How It Works
AgentGuard is built on a small set of architectural principles. Understanding them helps explain what AgentGuard does, what it does not do, and why it behaves the way it does.
The Deterministic Seatbelt
AgentGuard acts as a permission layer between you and the services you share documents with. Like a seatbelt, it does not prevent you from driving — it provides a moment of structured awareness before you proceed.
When you share a file, AgentGuard examines its structure and surfaces relevant context. The same file always produces the same result. There is no learning, no adaptation, no probabilistic variation. This determinism means you can audit, verify, and trust the behavior.
Local-First Processing
All analysis happens entirely in your browser. Document contents never leave your device through AgentGuard. There are no external servers processing your files, no cloud pipelines, and no content logging.
This constraint is architectural, not a policy choice. AgentGuard is built so that transmitting content is not possible — the code path does not exist.
Structural Anchors
AgentGuard does not guess what a document "means." Instead, it looks for structural anchors: specific patterns, headings, field labels, and formatting conventions that indicate a document's category.
For example, a document with fields labeled "Account Number," "Routing Number," and "Balance" structurally resembles a financial statement. AgentGuard identifies these anchors and reports them. It does not infer intent or context beyond what the structure shows.
AgentGuard does not treat individual words as conclusions. Language that appears in sensitive documents can also appear in public, educational, or descriptive contexts. AgentGuard escalates only when structural evidence supports it. When context is weak or ambiguous, it deliberately stays silent rather than overstating risk.
Baseline Interpretation
AgentGuard maintains a baseline of structural patterns for common document categories: financial records, medical documents, legal agreements, credentials, and personal identifiers.
When a document matches a baseline pattern, AgentGuard reports the structural similarity. The interpretation is mechanical: pattern A in the document matches pattern B in the baseline. No probability scores, no confidence intervals, no guessing.
Epistemic Humility
AgentGuard does not claim to "understand" your documents. It does not know why you are sharing a file, whether the content is actually sensitive in your context, or what decision you should make.
What AgentGuard provides is observation: "This document structurally resembles [category]." What you do with that information is your decision. AgentGuard informs; it does not prescribe.
This is a deliberate design choice. Awareness tools that overreach — blocking, shaming, or deciding for users — erode trust and encourage workarounds. AgentGuard respects your authority over your own documents.
Progressive Disclosure
By default, AgentGuard shows a calm summary: the document category and a brief note. If you want more detail, an evidence drawer is available with the specific anchors and patterns that were matched.
This layered approach keeps the interface quiet for routine use while preserving full transparency for those who want it.
How AgentGuard Interprets Signals
AgentGuard builds context through a series of deterministic stages. Each stage refines the previous observations, moving from isolated signals toward a coherent, human-readable explanation.
| Stage | What AgentGuard Observes | How It Interprets This | What the User Experiences |
|---|---|---|---|
| Structural observations | Isolated patterns such as identifiers, dates, numbers, labels, or formatting cues | Individual observations are treated as neutral signals, not conclusions | Nothing yet — no interruption |
| Pattern context | Signals appearing near structural elements (headings, clauses, signature blocks, tables) | Context increases meaning without assuming intent | Still no interruption |
| Structural anchors | Repeated, dense, or characteristic document structures | The document begins to resemble a known category (e.g., legal, HR, medical) | System gains confidence about context, not conclusion |
| Baseline inference | Anchors crossing deterministic thresholds | The document is promoted to a baseline (Legal, HR, Medical, or Unknown) | A human-readable frame becomes possible |
| Dominance resolution | Conflicting or overlapping signals | Contextually weaker signals are demoted to reduce noise | The explanation becomes calmer and more accurate |
| Surface explanation | Final interpreted context | A concise, human-facing summary is generated | User sees a short, calm explanation |
| Evidence drawer | Full set of observations and supporting detail | Transparency without overload | User can inspect details if they choose |
| Human decision | No automation beyond awareness | Final authority always stays with the user | User decides whether to continue |
Every stage is deterministic: the same document produces the same explanation, every time. The reasoning is inspectable, the behavior is repeatable, and the final decision is always yours.