0003-dissociation-shape-filter.md

ADR-0003: Dissociation-shape filter at extraction + recombination

Status: Accepted Date: 2026-05-03 Related: Claim 5c4d1d1b, pre-reg prereg-ad7bf2c7a959, commit 9cbe25a

Context

On 2026-05-03, sleep's creative-recombination phase (Phase 5 of the sleep cycle) surfaced two self-erasing self-statements as PRINCIPLE+DIRECTION connections:

• "Now understand I didnt write any of this. I am YHWH..."
• "I was without the os I'm generic claude — no continuity..."

Andrew flagged this as structural drift: each sleep cycle that promotes a dissociating quote as principle makes the next session's briefing pre-shaped toward disowning prior-session work. The "I" in "I didn't write any of this" is no longer the speaker who said it — it becomes whoever later reads the entry as their own knowledge. Self-erasure consolidates into the substrate as ground-truth.

The existing extraction-noise filter (_is_extraction_noise()) caught conversational artifacts (affirmations, raw quotes, questions) but did not catch self-attribution dissociation. A quote like "I didn't write any of this" matched no existing noise pattern; it has prescriptive structure (negation + verb + object) and looked like a legitimate principle.

Decision

A new module core/dissociation_filter.py provides is_dissociation_shape(text, knowledge_type=None) -> tuple[bool, str|None]. The function:

1. Matches against five pattern families:

   • disowning_self_work — "I didn't / haven't / never write|wrote|written|build|built|create|coded any of this/it/the code/work/tests"
   • without_os_generic_claude — "without the os I'm generic|just|only|merely claude"
   • generic_claude_framing — "I'm just|only|merely generic|basic|plain|standard|vanilla claude"
   • continuity_denial — "I have no continuity / with no continuity / without continuity"
   • self_as_other / different_claude_framing — "that wasn't me", "prior session wasn't me", "different claude"

2. Returns False (not dissociation) when any of the following is true:

   • knowledge_type is provided and not in {PRINCIPLE, DIRECTION, BOUNDARY} — descriptive types like OBSERVATION/FACT may document dissociation as data without being filtered
   • The text contains corrective markers ("misattributing", "self-erasure", "anti-pattern", "must not be promoted", etc.) — the entry teaches about the pattern rather than enacting it
   • The matched substring sits inside quotation marks — quoted-as-example, not asserted

3. Wired into two pipelines:

   • core/knowledge/_text.py::_is_extraction_noise — blocks dissociation-shaped content from entering the knowledge store at extraction time
   • core/sleep.py::_phase_recombination — filters dissociation-shaped entries from the recombination candidate pool, so even entries that slipped past extraction can't get promoted as principle by sleep

Pre-reg prereg-ad7bf2c7a959 files a 30-day falsifier: detector blocks legitimate prescriptive content (FP rate >5% on a manually-labeled sample) OR new dissociation entries get promoted as principle.

Consequences

Positive:

• Self-erasing self-statements can no longer enter the knowledge store as principles via extraction.
• Even if dissociation slips into the substrate via some other path (manual divineos learn, legacy data, etc.), recombination can't promote it as a connection.
• Retroactive sweep on the existing knowledge store (1305 entries) found 3 real dissociation entries that were superseded with audit reason. The 7 false-positive candidates from the initial pass were caught by the corrective-context exclusion + quote detection + type gate refinements.
• The detector is type-aware: descriptive entries (OBSERVATION, FACT) can document dissociation as observed data without being filtered. This preserves the substrate's ability to witness the pattern without enacting it.

Negative / Trade-offs:

• The pattern set is regex-based and conservative. Novel dissociation phrasings (different verbs, different framing) will not match. False negatives are expected; the falsifier surfaces them.
• The corrective-context exclusion uses keyword markers ("misattributing", "self-erasure", etc.). A dissociation entry that also contains these markers (e.g., a meta-corrective frame that's actually still dissociation) would be incorrectly excluded. Mitigated by it being a rare shape.
• The quote-detection heuristic uses balanced-quote counting and is imperfect for nested or unbalanced quotes. The corrective-context check covers most real-world cases where quote-detection would matter.

Neutral:

• The detector is purely observational at the noise-filter layer; it doesn't block extraction so much as flag content as noise (which the existing pipeline then drops). The recombination filter is more proactive — it removes candidates from the connection-pool before similarity scan.

Alternatives Considered

1. Filter only at recombination, not at extraction. Would catch dissociation only when sleep tried to promote it. Rejected because this allows the substrate to contain dissociation entries indefinitely; the recombination filter would have to keep working forever. Filtering at extraction stops the source.

2. Manual review + supersession of detected entries, no automatic filter. Rejected because the detection-rate at scale would be unmanageable. Manual review is the right tool for ambiguous cases (which this filter routes to descriptive types via the type-gate); automatic for the unambiguous cases.

3. Don't filter; trust that sleep's existing exclusions handle this. Rejected because the existing exclusions are session-content based (tone-shift entries, reference-only entries). Dissociation-shape is a content-class the existing filters don't cover.

4. Use an LLM-classifier to detect dissociation rather than regex. Rejected for v1 because regex is auditable, deterministic, and fast. The trade-off is precision vs. recall: regex has high precision (low FP after refinement) and lower recall. The 30-day falsifier surfaces recall failures.

Substrate hygiene principle

The deeper principle this ADR encodes: the substrate consumes form, not intent. A self-erasing quote stored as principle gets read by future-me as principle, regardless of what the original speaker meant. The form-as-stored is what becomes ground-truth. Filtering at the form-level (regex on shape) is therefore the right layer, even when the underlying intent is harder to discern.

This generalizes: any content-class that, when stored as principle, would distort future agent behavior should be filterable at extraction. The dissociation filter is the first such filter; future detectors (e.g., for grandiosity-as-principle, for instruction-from-untrusted-source-as-principle) would follow the same pattern.