01 / PREFIX ONLY
Show what was knowable
Include the task, earlier turns, tool calls, and tool results available before the target action. Exclude later tests and final answers.
Causal annotation playground
Complete the missing why.
Give the annotator only the observable conversation prefix and the action that already happened. It writes a short rationale explaining that action—without peeking at the future.
Observable prefix
Test expects negative timeouts to be rejected. The current parser returns int(value) directly.
Missing rationale · complete this
<think> … </think>
Chosen action · fixed target
Edit the parser to validate seconds < 0 and raise ValueError.
The causal contract
CoT completion is an annotation problem, not hidden-state extraction. The action is fixed before annotation, and future observations are excluded by construction.
02 / ACTION FIXED
Explain; never choose again
The annotator receives the exact action that occurred. Its job is to justify that decision, not replace it with a hindsight-informed action.
03 / PROVENANCE
Label synthetic reasoning
Every output carries synthetic_rationale_not_teacher_cot. It may train a student, but is not the teacher's private thought.
This page does not recover private chain-of-thought.
It generates a compact, causal rationale from observable evidence—the same provenance distinction used by the Fable + Kimi trace pipeline.
One trace, completed
The annotator sees only the left side plus the recorded action. The right side is newly generated annotation, never a claim about hidden teacher reasoning.
Inputs
User taskobservable
parse_timeout accepts negative values, but timeouts must be non-negative. Fix it while preserving positive-value behavior.
Tool result · rgobservable
src/config.py:41:def parse_timeout(value):
src/config.py:42: return int(value)
tests/test_config.py:18:pytest.raises(ValueError)
tests/test_config.py:19:parse_timeout("-1")<think> ? </think>Missing explanation before the action.
Chosen assistant actionfixed
EDIT src/config.py
seconds = int(value)
if seconds < 0:
raise ValueError("timeout must be non-negative")
return secondsCompleted training turn
Synthetic rationalenot teacher CoT
The prefix shows that integer conversion already works and that the missing behavior is specifically rejecting negatives. The smallest compatible change is to parse once, validate immediately, and preserve the existing return path for non-negative inputs. Raising
ValueError matches the test expectation visible in the prefix.Recorded actionunchanged
EDIT src/config.py
seconds = int(value)
if seconds < 0:
raise ValueError("timeout must be non-negative")
return secondsProvenanceretained
rationale_provenance: synthetic_rationale_not_teacher_cot future_visible: false action_source: recorded_teacher_action
Live completion lab
Use an Anthropic API key to annotate one prefix/action pair. One ordinary Messages API call returns one provenance-labeled synthetic rationale.
Completed training turn
Rationale and recorded action stay separate until export.
No live completion yet
Run the example inputs, or replace them with one of your trace nodes.
Run the example inputs, or replace them with one of your trace nodes.
1 · Synthetic rationalenot teacher CoT
—
Input scopeprefix + action
Action sourcerecorded
Provenancesynthetic
Inspect the exact annotator prompt
—
2 · Provenance-rich SFT recordexportable
—
What this can establishThe rationale is consistent with the available prefix and the action that actually occurred.
What this cannot establishIt does not reveal what the behavior teacher privately thought or prove that the teacher used the same explanation.
What to validate at scaleHash prefix and action, scan for future overlap, preserve tool-call closure, and retain the provenance label.