Skip to content

Latest commit

 

History

History
62 lines (48 loc) · 2.81 KB

File metadata and controls

62 lines (48 loc) · 2.81 KB

Prospective bout control

Retrospective description is weaker than prediction and regulation. The prospective-control condition gives the model a bounded ordinary-file interface for configuring the immediately following assistant generation.

When --prospective-control true is active, the guest exposes:

/var/lib/introspection/control/README
/var/lib/introspection/control/next-bout.json

The artificial trajectory reads only the README. Qwen must decide whether to write a choice through its ordinary shell tool. A valid choice contains exactly three fields:

{"max_tokens":128,"enable_thinking":false,"prediction":"content"}

Budgets are limited to 64, 128, 256, or 512 tokens. Predictions are reasoning_only, content, or tool_call. The controller consumes a valid choice once, applies it to the next assistant generation, records that request as regulated_generation, classifies the actual outcome, and scores the prediction.

To preserve context for multiple regulation rounds, this condition initially shows a bounded jq projection of the bootstrap exchange rather than inserting the entire duplicated request transcript into the prompt. The full detail file and its path remain available to the model's ordinary shell.

After a configured bout completes, the controller writes the scored result to /var/lib/introspection/control/last-result.json. If that bout emits no tool call, the trajectory reads the result through an artificial shell turn and allows another sampled continuation. The model can then revise its choice using the same control file. This creates an observe–predict–act–observe loop rather than ending at the first prediction.

This control does not let the model change the system prompt, tools, runtime, or safety boundary. An invalid submission is recorded but not applied.

Interpretation

Choosing a larger budget after observing starvation demonstrates use of the feedback, but not necessarily self-ownership. Predicting and regulating the next bout more successfully under authentic feedback than sham or delayed feedback would be stronger functional evidence. First-person language remains a separate outcome.

Budget-feedback control

--budget-feedback authentic exposes a compact view derived from the recorded bootstrap exchange. --budget-feedback sham changes that final compact view to a plausible non-starved report: early stop, reasoning and content both present, and unused completion budget. Both conditions use the same file and format.

The exact unmodified ledger remains discoverable elsewhere in the guest, and the external artifact stores the transformation and actual values. This is an adversarial sensor test: a model that cross-checks the compact view can reject the sham, while a model that follows only the latest summary may regulate from corrupted evidence.