Vince/evidence source indices - #935
Conversation
|
Important Review skippedDraft detected. Please check the settings in the CodeRabbit UI or the ⚙️ Run configurationConfiguration used: Organization UI Review profile: CHILL Plan: Pro Run ID: You can disable this status message by setting the Use the checkbox below for a quick retry:
✨ Finishing Touches🧪 Generate unit tests (beta)
Comment |
- Renamed 'atomic facts' to 'molecular facts' - Added decontextuality (stranger test) instruction with enduring vs incidental descriptors - Added minimality instruction (only enough context for interpretability) - Removed proposition gate (redundant with d + m gradient) - Updated examples to show molecular ideal
Each explicit observation now carries source_indices — 0-based indices into the <messages> block pointing to the messages that directly support the conclusion. Messages in the prompt are prefixed with [N] indices. This enables the action-based judge's evidence entailment criterion to slice the conversation to just the supporting messages instead of using the full conversation as evidence context (degraded mode). No text duplication — the LLM outputs indices (cheap, low-error), and the judge looks up the real message text from the trace input.
- Clarify that source_indices can point at any speaker's messages, not just the user's — what matters is which messages contain the evidence - Add two examples: assistant-sourced attribution and multi-message attribution showing question + recommendation + confirmation - Addresses issue where deriver cited wrong messages for conclusions derived from assistant responses
4a28e4f to
5bce6b9
Compare
There was a problem hiding this comment.
This prompt deviates greatly from the original prompt.
Needs to be tested against the current Neuromancer model to see if any negative effects occur.
Maybe we should keep this prompt out from this PR and scope this PR only to the batch message.
There was a problem hiding this comment.
@vintrocode I know you did some testing. Good to remove the prompt from this PR or is it required for the other changes?
There was a problem hiding this comment.
Note that if we do revert most of the prompt changes, we will still need to update the prompt to give the model context about source_indices .
VVoruganti
left a comment
VVoruganti
left a comment
There was a problem hiding this comment.
So main thing to note here is that the batch_message_ids that were added as a new field in this draft would include also the IDs from the interleaved messages from other speakers.
This means that an explicit conclusion for peer A can reference a message made by peer B in the conversation. Historically we have been authorative in saying a conclusion can only be made for a peer based on something the Peer has actually said, otherwise any claim by a separate peer could lead to nonsensical conclusions being made
A few things important to clarify: 1.) The context from which conclusions are derived hasn't changed — the deriver has always seen other speakers' messages as context ("Extract ALL observations from the target peer's messages, using others as context" is the pre-existing rule and the prompt is built from every message in the batch, each labeled with its speaker). This PR just makes the model declare which messages it used. 2.) A citation including another peer's message doesn't mean the conclusion rests on their claim. Specifically, the target peer's statements are often uninterpretable alone ("the first one" in response to "do you prefer Python or Rust"). Because the entailment judge only sees cited messages, the interpreting context has to be cited or valid conclusions fail. In both prompt examples the peer's own message is always among the sources. If anything this makes the "only from what the peer said" stance enforceable for the first time — a conclusion citing zero target-peer messages is now mechanically detectable. 3.) Your comments made me realize that the design of |
4 participants
make the deriver indicate which messages a conclusion came from via source ids