0. Document information
This model card covers Dossier's Autofill and Summarize features. It does not describe BEST, Toolbox, or every AI capability available through Palmera. It is an information reference for users, not a contract, certification, clinical recommendation, or legal classification.
- This full web edition follows the August 2026 background card's section structure and updates its product behavior, live managed-model settings, published provider terms, and internal benchmark results.
- The CDA-AMC Position Statement informs the user's reporting context; this model card is not CDA-AMC endorsement or approval. Users remain accountable for their own submissions and required AI disclosures.
- Date
- 8 October 2026
- Scope
- Dossier Autofill and Summarize
- Intended readers
- Health economists, medical writers, market-access and regulatory teams, and HTA consultants
- Document status
- Published web reference; product facts reflect current Dossier behavior and managed-model configuration as of the date above.
1. Feature overview
Autofill proposes draft values for fields in a Dossier deliverable. For each run, the user selects one project source and one PICO, or explicitly chooses No PICO. Suggestions appear with their source evidence when available; the user can review, edit, apply, or reject them. Starting a run does not change the document.
Summarize rewrites the existing content of one prose-capable field, using its accepted content and associated evidence. It does not summarize the full source library or combine multiple source documents into a new evidence review.
- Intended uses: help draft structured clinical and health-economic submission fields, reduce manual extraction and transcription, and make evidence easier to review.
- Examples include first drafts for jurisdiction-specific submission forms such as CDA-AMC, INESSS, AMCP, and Ontario Health templates supported by the organization.
- Summarize helps a reviewer inspect or rewrite existing field content and identify evidence gaps; it does not create a new evidence review from several source documents.
- Autofill is not an independent clinical or health-economic analysis and does not make reimbursement decisions, treatment recommendations, or regulatory determinations.
- Autofill does not generate a submission end to end, perform meta-analysis or network meta-analysis, estimate treatment effects, or arbitrate conflicting evidence.
- Summarize is not a formal systematic review, meta-analysis, evidence grading, study-quality assessment, or risk-of-bias evaluation.
- Neither feature replaces expert review, editing, approval, or the user's responsibility for the submitted content.
2. Intended users and use context
Dossier is used by health economists, medical writers, market-access leads, regulatory-affairs professionals at pharmaceutical and biotechnology organizations, and external HTA consultants working for sponsors. Teams prepare reimbursement submissions collaboratively, with different people contributing under organization membership and project permissions.
Sources may include proprietary clinical study reports, pharmacoeconomic models, sponsor dossiers, published literature, guidelines and product monographs. Palmera does not independently verify whether an uploaded source is accurate, complete, or appropriate for a particular field.
- Work context
- Collaborative project work with organization membership and project-level access controls.
- Human oversight
- A qualified person reviews each proposal against its source and decides whether to use it.
- User responsibility
- The user chooses suitable sources, checks the output, and ensures the submission meets the relevant HTA body's requirements.
- Palmera responsibility
- Maintain the product, evidence controls, data security, and model configuration for the managed route; this does not transfer submission accountability from the sponsor.
3. AI model description
The Palmera-managed Dossier route currently calls Google Vertex AI through a us-central1 endpoint. Live settings identify Gemini 2.5 Flash for Autofill and Gemini 2.5 Pro for Summarize. These are the configured model identifiers as of this card's date; model availability and settings can change.
Palmera uses provider-hosted foundation models and does not train or fine-tune them. The model provider determines the composition of its foundation-model training data.
Organizations can configure bring-your-own-key (BYOK) providers and models. In BYOK mode, the organization-selected provider receives the request under its own service terms. The Google terms in Section 7 apply only to the Palmera-managed Google route.
- Autofill uses AI for field classification and field-level extraction from retrieved context. Its prompt combines the user's PICO/scope with target-field name, authored guidance, native type, and selected source passages.
- The model does not read the full source at Autofill time: supported source documents are prepared/indexed separately and retrieval supplies relevant passages. Source preparation and indexing are separate AI/data-processing steps from the Autofill generation call.
- Evidence checks, citation materialization, application, and the review interface also use conventional software. A failure to locate a supporting passage can result in a withheld proposal or warning.
- Summarize uses an LLM to propose alternate wording for one existing field. Verification and repair steps include deterministic checks for references, figures, protected terms, and source content; these are not proof of correctness.
- Autofill · Palmera-managed
- Google Vertex AI · Gemini 2.5 Flash · us-central1
- Summarize · Palmera-managed
- Google Vertex AI · Gemini 2.5 Pro · us-central1
- BYOK
- Provider and model can vary by organization configuration.
- Fine-tuning by Palmera
- None in the current Dossier implementation.
4. Data
Dossier prepares supported uploaded source files, including PDF and DOCX, for retrieval. Scanned or poorly structured files, very long documents, non-standard terminology, and conflicting evidence can reduce retrieval or extraction quality.
For the Palmera-managed Google route, provider terms and deletion details are described in Section 7. Palmera's application storage, retention, and deletion practices are governed separately by Palmera's own terms and controls. In BYOK mode, data handling depends on the configured provider and agreement.
- User-uploaded documents and generated answers are not used by Palmera to train or fine-tune the provider's model.
- The feature does not browse the public web or other organizations' project data as part of an Autofill run.
- The provider determines the underlying foundation model's training data. Its complete training-corpus composition is not determined by Palmera.
| Feature input | What is used |
|---|---|
| Autofill source | One source document selected for the run. Retrieval supplies relevant passages; the model call is not unrestricted web search. |
| Autofill scope | A selected project PICO (population, intervention, comparator, outcomes and related context), or explicit No PICO, plus target field name, authored guidance and native type. |
| Template context | The selected document's authored field and table structure define the expected answer type and target. |
| Autofill output | A field-appropriate draft and source references when available; unsupported or unsafe proposals can be withheld or flagged. |
| Summarize input | Existing accepted content in one prose-capable field plus its associated source references and context. |
| Summarize output | Proposed alternate wording for that field, with review warnings and evidence context. |
| Source validation | Palmera does not validate the factual accuracy or completeness of user-uploaded documents. |
5. Performance and evaluation
The Palmera issue report contains thousands of recorded runs and workflow checks across different models, fields, sources, and purposes. Those records are not one uniformly scored dataset, so run volume is not an accuracy rate.
The retained Autofill gold benchmark uses a 168-chunk source dossier and a 35-field Ontario Health Innovation Pathway intake form. Each field has an expected answer state, rationale, and source anchor. The scorer checks whether the output state is appropriate and whether the answer or citation points to gold evidence. A right result or an acceptable blank counts as successful; unsourced answers are reported separately; pipeline errors are excluded from the scorable denominator.
The background card's proposed evaluation dimensions were field-level factual support, source-attribution support, substantive completion, unsupported-content rate, and representative coverage across jurisdictions. The retained numeric score below implements the gold-set state/evidence rule; it is not a published result for every proposed dimension.
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.
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.
- The two model runs covered different numbers of fields; these percentages are not a controlled head-to-head comparison.
- This is one internal gold-set benchmark, not general model accuracy, clinical validation, or independent external validation.
- A separate 29 September 2026 workflow check on a 4,120-page clinical study report found that Autofill missed a study phase stated near the beginning of the report. In the same scenario, Summarize omitted a dosing frequency and showed only a near-verbatim warning; the reviewer rejected the draft. This is a documented failure case, not an overall error rate.
- Many additional issue-report runs are scenario tests with different scopes; they should not be combined into a single score unless they share a defined gold set and scoring method.
| Internal run · 27 September 2026 | Coverage | Gold score | Other outcomes |
|---|---|---|---|
| Gemini 2.5 Flash · current Autofill model | 25 of 35 fields | 18/24 scorable · 75% | 17 right · 4 wrong · 2 missed · 1 acceptable blank · 1 pipeline error excluded |
| GPT-6 Sol · reference run, not current managed model | 35 of 35 fields | 30/35 · 85.7% | 28 right · 0 wrong · 3 missed · 2 acceptable blanks · 2 unsourced |
6. Risks, limitations, and mitigations
| Risk | Description | Mitigation and limit |
|---|---|---|
| Unsupported or incomplete output | A proposal may be unsupported, incomplete, or omit an important detail. | Source-grounding checks, citations and review warnings help users inspect results; they do not prove all relevant evidence was found or correctly interpreted. |
| Source misattribution | An answer can point to the wrong source document or passage. | Dossier records source references and checks evidence associations; users should verify important claims against the original. |
| Incomplete extraction | Autofill may miss relevant material, especially in a long or complex source. | Users review every proposed answer and check empty or flagged fields; an absent suggestion is not proof that the source contains no evidence. |
| Conflicting evidence | Source documents may disagree. | Autofill uses one selected source per run and does not arbitrate across documents; users choose and reconcile sources. |
| Overreliance | Users may accept a draft without adequate expert review. | AI proposals are not applied automatically. A user must review and explicitly apply them; warnings are not correctness verdicts. |
| Sensitive source content | Uploaded files may contain confidential sponsor or clinical information. | Dossier applies organization/project access controls. Provider processing is subject to the selected route's terms; BYOK terms vary by provider. |
| Jurisdiction-specific framing | A draft may not follow the target authority's wording conventions. | The selected PICO, template and guidance constrain context; jurisdiction expertise and human review remain necessary. |
| Model changes | Provider changes may affect output characteristics. | The managed model settings and card date are shown here; material changes should prompt an update to this reference. |
7. Data privacy and provider terms
Dossier protects confidential project material with organization and project access controls. Palmera does not use uploaded documents or generated answers to train or fine-tune AI models. For Palmera-managed AI requests, Google's terms also prohibit model training on customer data without prior permission or instruction.
Palmera-managed AI requests use Google Vertex AI through the us-central1 endpoint. Palmera has disabled Google's AI request caching for this project, and optional request-response logging is not enabled for the current managed Gemini models.
The managed service uses Google's standard abuse-prevention safeguards. If safety systems flag suspected misuse, Google may retain the flagged prompts for up to 90 days for investigation by authorized staff. These security records are not used for model training.
Your organization can choose a supported provider and model using its own API key (BYOK). That provider's terms and account settings govern its model requests; Palmera's application access controls and storage practices continue to apply. Source preparation and indexing are separate processing steps and can use Palmera-managed services.
Dossier retains source files, saved work and processing records as part of the application, under Palmera's own controls and terms. Teams should use source material they are authorized to process and select an AI provider consistent with their confidentiality obligations. The table below distinguishes these application records from provider processing.
- Provider certifications apply only to Google services listed within Google's applicable audited-service scope; they are not Palmera certifications.
- The exact provider, model, and data-processing terms can differ when an organization selects BYOK.
| Layer | Protection and scope |
|---|---|
| Project access | Organization and project permissions control access to Dossier content. |
| Provider model training | Google's terms prohibit model training on customer data without prior permission or instruction. |
| Managed AI caching and logging | Google AI caching is disabled for Palmera's managed project. Optional request-response logging is not enabled for the current managed Gemini models. |
| Security monitoring | Google may retain prompts flagged for suspected abuse for up to 90 days for restricted investigation; these logs are not used for training. |
| Provider deletion | Google's DPA requires deletion instructions to be completed as soon as reasonably practicable, within 180 days, subject to law. This is a maximum deletion-completion window, not the routine retention period for prompts. |
| Application storage | Dossier's source files, saved work and processing records are managed under Palmera's application controls and terms. Disabling provider caching does not delete application records. |
| BYOK | The selected provider's terms and account settings apply to its model requests. Palmera's application controls continue to apply. |
8. Human oversight and accountability
- Autofill proposals are field-level drafts. Users inspect proposed content and evidence, then choose to edit, apply, or reject it.
- Summarize is a draft of existing field content. Users can keep the original, edit the proposal, or accept it.
- Warnings are not final correctness judgments. No warning does not prove the answer is correct.
- Dossier permits export of unfinished drafts; export does not require every field to be complete or reviewed.
- Users are responsible for verifying claims against cited source material, selecting appropriate inputs, and meeting the relevant authority's submission requirements.
- Palmera is responsible for maintaining the software, model configuration for the managed route, and platform data controls. This does not transfer the sponsor's responsibility for the submitted content.
9. Transparency and explainability
Autofill proposals can retain evidence associations identifying the source, passage, location, excerpt and accepted answer fragment. Visible citation fields and internal evidence provenance are related but are not the same object.
Summarize retains original field material and supporting references for review. The Verify checks look for changes such as missing figures, terms, references, or source content; they do not prove semantic correctness.
Dossier does not provide token-level or attention-based model explanations. Source passages, generated-output metadata and review warnings make the workflow inspectable, but do not reveal the model's complete reasoning.
Confidence and quality indicators are review cues, not calibrated probabilities. A high indicator is not a guarantee; a low indicator does not by itself prove an answer is wrong.
10. Regulatory and ethical context
Dossier's Autofill and Summarize are documentation and evidence-review tools. They do not make clinical treatment recommendations or reimbursement decisions. This model card does not determine medical-device status or an EU AI Act risk classification; organizations should assess their own obligations with qualified counsel.
Users should follow the relevant HTA body's AI reporting requirements, disclose AI use when required, and verify submission content against source evidence. The CDA-AMC Position Statement is relevant to users preparing submissions but does not endorse or approve Dossier.
11. Keeping this information current
Update this card when Palmera-managed model identifiers, provider routes, prompts or retrieval behavior, provider data terms, evaluation evidence, or user-visible review behavior materially changes. The as-of date identifies when this information was reviewed. BYOK settings can differ by organization.
- As of
- 8 October 2026
- Palmera-managed models
- Gemini 2.5 Flash for Autofill; Gemini 2.5 Pro for Summarize
12. References
- Mitchell et al. (2019), Model Cards for Model Reporting, ACM FAccT.
- AMCP (2026), AI Transparency in Health Economics webinar materials.
- FDA (January 2025), Draft Guidance: Artificial Intelligence-Enabled Device Software Functions, Appendix E.
- EU AI Act, Regulation (EU) 2024/1689.
- ISO/IEC 42001:2023, Artificial Intelligence Management Systems.
- Palmera Dossier Autofill gold harness and issue-report evaluation records, September 2026.
Annex A. Scope of the evidence in this card
- The managed model settings are read from the deployed Dossier API/worker configuration as of 8 October 2026; they may change.
- The 75% Gemini gold-set result is limited to 24 scorable fields in one internal gold test. It is not a general accuracy score, and the comparison run used a different field subset.
- Thousands of issue-report runs document product scenarios; unless a run uses an agreed gold answer and scoring rule, it is not a benchmark score.
- Google provider terms cited here apply to the managed Vertex route. BYOK traffic follows the organization-selected provider's terms.
- This reference does not claim zero retention, Palmera security certification, independent clinical validation, or legal risk classification.