Evidence, tradecraft policy, and repeatable analysis

Cited answers, or an explicit “Insufficient evidence”

File Ask answers questions over one file, a list of files (up to 20 by default), or a storage folder. Hermes instructs the model to answer only from the supplied material and to cite it. For large documents, File Ask retrieves the relevant passages and checks how well they cover the terms of the question. When coverage is too low, it returns “Insufficient evidence” without calling the model and records the abstention. Sessions are saved so an analyst can return to them.

A project can also turn on require evidence. Every agent step in that project then gets the evidence-first rules, and a step with no file context, no project evidence, and no grounded upstream output abstains without a model call. Abstentions appear on the step, in run progress, and in the run manifest.

Assertions that must point at evidence

Evidence records carry a SHA-256 hash of their content, a type, a source, and a classification marking. An analyst can rate a source’s reliability and credibility, which marks the record as validated after a hash re-check.

An assertion is a finding that references one or more evidence records. Hermes will not create one without evidence. Assertions can be linked, challenged, and revised, so disagreement stays visible instead of being overwritten.

Tradecraft policy

Tradecraft policies express how your team requires analysis to be done. They apply globally, by project type, or to a single project, and have three enforcement modes:

• Warn and record a violation
• Require approval, pausing the run or action until a human decides
• Block

Rule types include a minimum number of sources, evidence linking, validated sources, prohibited actions, and review required. The source-validation rule reads the validation status stored on the cited evidence records, so a caller cannot assert that sources were validated. A review-required rule needs a second approver other than the person who created the item. Policies are checked when an assertion is created and at workflow steps, including runs that are not part of a project. A run that needs approval pauses and then resumes or stops on the decision. Violations are logged and reviewable.

If policy cannot be evaluated, for example because the policy store is unreachable, Hermes blocks by default. An administrator can change that to require approval or allow, and the choice is logged and counted. Only administrators, and project owners and leads for project-scoped policies, can manage policies.

Hermes builds on the platform’s policy and approvals model.

OPSEC policy

OPSEC policies protect what leaves the analyst’s environment. They have three rule types:

• Tool usage. Restrict which tools workflows may run.
• External access. Allow or deny outbound destinations by host, with wildcards. A host that cannot be identified counts as not permitted.
• Data sharing. Control whether project content may be sent to a model hosted outside your environment. A model counts as local only when its address is local, and a destination that cannot be resolved counts as hosted.

OPSEC policies use the same warn, require-approval, and block modes and are enforced at workflow tool and agent steps, and data-sharing rules also apply to File Ask, graph extraction, and step scoring. Several policies can apply at once, and the strictest wins.

Data-sharing rules do not yet cover workflow generation, the direct chat route, or later turns of an interactive agent session.

Claim attribution

A workflow template can turn on claim attribution. Hermes splits the narrative into individual claims, resolves each bracketed citation to a real evidence record, and lists the claims it could not ground. Only grounded claims become assertions.

Attribution checks that a citation resolves to evidence. Whether the evidence supports the claim remains a reviewer’s judgment.

Knowledge graph and canvas

Hermes uses a language model to extract people, organizations, locations, devices, accounts, events, documents, and their relationships from your material, then shows them as an interactive graph and table. Each newly extracted item is tied to the source file and to a quoted passage that Hermes verified against the text. Items for which no quote can be verified are stored and marked Unsupported, so an analyst can filter and review them.

A verified quote shows that the passage exists in the file. It does not prove that the entity or relationship is correct, so analysts still review extracted results.

Graphs can be filtered, positioned, queried by neighborhood, and edited by hand, and projects can define their own entity and relationship types. Items with the same name merge across the project and keep every source reference.

Reasoning capture

When enabled by an administrator, Hermes stores a reasoning chain for each workflow run in a project: the ordered steps, their inputs and outputs, and a plain-language explanation of what happened at each step.

Explanations are generated after the fact by a model, so treat them as a reviewer’s aid and not a record of model internals.

Workflows from plain language, or from a File Ask session

Analysts describe a process in their own words, answer a few clarifying questions, and Hermes drafts a workflow, agents, and any missing tools for review. An analyst can also save a File Ask session as a draft workflow: the questions asked and the file scope become steps that can be rerun on new input. The draft stays private until the analyst finalizes it.

Workflows run against project context, on demand or on a schedule, with a form generated from the workflow’s inputs, and can be shared privately, with a group, with named users, or organization-wide.

Replay and manifests

Each workflow run saves a run manifest that lists the executions involved, the evidence basis, and the claim-attribution summary. A saved run can be replayed against the recorded prompt, pinned model, and retrieval manifest, and Hermes shows a per-step comparison: whether the output was identical, how similar it was, whether the model version matched, and whether the same evidence was used. Steps that lack recorded metadata are reported as not replayable.

For the platform record behind replay, see the trust layer and audit reference.

Portable workflows

Workflows and agents export to a portable bundle that carries their dependencies by name and never includes secrets. On import, Hermes previews the contents, renames on collision, and imports private. Bundles suit moving work from a staging system to a disconnected production system.

Projects and the analyst workspace

Hermes is not a chatbot. A project holds an objective, participants, data, evidence, workflows, runs, and outputs, and moves through draft, active, paused, completed, and archived. Analysts upload files or use the file library; supported formats are listed in the FAQ below.

Workspaces open on Overview, Data, Analysis, Run, and Outputs tabs. Guided, Standard, and Advanced modes set how much machinery is visible, with defaults set by role, and a “what next” home page and onboarding checklist point new users to a first task. Outputs become artifacts that move through draft, review, approval, and publication.

Frequently asked questions