Preserve evidence boundaries when turning research into slides in Orkas

A research handoff can preserve prose and lose its basis

A report and its slide deck can repeat the same sentence while meaning different things. A qualified finding such as “offline use depends on the selected model endpoint and tools” can become “works offline” during compression. A separate authoring stage needs the supporting excerpt and its qualification, not just a polished paragraph.

Orkas documents a desktop workflow in which Commander coordinates DeepResearcher, ContentWriter and PptMaker around a research goal. Its public README also describes local workspace files and a choice of supported model providers. Those capabilities provide a setting for the proposed handoff below; they do not establish that every generated citation or summary is correct.

Disclosure: This original, AI-assisted contribution was prepared on behalf of the Orkas project. It describes a source-based method and a fictional worked example, not an independent product review or a recorded execution.

Carry four things across the boundary

A useful handoff contains the question, the evidence, the permitted claim and the unresolved part. Keep the original reference material with the report so a later author can inspect the basis without reconstructing it from memory.

For each finding, record a short identifier, a source address or supplied document reference, the relevant passage, the interpretation and its limits. The identifier is a convention in the working files, not a claimed built-in Orkas citation format. If a source changes, record the checked date and review the interpretation again.

Commander can coordinate the research and writing stages, while the task asks each stage to retain those references. Reusable project context and agent memory may help subsequent work, but a remembered statement is not a replacement for the cited material.

Worked example: a bounded offline-use comparison

The following three excerpts and products are invented teaching inputs. They are not descriptions of actual vendors or results produced by running Orkas.

The brief asks: “Which candidate could support a document-summary demonstration on a disconnected laptop?” It supplies these excerpts:

• Reference A, fictional Cedar manual: “The desktop interface starts without a network connection. Summaries use a hosted model.”
• Reference B, fictional Birch setup guide: “A compatible local model endpoint may be configured. The endpoint and its model must already be installed.”
• Reference C, fictional Maple product sheet: “Local workspace storage.” No statement about inference or external tools is supplied.

An illustrative workflow asks DeepResearcher to extract what each excerpt establishes, ContentWriter to turn the findings into a short decision note, and PptMaker to prepare three slides from that note. The instruction restricts conclusions to the supplied excerpts, preserves reference identifiers and labels missing facts as unknown. It does not authorize new accounts, purchases, downloads or external data transfer.

A complete illustrative findings table is:

Candidate	Supported statement	Missing check	Decision for this demonstration
Cedar	The interface starts offline; summarization uses a hosted model.	Whether another inference route exists.	The supplied excerpt does not support disconnected summarization.
Birch	A compatible local endpoint can be configured after installation.	Whether this laptop has a working model, sufficient resources and network-independent tools.	Conditional candidate; verify the actual setup before selecting it.
Maple	Workspace storage is local.	Where inference runs and whether tools need a network.	Insufficient evidence for the requested demonstration.

The decision note can therefore conclude: “Only Birch has supplied evidence of a possible local inference route, and that route remains conditional on the installed setup. Cedar's supplied summarization route is hosted; Maple's inference route is unknown.”

The three-slide output has a concrete content boundary:

1. Question and evidence: the disconnected-summary requirement and references A–C, explicitly labelled fictional.
2. Candidate assessment: the three table rows, retaining Birch's installation condition and Maple's unknown inference location.
3. Decision and next check: the qualified conclusion above, followed by an actual offline summary check on an authorized local setup before making any capability claim.

These are proposed output contents, not files claimed to have been generated.

Inspect the report and deck together

Check that each slide assertion still has the same reference and qualification as the decision note. “Local storage” must not become “local inference”; “can configure” must not become “already configured.” If a slide omits a limit for space, shorten the claim as well or retain the limit in an accessible note.

For an actual task, preserve the report, deck and selected source material as inspectable artifacts. Check that cited addresses open and that the relevant passages support the wording. An attractive deck and matching wording across files do not themselves validate the source or the conclusion.

Storage, execution and cost remain separate

Orkas is a local-first desktop workspace, rather than an inference model. Using a remote provider still sends the relevant request to that provider; choosing a compatible local endpoint does not make every research tool network-independent. The MIT-licensed core and model-service charges are also separate: a free application does not imply free provider usage.

The official repository's README documents the supported setup and the roles used in this example. Its documentation is the starting reference for checking an installed configuration. The method here does not establish product performance, automatic evidence preservation or the feasibility of a particular laptop's offline workload.