Health economists, medical writers, market-access and regulatory teams, and HTA consultants.
Intended use — Autofill
Propose draft answers for structured submission fields from one selected source document and one project PICO, or an explicit No PICO scope. Users review, edit, apply or reject each proposal.
Intended use — Summarize
Rewrite the accepted content of one prose-capable field using its associated evidence and context. It does not synthesize the full source library or perform a new review across documents.
Out of scope
Clinical or health-economic analysis; treatment or reimbursement decisions; systematic review, meta-analysis and evidence grading; use without expert review.
Required oversight
Expert review is required before submission. AI generation does not apply answers automatically. Dossier permits unfinished draft exports; there is no requirement to lock every section before export.
2. AI Model Facts
Field
Palmera Dossier
Model type
Provider-hosted large language models. Autofill retrieves relevant source passages before extraction; conventional software validates, presents and applies results.
Model provider
Managed route: Google Vertex AI. Organizations may instead configure a supported provider and model with bring-your-own-key (BYOK).
Model name & version
Managed Autofill: Gemini 2.5 Flash. Managed Summarize: Gemini 2.5 Pro. Source/query embeddings: text-embedding-005. BYOK model identifiers depend on the organization’s configuration.
Fine-tuned?
No Palmera training or fine-tuning. The managed route uses provider foundation models.
Hosting / region
Managed model requests use Google Vertex AI through the us-central1 endpoint. BYOK hosting depends on the chosen provider.
Input data
Autofill: selected-source passages, PICO or No PICO scope, field name, guidance and native type. Summarize: the field’s accepted content, associated evidence and context. Source preparation and indexing are separate processing steps.
Training data
The provider determines the foundation model’s training data. Palmera does not use uploaded documents or generated answers to train models. Google’s terms prohibit training on customer data without prior permission or instruction; BYOK terms vary.
3. Performance & Evaluation
Field
Palmera Dossier
Autofill accuracy
Recorded gold baseline, 27 September 2026: Gemini 2.5 Flash 18/24 scorable fields (75%); GPT-6 Sol comparison 30/35 (85.7%). The Gemini run covered 25 fields and excluded one pipeline error. These are answer-state/evidence scores with different coverage, not a controlled model comparison.
Source attribution accuracy
The gold scorer checks expected source anchors or cited chunks. Gemini recorded 0 unsourced verdicts; the GPT comparison recorded 2 correct checkbox selections without citations. An unsourced verdict is reported separately from an incorrect answer or missed evidence.
Hallucination rate
The Gemini gold baseline recorded 4 incorrect selections among 24 scorable fields (16.7%), all false-positive technology-type checkboxes; the GPT comparison recorded 0 wrong selections among 35 fields. These observed field errors are reported precisely, rather than relabelled as a universal free-text hallucination rate.
Summarize faithfulness
Retained scenario checks assess source content, figures, terms and references. In the 29 September TAVI journey, a reviewed two-paragraph summary retained one native citation and two provenance records through save, reopen and DOCX/PDF export. A separate CSR review exposed an omitted dosing frequency, demonstrating a completeness limit.
Evaluation method
A 35-field Ontario Health gold set with expected answers, rationales and source anchors was scored against a 168-chunk dossier. Gemini: 17 right, 4 wrong, 2 missed, 1 acceptable blank, 1 pipeline error. Thousands of workflow checks and retained model responses supplement the gold results, including review, correction and export evidence.
Known limitations
Poor scans or document structure; long or ambiguous sources; missed evidence and conflicting source material. Language and jurisdiction coverage do not have separate validated accuracy scores.
4. Data & Privacy
Field
Palmera Dossier
Source docs used for training?
Palmera does not train models on uploaded documents or answers. Google’s managed-service terms prohibit training on customer data without prior permission or instruction. An organization’s BYOK provider has its own terms.
Data retention policy
Dossier stores source files, saved work and processing records under Palmera’s application controls. Managed Google AI caching is disabled; optional request-response logging is not enabled for the current Gemini models. Standard security monitoring may retain flagged prompts for up to 90 days. Google’s DPA allows up to 180 days to complete a deletion instruction, subject to law; this is not routine prompt retention.
Third-party data sharing
Yes: the model provider receives the content required for the selected AI operation. Managed requests follow Google Cloud’s service terms and Data Processing Addendum. BYOK requests follow the selected provider’s terms. Source preparation and indexing can still use Palmera-managed services.
Data processing region
Managed model inference: us-central1. This identifies the model endpoint, not every application store or subprocessor. BYOK regions depend on the selected service.
Access controls
Organization membership, operation permissions and project access govern content access. Private-project permissions and tenant-scoped source processing protect confidential project material.
Compliance
This card does not assert Palmera SOC 2 or ISO 27001 certification. Google’s certifications cover its listed services, not Palmera itself. Applicable privacy obligations, including GDPR or PIPEDA where relevant, depend on the organization’s processing.
5. Key Risks & Mitigations
Field
Palmera Dossier
Hallucination
Source-grounding checks, evidence citations and review warnings help identify unsupported proposals. They do not guarantee correctness; users verify and apply draft content deliberately.
Source misattribution
Source references and passages support field-level verification. Reviewers check that the cited evidence answers the field and refers to the intended study or population.
Overreliance
AI content is presented for review, editing or rejection. Expert review remains required before submission, including when a draft has already been exported.
Confidential data
Organization/project access controls, tenant-scoped source processing, managed-provider data terms and disabled Google AI caching protect confidential material. BYOK lets organizations select their supported model provider.
Jurisdiction framing errors
PICO, template structure and authored guidance constrain the task. Reviewers check authority-specific requirements and the appropriateness of the selected source.
Model version changes
Managed model identifiers and the information date are recorded here. Revisit affected workflows and update this card when models or feature behavior change. BYOK choices can differ by organization.
6. Oversight & Transparency
Field
Palmera Dossier
Workflow enforcement
Generating suggestions does not change the submission. Applying a suggestion requires a user action. Quality issues may warn or block application, but export does not wait for every section to be reviewed or locked.
Source attribution shown?
Yes. Autofill offers show source evidence and references when available. Applied answers can retain evidence provenance and native citations; reviewers verify the final wording and citation together.
Confidence indicator?
Quality flags and confidence signals guide review. They are not calibrated probabilities or guarantees of correctness; an absent suggestion can reflect missing evidence or a processing failure.
Summarize traceability?
Summarize uses the field’s accepted content and associated evidence. Source context, reference checks and review warnings support comparison with the original; they do not prove that every relevant detail is preserved.
AI disclosure to users
Autofill and Summarize are named AI drafting actions with source/context choices and review interfaces. This public model card describes the model route, limitations, evaluation and privacy practices.
User responsibility
Check proposed content against the source, resolve omissions and conflicts, and approve the final submission. Declare the AI features actually used and the human review performed.
7. Regulatory & Ethical Alignment
Field
Palmera Dossier
Medical device regulation
Dossier’s intended function is submission drafting and documentation. It does not diagnose, recommend treatment or make clinical decisions. This card is not a regulatory classification.
EU AI Act
This card documents intended use, model providers, oversight and limitations. It does not assign a legal risk category or assert AI Act compliance.
ISO/IEC 42001:2023
A reference framework for AI governance documentation. This card records the feature’s scope and controls; it does not claim certification to the standard.
AMCP AI framework
The card covers model identity, intended use, evaluation, limitations, privacy and oversight for health-economic users. It does not claim AMCP approval or endorsement.
FDA draft guidance (Jan 2025)
The January 2025 draft is a documentation reference for model-card structure. This card does not imply FDA approval or determine medical-device status.
CDA-AMC disclosure
Use this card to explain the actual AI-assisted extraction and rewriting performed, the sources and scope selected, and human review. The submitter remains responsible for AI declarations and programme requirements; CDA-AMC has not endorsed Dossier.
Next review date
Review when managed models, feature behavior, provider terms or evaluation evidence materially change. This card does not set a separate annual review deadline.
Warnings & Out-of-Scope Uses
Review and approve AI-assisted content before using it in a submission.
Use formal evidence-review and analytical methods for systematic review, meta-analysis, evidence grading and treatment-effect estimation.
Check source quality before relying on extraction from scanned, poorly structured or unverified documents.
Verify important claims, citations, numbers and omissions against the original source. Exporting a draft is not expert sign-off.
Supporting references
This card supports users who need to describe AI-assisted work in an HTA submission. The detailed background explains the workflow, benchmark results and scope of the available evidence.