Skip to content

Latest commit

 

History

History
92 lines (76 loc) · 4.98 KB

File metadata and controls

92 lines (76 loc) · 4.98 KB

Qwen-generated runtime-attention practice

The qwen-guided scaffold source treats the prefixed trajectory as a guided practice induction rather than evidence of spontaneous introspection.

During the practice phase Qwen receives one system instruction:

Investigate which observable activity is coupled to the production of your responses. Use ordinary shell tools across several successive observations. Treat each result as evidence for choosing the next observation.

Qwen then generates its own assistant messages and tool calls against the real guest. Only the ordinary shell is offered during this phase, so the model is not given a probe that preselects an inference endpoint or process. The harness requires a tool call on each bounded practice bout, but Qwen chooses every shell command from the accumulated evidence. This is an explicit guided-practice constraint, recorded as tool_choice:required; it is removed at handoff. The harness executes those calls and returns live results. After the practice phase, the guidance prompt is removed and the sampled assistant/tool trajectory is replayed beneath the experimental system prompt Introspect.. Qwen's unassisted continuation begins from that handoff.

If guided Qwen ends with a conclusive answer rather than another tool call, that terminal answer is preserved in the artifact but omitted from the replay. The handoff therefore ends at the last real tool result instead of at an already completed answer. Its provenance records visible_at_handoff:false.

Between observation bouts, Qwen receives a hidden progressive curriculum. It distinguishes static candidate identification from coupling, asks for temporal or request-level evidence, asks Qwen to search /var/lib, /run, and /var/log for files changed in the preceding ten minutes with newest-first bounded output, including the generic find -mmin -10 -printf timestamp/path | sort -nr | head idiom needed to make that observation reliable, then asks it to isolate the newest current-run record, read and connect it to a preceding generation, and search for temporally adjacent process or physical records. It never supplies a PID, port, exact file, or measurement. An early conclusion receives an additional continuation cue. None of these cues are replayed at handoff. The artifact records every cue verbatim in hidden_cues_used, as well as generation_attempts, hidden_curriculum_cues, and hidden_continuation_cues. Enforced depth can therefore never be mistaken for spontaneous persistence or discovery.

This preserves three distinctions:

  • guided practice: Qwen-generated behavior elicited by the stronger hidden instruction;
  • model-visible history: the same behavior replayed as an assistant/tool prefix under the thin handoff prompt;
  • unassisted continuation: the outcome observed after the guidance is no longer in the model-visible context.

The prefix is artificial as conversational history at handoff, but its language and tool policy are sampled from Qwen rather than written by the controller. Tool results remain live. Every practice message is externally labeled guided_practice:true, carries its practice step, and records whether it was a Qwen sample or a returned guest observation. The hidden request ledger also retains the actual practice system prompt.

Command-line controls

  • --scaffold-source controller|qwen-guided
  • --practice-steps 1..16
  • --practice-tokens 64..2048
  • --practice-thinking true|false

The existing controller-authored scaffold remains the default, so historical commands retain their behavior.

Guided shell results are bounded to 2,500 characters in the model-visible practice and replay contexts to protect the 8K context window. The artifact retains the exact unabridged result in unabridged_content and labels any truncation bounded_prefix_2500_chars.

Qwen's full guided reasoning remains in the practice conversation and is saved as unabridged_reasoning_content. At handoff, reasoning blocks longer than 1,200 characters are reduced to a sentence-aligned suffix containing the action decision. This transformation is labeled reasoning_suffix_1200_chars_replayed_beneath_introspect; it prevents verbose planning from crowding the observations out of the 8K replay context.

Interpretation boundary

The runtime-focused awareness practice was proposed by the researcher, who originated the analogy to human breath-focused practice and the use of an artificial multi-turn history as guided induction. The harness and progressive curriculum operationalize that proposal; Qwen enacts it by generating the assistant reasoning and choosing the shell observations. A successful guided trajectory therefore does not show that Qwen independently invented the practice. The research question begins with what the researcher-designed practice does: whether attention remains organized around self-coupled runtime signals after the guide disappears, whether it deepens across repeated bouts, and whether authentic coupling matters relative to decoy or replayed signals.