html             # rotatable browser scene
├── ai-reads/                   # OFF-PROTOCOL AI reads (consent memlog + read-*-OFF-PROTOCOL.md) — never a report
└── board-<date>/
    ├── memlog.md               # append-only orchestration log
    ├── round1-<member>.md      # verbatim blind assessments
    ├── round2-challenger.md    # the attack
    ├── round2-<member>.md      # concede / rebut / revise
    ├── round2-steward.md       # staged plan
    ├── round3-*.md             # re-consultations (if the gate demands them)
    ├── gate-A..D*.md           # Reviewer Gate audit reports
    └── minutes.md              # the deliverable

Everything is Markdown. The transcripts are the audit trail; the minutes may organize them but never replace them.

Templates

Two templates drive all structured output (both in templates/):

case-file.md — frontmatter carries machine-readable state:

---
case_id: ...
language: ...          # board writes minutes in the case's language
intake_status: pending # pending | complete | complete-with-gaps
gaps_accepted: [...]   # every gap the user chose to proceed with
---

plus: ranked problem list · timeline table (one dated fact per row; corrections are new rows flagged [CORRECTS <date>], never overwrites) · labs with printed reference ranges · imaging findings from written reports only · inconsistency flags · source map (every fact → its document).

board-minutes.md — 12 sections + appendix; the ones that make Konsil different:

  • §7 Dissent Ledger — every unresolved disagreement, both positions verbatim, each side's "what would change my mind." An empty ledger on a complex case is itself flagged as a red flag.
  • §8 Can't-Miss Audit — coverage grades, orphaned findings, time-sensitive items with safe intervals.
  • §9 Stewardship — Stage-0 (zero-cost, same-day) → Stage-1 (the single uncontaminated draw/test set) → conditional branches, each expensive step behind an explicit gate; rejected actions listed with reasons.
  • §10 Questions for the treating physician / radiologist — including report inconsistencies to re-measure.
  • §12 Sources — every citation attributed to the member whose session verified it; [UNSOURCED]/[ESTIMATE] items listed separately.
  • The disclaimer block is untouchable — Gate C diffs it byte-for-byte against the template.

Extending Konsil

Add a specialist type. Specialists are generic: the orchestrator passes the specialty in the spawn prompt, so "add pediatric neurology" is a roster decision, not a code change. To give a specialty standing knowledge (e.g., a guideline file), add a persona file under agents/ following specialist.md's structure and reference it from the orchestrator skill.

Add an officer. Copy an officer charter (challenger.md is the best template), define: the failure mode it exists to catch, its procedure, its output format, and its evidence obligations. Wire it into skills/board-orchestrator/SKILL.md Phase 2/3.

Change the report. Edit templates/board-minutes.md — then update Gate D's completeness expectations in agents/reviewer.md to match. The gate must always chase the template.

Add a gate. New auditor checklist in agents/reviewer.md (E, F…), then list it in skills/board-report/SKILL.md. Gates are cheap; silent failure modes are not.

Tune evidence sources. skills/literature-protocol/SKILL.md is the single place agents learn where (and where not) to search. Keep the verification step non-negotiable.

Literature & data APIs used

API Use Auth / limits Notes
Europe PMC REST Primary search + open-access full text none / 10 req·s⁻¹ The workhorse
PubMed E-utilities Secondary search, record verification none / 3 req·s⁻¹ (10 with free key)
NLM Clinical Tables LOINC mapping for lab intake none Reference ranges still come from the printed report — they are lab-specific
Guideline bodies (EAU, AUA, NICE, USPSTF…) Guideline anchoring via WebFetch varies
Deliberately not used Semantic Scholar (restrictive API license), WikEM (ToS prohibits AI use), UpToDate / OpenEvidence (closed) Documented in the protocol so agents don't wander into them

Safety & regulatory posture

Konsil is built for personal / educational use and positions itself, by construction, on the non-device side of the clinical-decision-support line:

  1. Inputs are text: history, discrete lab values, and the radiologist's written findings. Software that analyzes medical image pixels and renders findings is a regulated medical device in the US (FDA CADx classifications), EU (MDR Rule 11, Class IIa+), and jurisdictions that mirror the MDR. Konsil's imaging pipeline is deliberately deterministic and non-interpretive.
  2. Outputs are recommendations with reviewable basis — ranked possibilities with confidences, verified sources, and preserved dissent — addressed to a clinician, never directives to a patient. (US readers: this maps onto the FDA's non-device CDS criteria, including the January 2026 guidance revision that explicitly covers synthesis of a radiologist's written findings. This README is not legal advice.)
  3. Not for time-critical use. The minutes template says so; Gate C enforces the framing.
  4. If you fork this into something patient-facing or pixel-reading, you are building a medical device. Plan accordingly, in every jurisdiction you ship to.
  5. Model-provider terms apply independently (e.g., usage policies treating medical diagnosis as high-risk requiring qualified professional review). Konsil's clinician-review framing is designed to comply; keep it.

Known limitations (measured, not hypothetical)

  • AI image reading is unreliable and is therefore excluded from board reasoning. Published evaluations and our own development scoring agree: acceptable on negatives, weak at characterization, blind to subtle findings. If you enable any experimental image viewing for personal use, every observation must carry the unverified-observation label and converts to questions, never findings.
  • No formal benchmark yet. Blinded end-to-end development runs are encouraging but few, and retrospective published material is always subject to training-data-contamination concerns (file-blinding ≠ memory-blinding). The honest benchmark — board vs. --solo single agent on ≥10 blinded retrospective cases with independent evaluators — is designed but not yet run.
  • Cost & latency. A full board is 10–15 agent sessions with live retrieval. Minutes take tens of minutes, not seconds. --solo exists for cheap runs — and for keeping the board honest.
  • Language coverage is whatever the underlying model handles; the system passes the case language through explicitly (tested: English, Turkish).
  • Segmentation scope. The default (fully open) TotalSegmentator tasks cover trunk organs and major bones — not menisci, ligaments, or small joints. Specialized subtasks exist under a non-commercial license (audit before any commercial redistribution; see below).

Roadmap

Grounded in the measured gaps, roughly in order:

  1. Kill-test #1: board vs. --solo on ≥10 blinded retrospective published cases, independent blinded evaluators. The result decides whether the multi-agent board earns its cost — the repo ships the instrumentation either way.
  2. Protocol-adequacy gate: "this study cannot answer this question" detection from DICOM metadata (missing sequences/contrast/FOV) at intake — extends the two categories that measured 100%.
  3. Quantification tools: user-guided measurement on slices (system measures, never characterizes); automatic per-structure volumetrics into the case file.
  4. Longitudinal difference detector: rigid registration between serial studies + change reporting — change detection is an easier, more honest problem than characterization.
  5. Targeted closed-question reading experiment (pre-registered): pipeline-cropped ROIs + narrow yes/no questions vs. the measured free-reading baseline; enters the product only if it clears pre-set thresholds.
  6. Human-in-the-loop second-read integration: package slices + the board's questions for a real radiologist's paid second read — the only currently honest answer for subtle findings.
  7. Accuracy scorecard automation: when a written report arrives after AI observations were made, auto-score them so the system accumulates its own honesty data.

Repository layout

konsil/
├── .claude-plugin/          # plugin + marketplace manifests
├── commands/                # /konsil:case | board | update | imaging | okuma
├── skills/
│   ├── case-intake/         # intake + Missing Data Gate
│   ├── board-orchestrator/  # roster, blind rounds, cross-exam, phases
│   ├── board-report/        # minutes drafting rules + Reviewer Gate loop
│   ├── literature-protocol/ # evidence rules + API cookbook
│   ├── case-update/         # longitudinal delta mode
│   ├── imaging-3d/          # deterministic DICOM → 3D scene
│   └── imaging-ai-read/     # OFF-PROTOCOL consent-gated AI read (separate from imaging-3d)
├── agents/                  # specialist, challenger, checklist, steward, reviewer
├── templates/               # case-file.md, board-minutes.md
├── scripts/                 # setup_imaging.sh, dicom_to_scene.py, dicom_to_png.py
├── docs/assets/             # README media (from the author's own anonymized MRI)
└── cases/                   # your cases live here (gitignored — always empty upstream)

cases/ is gitignored on purpose: never commit health data — yours or anyone else's. Board runs, transcripts, gate reports and minutes all live under cases/<id>/ and stay local.

License & attributions

MIT — see LICENSE (with a non-binding medical-disclaimer addendum).

Standing on the shoulders of:

  • TotalSegmentator (code & default task weights: Apache-2.0; several specialized subtasks are non-commercial — audit per-task before commercial redistribution)
  • dcm2niix (BSD-style), NiiVue (BSD-2), nibabel / scikit-image / trimesh (BSD/MIT)
  • Europe PMC and NCBI E-utilities for open literature access; NLM Clinical Tables for LOINC
  • Orchestration patterns informed by BMAD-METHOD (multi-persona subagent party-mode, reviewer-gate document pipelines) and plugin packaging by pm-skills
  • Built with Claude Code; board members are Claude subagents

The hero and gallery GIFs derive from the author's own knee MRI (exported from a national teleradiology portal with identifying metadata already stripped), published here with the owner's explicit consent. Do not reuse them for model training or clinical purposes.


Konsil renders opinions with receipts, disagreements included — and the final word always belongs to your physician.

About

Adversarial AI medical case-review board for Claude Code — blind specialist rounds, verified citations, dissent-preserving minutes, deterministic DICOM-to-3D. Not a medical device.

Topics

Resources

Stars

3 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages