> ## Documentation Index
> Fetch the complete documentation index at: https://wuweism.com/llms.txt
> Use this file to discover all available pages before exploring further.
# PDF Synthesis
> Extract numeric evidence and causal claims from a single PDF with structured, uncertainty-labeled output.
PDF Synthesis is a focused extraction surface: give it one document, and it returns a structured breakdown of the quantitative evidence inside it and the causal claims that evidence can support. Where [Hybrid Synthesis](/workbench/hybrid-synthesis) is designed for cross-document conflict and novelty, PDF Synthesis is designed for depth — extracting everything causally relevant from a single source and presenting it in a form that feeds directly into the Claim Ledger.
Navigate to PDF Synthesis by selecting **PDF Synthesis** in the sidebar, or by opening `/pdf-synthesis` directly.
## Uploading a document
The PDF Synthesis surface opens with a dropzone as its primary interface. You will see a large upload area in the center of the screen.
Drag your PDF file onto the dropzone, or click **Choose file** to browse your filesystem. Only `.pdf` files are accepted.
PDF Synthesis runs extraction automatically after upload. A progress indicator shows parsing and analysis stages as they complete. Extraction time depends on document length and numeric density.
When extraction completes, the output panel populates with three structured sections (described below).
PDF Synthesis processes one document per run. If you need to synthesize across multiple documents simultaneously, use [Hybrid Synthesis](/workbench/hybrid-synthesis) instead.
## The three-section output
Every completed PDF Synthesis run returns the same three-section structure. This contract is fixed — you can rely on it for downstream processing or integration into reporting workflows.
### Section 1 — All explicit numbers
Section 1 is an exhaustive inventory of every number that appears in the document in a causal or scientific context. This includes:
* Measured quantities with units
* Effect sizes and odds ratios
* Sample sizes and study parameters
* Percentages, rates, and proportions
* Confidence intervals and standard errors
* p-values and test statistics
* Dates and durations relevant to causal timing
Each entry records the number, its unit (if applicable), the sentence it appeared in, and the page or section location in the source document.
Section 1 is intentionally exhaustive and unfiltered. Its purpose is to ensure nothing is missed before the filtering step in Section 2.
### Section 2 — Claim-eligible numerics
Section 2 is a filtered subset of Section 1: only those numbers that could support a causal claim. A number is claim-eligible if it:
* Represents an effect, association, or outcome (not merely a reference or label)
* Has sufficient context to identify what was measured and under what conditions
* Could, in combination with a causal model, support a directional assertion about a relationship between variables
Each entry in Section 2 carries an **evidence class label**:
| Label | What it means for this number |
| ------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
| `bibliographic/structural only` | The number appears in a citation, reference list, or structural element. It cannot be used as primary evidence for a claim. |
| `mixed` | The number is present in the body but lacks full measurement context (no reported conditions, missing units, unclear operationalization). Usable with caution at Rung 1. |
| `metric-bearing` | The number is a well-formed measurement with sufficient context to support Rung 2 claims and, with appropriate SCM framing, Rung 3 inference. |
Use Section 2 to quickly assess whether a document's evidence base can support the causal claims you want to make. A document that returns mostly `bibliographic/structural only` labels in Section 2 is a thin evidentiary source regardless of how substantial it appears.
### Section 3 — Three claims with uncertainty labels
Section 3 is the synthesis output: three candidate causal claims derived from the Section 2 numerics, each labeled with an uncertainty level.
Each claim includes:
* **Claim statement**: a falsifiable causal assertion grounded in the document's evidence
* **Supporting evidence**: the specific Section 2 numerics that underpin the claim
* **Causal rung**: Rung 1, 2, or 3, reflecting the epistemic type of the claim
* **Uncertainty label**: one of `low`, `moderate`, or `high`, reflecting how confidently the evidence supports the claim
The three claims are not exhaustive — they represent the three best-supported and most causally specific claims the extraction engine can construct from the document. They are starting points for investigation, not final conclusions.
Uncertainty labels in Section 3 must be preserved when recording claims to the Claim Ledger. Stripping uncertainty labels is a governance failure and will cause the [Epistemic Dashboard](/workbench/epistemic-dashboard)'s uncertainty disclosure benchmark to fail.
## How results feed into the Claim Ledger
After Section 3 is generated, each claim can be recorded to the [Claim Ledger](/concepts/claim-ledger) by clicking **Record claim** next to it. The recorded entry carries:
* The claim statement
* The uncertainty label
* The source document (filename and extraction timestamp)
* The supporting Section 2 numerics as provenance
Claims recorded from PDF Synthesis are traceable to their source document at the numeric level — not just to the document as a whole. This fine-grained provenance is what distinguishes governed claims from informal notes.
## Difference from Hybrid Synthesis
PDF Synthesis and [Hybrid Synthesis](/workbench/hybrid-synthesis) share extraction infrastructure but serve different purposes:
| | PDF Synthesis | Hybrid Synthesis |
| ------------------- | ------------------------------------------------ | ------------------------------------------------ |
| **Input** | One PDF | Up to 6 PDFs and/or 5 company names |
| **Goal** | Deep extraction from a single source | Cross-source conflict resolution and novelty |
| **Output** | Three-section structured extraction | Novel hypotheses with confidence scores |
| **Claim recording** | Manual (click to record) | Automatic for top claim; manual for others |
| **Best for** | Auditing a single document's evidentiary content | Finding what multiple sources collectively imply |
If you have a single paper and want to know what causal claims it can support, use PDF Synthesis. If you have a corpus and want to know what those sources collectively imply beyond their individual conclusions, use Hybrid Synthesis.
## Related pages
Multi-source synthesis for reconciling conflicting claims across a corpus.
Understand how recorded claims are governed and exported.
Track evidence class distribution and uncertainty disclosure across your work.
Use extracted claims as starting points for SCM-grounded causal dialogue.