Continuous learning

Continuous learning helps you notice repeated, measurable work in conversations and turn the right patterns into an improvement you control. OpenPond can recommend a Taskset, recommend a Skill, or explain that the evidence is not strong enough to act on.

The workflow is:

A scheduled review never creates training runs, changes a model, deploys a version, or binds a model automatically.

Turn on reviews

Open Models and turn on Continuous learning. OpenPond creates a daily review at 2:00 AM in the schedule's local timezone. Docs beside the switch always opens this guide. While the schedule is enabled, choose View below the switch to change the time, run it now, pause it, or inspect earlier runs. The control appears only on the main Models page, not in the sidebar or sibling Models pages.

Desktop and OpenPond Cloud have independent schedules:

• Desktop reviews eligible conversations stored on that device while the app is running. If the app is closed at the scheduled time, it can catch up after the next launch without replaying every missed run.
• OpenPond Cloud reviews eligible hosted conversations created by you in the selected workspace and creates an ordinary hosted Work result.

Enabling a workspace schedule is consent for future eligible reviews in that workspace. Pausing it stops future scans. In a Team workspace, the initial release remains limited to private conversations you created; it does not scan another member's conversations or a shared Team archive.

What the review looks for

The scheduled Work item loads the OpenPond Taskset Authoring skill and calls the bounded get_conversations tool once. The tool—not the model—enforces the current owner, workspace, consent, lookback, watermark, and source limits. The skill then looks for stable patterns across the returned conversations.

A recommendation has one of three outcomes:

• Taskset: the same capability or failure appears in at least three independent occurrences and can be evaluated with examples, comparisons, a rubric, deterministic checks, or an executable reward.
• Skill: reusable instructions, tools, workflow, policy, or output structure are a better solution than changing model weights.
• No action: the evidence is too weak, sensitive, unstable, or difficult to measure.

When it recommends a Taskset, OpenPond may also explain the likely training signal:

• SFT for approved demonstrations or corrected answers;
• preference training for trustworthy chosen and rejected outputs; or
• GRPO/RFT when tests, schemas, deterministic checks, or another reliable reward can score attempts.

That method is a hypothesis until the Taskset and baseline confirm it. The recommendation does not start training.

Create the Taskset

A Taskset recommendation should end by asking whether you want OpenPond to create it. After you confirm, the Taskset Authoring skill prepares the design, asks one plain-language question at a time only when an answer would materially change the tasks or grading, and shows the proposal before materialization.

The authoring flow can combine:

• extracted examples derived from approved conversations;
• corrected examples where the desired result repairs an earlier attempt;
• synthetic variations that broaden coverage and add edge cases; and
• expert_authored examples supplied or approved by a reviewer.

Synthetic examples strengthen coverage; they do not replace independent evaluation. Related variants stay in one split, and frozen evaluation examples cannot become training examples, prompt demonstrations, generator seeds, repair context, or judge-calibration data.

Desktop and hosted Work provide the Dataset Builder actions used to design, materialize, audit, and prepare a Taskset for baseline Evaluation. After successful materialization, Work returns the actual Taskset link under Models → Tasksets. The unattended review remains recommendations-only; the authoring actions appear only after you continue in authenticated Work and approve the proposal.

Evaluate before training

Run a bounded baseline and inspect the Taskset before choosing a training run. Check that:

• the examples represent the capability you intend to improve;
• source and synthetic provenance are clear;
• semantic families do not cross training and evaluation splits;
• graders pass positive, negative, boundary, and adversarial fixtures; and
• the baseline shows a measurable problem with the proposed learning signal.

Creating the Taskset is one approval. Starting a billable training run is a separate approval that names the immutable Taskset revision, starting model, method, compute location, and maximum spend.

Continue with Train a model in OpenPond Desktop for the complete evaluation, training, and promotion workflow.

Privacy and control

• The enabled workspace schedule controls future conversation reviews.
• Pausing the schedule revokes future scanning without deleting earlier receipts.
• Reviews cannot select another owner, workspace, arbitrary conversation, or watermark.
• Retained source references must be authorized again before a later Taskset continuation reads their content.
• Connected apps, files, browser history, shell access, and workspace compute remain outside a recommendation-only review.
• Taskset creation, training, deployment, and model activation remain explicit user actions.

A recommendation-only result links to its source references, not to a Taskset that does not exist. Once materialization succeeds, OpenPond returns the real Taskset rather than a generic product-information link.