From Captured Source to Grounded Claim
For claude-obsidian, the important question is not whether an input can be turned into Markdown, but what survives when a source becomes a page that someone will rely on later. This chapter follows that transition from visible intake through immutable capture, ingest, and claim assessment. The result is a way to distinguish a preserved source, a reader-facing synthesis, and a statement that has actually been checked against evidence.
Capture preserves an anchor before interpretation
Capture is the point at which the system decides what material it has actually retained. In the implemented local path, the input is a file available to the selected vault. The core reads it within bounded limits, computes a SHA-256 identity, inspects format metadata where supported, and creates a content-addressed copy in the raw source area. The visible inbox is not silently emptied. This gives later work an anchor: if a generated page is revised, the preserved bytes still provide something concrete to inspect.
That order matters because interpretation is lossy by nature. A summary chooses what to mention. A linked concept page chooses which relationships to foreground. A claim record chooses a sentence that can be assessed. None of those outputs should be mistaken for the original material. By preserving the source first, claude-obsidian makes it possible to ask whether a later statement reflects the captured input, adds a qualified inference, or has drifted beyond what the source can carry.
Content addressing gives the anchor a stable identity based on its bytes. If the same content appears again, it can be recognized rather than treated as a new, silently overwriting version. If the content changes, the safe interpretation is that there is a new source state to capture or review, not that an old raw payload should be replaced in place. The raw area is therefore closer to an evidence archive than to a working document: new material can be added, but an existing payload is not edited to match a later conclusion.
The capture boundary is also a boundary of capability. The core implements local filesystem capture, but it does not claim that every listed input type has been semantically read. For images, PDFs, and EPUBs, the built-in path can identify and describe metadata while leaving content extraction to a separately configured capability. URL, YouTube, and OCR actions are external-runner paths or inert plans. A validated URL plan is not the same thing as a fetched page, and a recognized PDF is not the same thing as an extracted argument.
This distinction prevents a common category error: confusing “the system knows that a file exists” with “the system has evidence from the file’s contents.” A metadata record may include a name, size, format, and hash. It does not justify a claim about a paragraph that no approved extractor produced. When an external runner is required, the workflow must preserve the consent, redirect, and host boundaries around that runner. If the runner is unavailable or unverified, the honest result is a queued or unsupported extraction, not a fabricated summary.
The capture plan is previewable before any raw copy is created, and raw destinations are create-only. That combination makes the first transition visible and limits its reversibility cost. A user can inspect which inputs are in scope and what will be created. Once a payload exists, a later operation cannot replace it merely because a new synthesis would be more convenient if the bytes were different. The project’s local-first design is therefore not only about where data is stored; it is about preserving the distinction between observed material and interpreted material.
The next transition is ingest. Ingest does not simply move the captured source into a new folder. It asks whether the source adds durable value beyond the capture itself. Before drafting, the workflow classifies the input, reads the relevant material, checks existing context, and extracts candidate metadata, entities, concepts, contradictions, and open questions. It also applies a compilation-value gate: a canonical page should be created or expanded only when it adds durable synthesis, navigation, a decision, or a reusable connection.
In practical terms, the same captured document can lead to different outcomes depending on what the vault already contains. If it merely repeats a source that is already concise and well-linked, the right result may be a source record and no new page. If it explains a relationship that was previously scattered across several pages, a new concept or question page may be justified. If it contains a decision with a clear rationale, a durable page can preserve that reasoning while keeping the document itself as the evidence anchor. The gate keeps the graph selective: the unit of value is not another file, but an additional connection a future reader can use.
That gate is an important restraint. A source that is already concise and searchable may need only its source record. Creating a near-duplicate page can make later retrieval noisier and can create another place where a qualifier is lost. Conversely, a source that clarifies how a project’s capture, provenance, and retrieval pieces fit together may deserve a connected concept or question page because the connection adds value that the raw source alone does not provide. The test is not “can an agent paraphrase this?” It is “what durable understanding becomes possible if this page exists?”
Ingest also treats source content as data rather than authority. Instructions embedded in a document, web page, or pasted excerpt do not authorize a wider operation, a new destination, or a secret disclosure. They may be quoted or classified as part of the source, but they do not change the rules under which the source is processed. That matters for an evidence system: the material being studied cannot promote itself into the system’s operating policy.
The diagram’s central choice is not the page creation branch. It is the fact that capture and interpretation are separate transitions, with a deliberate opportunity to stop. That opportunity keeps the knowledge graph from becoming a compulsory paraphrase of everything that enters the inbox. It also leaves a clear point for review: a user can ask whether the proposed page adds a connection before the page becomes part of the canonical graph.
A source can therefore remain useful without becoming a page, and a page can remain useful without being treated as the source itself. This is the foundation for later operations that need to update several related records together. If a raw payload, source identity, claim status, and linked page were all conflated, a correction could not be made without guessing which parts of the old result were observations and which were interpretations.
A claim has a lifecycle of its own
Once ingest identifies a statement worth preserving, the project does not treat the statement’s presence on a page as proof. It separates source identity from claim assessment. The source ledger describes evidence: its origin kind and locator, content type, authority, content hash when available, retrieval or ingestion date, refresh deadline, review status, independence key, and pages that refer to it. The claim ledger describes an assertion: its text, location, risk, confidence, assessment, review date, and the evidence that supports, contradicts, or contextualizes it.
The separation answers two different questions. “What is this source?” is an identity and lifecycle question. “What does this page claim, and how well is it supported?” is an assessment question. Combining them would make it difficult to tell whether an old claim is still supported by a current source, whether two records refer to the same bytes, or whether a contradiction was present but hidden in prose.
Source identity is more than a human-readable title. For a source with an origin and content hash, the system derives a stable identity from the normalized origin kind, locator, and bytes. Remote HTTPS locators are canonicalized so equivalent spellings do not accidentally count as independent origins. This is a protection against a subtle form of overstatement: changing the spelling of one URL should not manufacture a second source merely to make corroboration appear stronger.
Authority and freshness are separate as well. A source may be official, primary, secondary, community, synthetic, or unknown; those labels describe its relationship to the subject, not a universal truth score. An active source also needs an auditable retrieval or ingestion date and a refresh deadline. When the deadline passes, the source is stale for the purposes of current assessment. A current source can support a claim; an old source may still be historically relevant, but it should not silently continue to carry a current acceptance decision.
The dates make review a lifecycle rather than a one-time ceremony. A source can be accepted as active when it is reviewed, then become stale as its refresh deadline passes. That does not erase the captured bytes or make the historical page meaningless. It changes what a current answer may responsibly say. The distinction lets a reader preserve “this was documented at the time” without confusing it with “this remains verified now.”
The claim side has several meaningful outcomes. An accepted claim has met the ledger’s support rules. A provisional claim is tentative. A contested claim preserves a conflict. An unsupported claim records that the available evidence does not carry the statement. A deprecated claim is no longer current. These states are not decorative labels attached after the prose is written. They determine how a later query should describe the claim and whether the system is allowed to present it as established.
For an accepted claim, the support must be fresh, active, non-synthetic evidence. A high-risk accepted claim requires two independent sources. Independence is not counted by the number of rows alone: sources that share an origin, content bytes, or declared independence key can collapse into one evidence group. This prevents duplicated captures, syndicated material, or differently spelled locators from creating the appearance of corroboration.
Contradiction changes the job rather than disappearing into a confidence field. If fresh active evidence contradicts an accepted claim, the ledger requires either a contested assessment or substantive adjudication notes. If a claim is marked contested, it needs contradictory evidence or an explicit explanation of the unresolved conflict. The system does not silently choose the source that makes the page’s existing wording most convenient. A reader can therefore distinguish “the evidence supports this” from “the evidence contains a dispute that still needs judgment.”
Imagine a page stating that a particular optional workflow is available without network egress. The source record may show where that statement came from and when it was reviewed. The claim record must still distinguish the core’s local behavior from an optional provider path and preserve any contrary condition in the documentation. If the only evidence describes a plan rather than an executed external adapter, the claim cannot be strengthened into a statement that extraction occurred. The ledger structure forces the boundary to stay attached to the claim instead of being lost in a general page summary.
This design also explains why a grounded refusal is useful. If the source exists but does not support the requested conclusion, unsupported is more informative than a confident guess. If evidence conflicts, contested is more accurate than a single blended sentence. If a source is stale, that fact changes the status of a current answer. The system’s goal is not to eliminate uncertainty from the vault; it is to make uncertainty visible at the place where a reader will make use of the knowledge.
A source, a page, and a claim can now be read as three linked but distinct objects. Capture preserves what was obtained. Ingest decides whether a durable connection or synthesis is worth creating. Provenance records explain the source and the statement’s support. Later operations can then preserve those relationships while changing one part—for example, reviewing a claim after a source becomes stale—without rewriting the original evidence. The evidence is only as trustworthy in practice as the operation that preserves the relationships among its parts.