Shared Reports¶

Shared Reports is the library you reach for when you remember the finding but not the thread that produced it. It collects durable outputs you created and reports other researchers shared with your organization.

A report might be a diagram you refer to in every model review, an explorer that makes a behavior easy to reproduce, or a longer analysis that settled a research decision. The library keeps those outputs useful after their original experiment is no longer active.

Treat it as a research library¶

The report title should help a future reader recognize the question or result, not merely the run that created it. “Probe sweep 04” may make sense today; “Where hallucination probes transfer across domains” will still make sense six months from now.

Use Shared Reports to:

revisit a result before planning related work
find the canonical explanation to send a teammate
compare reports that tested similar questions under different conditions
attach established evidence to a new thread
check whether a conclusion has been updated since you last used it

Find a report¶

Open Shared Reports and search by the scientific question, model, dataset, or intervention you remember. Filters narrow the library to all reports, reports you created, or reports shared with you.

The list shows title, creator, access, and last update. The update time matters: a familiar title may now point to a better explanation or a restored earlier version.

Inspect provenance¶

When a report is relevant to a new decision, first read it on its own terms. Identify the main claim, strongest evidence, limitations, and recommended next step. Then use its references to inspect the source thread or experiment when the assumptions matter.

You do not need to audit every source file for every casual read. Go deeper when you are about to spend compute, compare against the result, or communicate it as established evidence. At that point, verify the model checkpoint, data split, metric, controls, sample size, and relevant source files.

If the report conflicts with newer work, do not choose the more polished one by default. Compare the experimental conditions and ask the thread to synthesize the disagreement.

Open the report and choose the audience before copying its link. An organization report stays available to members and follows the current working state. A public report publishes a saved version that anyone with the link can open; later edits remain private until you update that public version.

Share the report when the output can stand on its own. Share the thread when the reader needs the discussion, decisions, or sequence of experiments that produced it.

Continue from a report¶

Attach a report to a thread when you want to build on it. State the new question and any changed inputs so Silico can treat the report as evidence rather than starting from its conclusion alone.

For example:

This report found a stable representation in the base model. Test whether it survives fine-tuning, using the same layer range and evaluation prompts but the checkpoint attached here.

The attached report is read-only. New work can cite and inspect it, but it does not modify the source report or give the new thread access to unrelated research.

Curate what deserves to last¶

Not every temporary visualization needs to become a long-lived reference. A report belongs in the shared library when another person could reasonably return to it for a result, method, comparison, or decision.

Before sharing broadly, give it a descriptive title, make the conclusion legible without the source conversation, and include the important boundary conditions. A small amount of curation now is what turns an experiment output into team knowledge.