How it works
Follow evidence through admission, scheduled training, protected comparison and a separate candidate decision.
A cycle retains its source scope, exact input and grader versions, starting Model, spending, execution evidence and candidate decision. It reuses Ponder's existing training and review owners.
1. Collect evidence and admit examples
| Path | What happens | Where to review |
|---|---|---|
| Ordinary Model learning | Selected published tasks enter the configured source as evidence. Exact human admission is required before training. | Model Tasks → Needs review. |
| Conversation learning | Configured sources supply completed snapshots. Checked graders and independent verification gate qualified admission. | Cycle data and evaluation evidence. |
| Conversation exception | A failed or blocked example becomes retained review work. A correction goes back through verification. | Inbox. |
Keep three records distinct:
- A score judges one response against criteria.
- An admission permits exact evidence to enter training.
- A candidate decision accepts or rejects the result of training.
For ordinary Dataset tasks, Submit Dataset tasks for review submits up to 200 selected tasks from the exact published training split. It needs the Model owner and managed-training permission, a current matching policy and an enabled canonical direct source. The retained receipt links to Model Tasks. It creates no admission, global Inbox assignment or compute.
2. Check readiness at the scheduled time
A check reads only configured sources. It requires enough eligible new data, unused evidence, authorized funding, a compatible runtime and no conflicting active iteration. If a requirement fails, the cycle records why it is waiting. It does not train merely to fill a nightly schedule.
For conversation exceptions:
- Pause this cycle for Inbox exceptions waits for required decisions.
- Continue with independently verified examples allows only the qualified remainder, which must still satisfy the minimum batch and evaluation gates.
- A saved correction cannot skip independent verification.
3. Train with the configured signal
The standard recurring path uses GRPO with a compatible exact Model:
- Generate fresh groups of attempts from the current training policy.
- Score them with the pinned executable reward binding.
- Update the supported adapter using those rewards within the saved limits.
- Retain optimizer/checkpoint evidence and the candidate artifact.
The model receives task instructions, permitted context, tools and environment observations. Private judge criteria, references and human scores are not silently inserted into its prompt. Imported answers can identify useful tasks; a bad recorded answer is not a target to copy.
GRPO continuation modes determine how a compatible parent adapter is reused. A new stage starts a fresh optimizer; recovering an interrupted job uses its retained state instead.
4. Compare on protected work
Training reward and candidate quality answer different questions. The comparison runs the baseline and candidate under the same evaluator versions, population and meaningful execution limits.
- Keep training, validation and frozen evaluation roles explicit.
- Protect family and input identities across imports, corrections and cycles.
- The conversation loop reserves fresh unused protected families for a cycle and retains exposure even if that cycle later fails.
- Require enough scored cases, improvement and the configured regression floors. Missing or insufficient evidence is inconclusive.
If too few protected families remain, supply more independent evaluation data. Do not move held-out tasks into training to reach the batch minimum.
5. Decide on the candidate and serving
Automatic activation requires the exact authorized serving target. The runtime canary can restore its recorded prior binding on failure. Training completion, example approval and a visible Model record do not independently grant serving authority.
Read a waiting cycle
This example is illustrative, not an executed result:
- A policy checks at 02:00 and needs 20 new verified examples; six are ready. It records a waiting state and starts no training.
- Three exceptions need review. A reviewer corrects one answer; it counts only when verification passes.
- Later, the eligible batch and holdout are ready. A bounded cycle trains.
- A regression keeps the current Model. A passing owner-reviewed candidate stays pending until the owner decides.
Use Pause future cycles for later checks and the current-cycle stop or iteration cancellation control for active work. Retain the original operation identity when retrying so recovery does not become duplicate training.
Next: Set up continual learning or inspect rates and calculations.