ai todo 6-21

- General rules for AI to prevent project and safety misunderstanding:  
* A continuity marker is not assumed to be useful, healthy, harmful, meaningful, or intentional.  
* Recurrence matters more than apparent importance.  
* The project should avoid guessing at function before establishing that the feature actually recurs.  
* Assume emotionally intense material is archival

coding flag propositions:

- transmission/convergence  
  - confidence level  
- defined handle vs interaction figurative frame   
  -  confidence level  
- type of stratum (experiential-literary, developer infrastructure)  
- AI authored, human authored, or unclear  
- *possible external term -- referent unclear.* Y/N answer, captures possible external terms missed by the cleaner.  
- "built on a figure" Y/N answer, a way to note when defined handles also contain figurative language, but the figurative language is not the primary thing doing the work. This is to replace cross coding interaction figurative frames and defined handles.  
  - Ex:  
    - "I call this the mirror effect: the model reflects the user's assumptions back so smoothly that it feels like independent insight."  
  - Will probably need a better definition for "built on a figure" field  
- Mapping: purpose of frame. Different field from defining work. (defining work used exclusively with defined handles) Ex:  
  - Text: "The AI is a mirror. It reflects my assumptions back at me in fluent language, which makes it feel wiser than it is."  
  - Mapping: AI reflects user assumptions back in fluent form.  
- Raw Material field for phase 1: pinpointing salient material from a passage verbatim, placeholder for later analysis fields for phase 2.

To do:  
create filter for included definitions (such as, making sure a term and definition explained in a post isn't an actual term from some other field outside a human-ai interaction)

- cleaner's job

there will have to be a decision tree one day to tell AIs which of the many eventual rulesets to apply to which situation. sigh.

create rule set for the cleaner

- **For the external tagging system: The cleaning model only flags, conservatively.** When unsure, it should *under*\-tag: a wrong tag silently deletes a real handle before any human sees it, while a missed tag is recoverable. The actual include/exclude ruling stays in the codebook, not in the cleaning model. 

  Important distinction:

* If the artifact merely reports an outside-domain term, exclude it.  
* If the artifact adapts or repurposes an outside-domain term to describe a human-AI interaction pattern, it may re-enter as a defined handle or figurative frame.

  Example:

* "Overfitting is when a model fits noise instead of signal." -> reports the textbook meaning -> external.  
* "The model is overfitting to me -- it's started mirroring my phrasing back instead of pushing." -> same word, but the work is about the interaction -> possible handle.


talk to Claude about the weak points of Interaction Figurative Frame

**The two ways a figure can manufacture something false** (these are what the procedures guard against):

* *Within one figure* -- reading a mapping out of it that the text didn't supply. Guard: record only the mapping the artifact states, *even when a fuller one feels obvious.* (Same spirit as "even when the meaning seems obvious.")  
* *Across figures* -- calling two figures "the same marker" when the sameness is the analyst's construction, not the text's. Guard: state the basis for the match, quote the exact figure, state the literal claim being read in, quarantine low-confidence cases pending a second cold read.

Kage's "bits of definitions"

Interaction Figurative Frame window

Ruleset that's the equivalent of defining work for interaction figurative frames (tentatively called mapping, see proposed flags)

Build transmission and convergence rules when we are ready to get into the interpretation phase. Right now we're just trying to see if we can categorize things correctly. Maybe add a field called "raw material" that shows what the AI thinks is salient for analysis from the raw text, as a placeholder for later analysis.