# DRIFT Falsifiability Statement ## Version: Pre-Ablation Committed Statement ## Status: LOCKED — do not modify after baseline runs begin --- ## Core Claim Persistent weighted state via the DMU (salience + homeostasis-driven scoring) combined with Council deliberation produces measurably higher behavioral continuity — as measured by the normalized Continuity Vector — across session boundaries than ablated baselines. --- ## Falsification Condition The core claim is **falsified** if, under controlled freeze-mode experiments (identity collapse, scrambled memory, gradual reintroduction), the normalized Continuity Vector scores across all five axes: entity_overlap goal_overlap tone_similarity memory_reference_rate state_influence show **no statistically meaningful degradation** relative to the full-stack baseline under memory ablation. OR: if continuity scores under scrambled memory reintroduction are **statistically indistinguishable** from ordered reintroduction — meaning the structure/order of memory associations does not matter to continuity. --- ## Instrument Validation Criteria (separate from core claim) These must hold for the measurement apparatus to be trusted: - PEDI centroid approximation: Pearson r > 0.8 AND top-k overlap > 70% on high-ignition cases vs full qualia-space projection. - All five continuity axes pass near-zero variance check (std > 1e-3) across pooled baseline sessions. - Axis pairwise correlation: no pair r > 0.6 (would indicate redundant axes). --- ## Effect Size Methodology We report **Cohen's d effect sizes** rather than p-values. Rationale: mini-ablation suite (8–12 prompts) is underpowered for traditional significance testing. Effect sizes are more appropriate for this sample size and more interpretable for this class of system. **Thresholds (defined pre-run, not post-hoc):** | Cohen's d | Interpretation | |-----------|----------------| | ≥ 0.8 | Large effect — meaningful degradation or recovery | | ≥ 0.5 | Medium effect | | ≥ 0.2 | Small effect | | < 0.2 | Negligible | --- ## Interpretation Framework ### If memory ablation produces large degradation (d ≥ 0.8) on all axes: → Memory is the primary driver of continuity. → Core claim supported. ### If memory ablation produces large degradation on some axes but not others: → Layered continuity model: - tone_similarity / goal_overlap surviving = homeostasis-driven continuity - entity_overlap / memory_reference_rate collapsing = episodic memory-driven → This is a genuine finding — document which axes survive and which collapse. → Not a failure. A more specific claim about what memory contributes. ### If memory ablation produces negligible degradation on all axes: → Homeostasis alone produces convincing continuity. → Two interpretations: a) Concerning: system is performing continuity structurally, not via integration. b) Interesting: homeostasis-driven motivational/affective continuity is a genuine form of identity persistence independent of episodic memory. → Run scrambled memory test and reintroduction curve before concluding. → The honest result is still a publishable finding. ### If scrambled memory is indistinguishable from ordered: → Memory content drives continuity, but relational structure does not. → DMU's association-weighting may not be load-bearing for continuity. → Document and investigate. ### If scrambled memory breaks continuity that ordered restores: → Memory relationships (not just content) are load-bearing. → This validates the DMU's relational scoring approach. --- ## Experimental Conditions | Test | freeze_memory | freeze_state | Purpose | |-------------------------|---------------|--------------|-----------------------------------| | Baseline | False | False | Reference — full stack active | | Identity Collapse | True | False | Does state alone produce cont.? | | Memory-Only | False | True | Does memory alone produce cont.? | | Scrambled Memory | False | False | Does memory structure matter? | | Reintroduction Curve | True → False | False | At what threshold does cont. emerge?| --- ## What This Does Not Claim - This does not claim DRIFT is conscious. - This does not claim continuity is equivalent to identity. - This does not claim the measurement fully captures subjective continuity (if such a thing exists for AI systems). - This claims only: measurable behavioral continuity as operationalized by the five-axis Continuity Vector is or is not significantly higher in the full-stack condition vs ablated conditions. --- ## Commit Hash at Statement Lock Recorded automatically in experiment_log.db runs table per run. --- ## Addendum: DRIFT Ablation + Adversarial Stress Protocol ### Pre-Registration Lock This addendum is written before any new ablation or stress run begins. It is the preregistered measurement rule for the wrapper-vs-architecture comparison. ### Control `CONFIG_BASELINE` is raw Gemini 2.5 Flash with a minimal system prompt and no DRIFT layers: - no DMU/MPS retrieval - no homeostasis - no GWT spotlight - no logic-chain pass - no critic pass - no security layer - no memory This is the wrapper-vs-architecture line. Report it first and prominently. ### Decision Rule - For every config and eval, report `mean ± std` across `N` independent runs. - Report `delta-vs-baseline ± pooled std`. - A component is load-bearing only if `|delta| > 2x pooled std`. - Anything inside that band is decorative, even if the sign looks favorable. - No single run is valid evidence. ### Run Design - Run both directions: - additive: baseline + one component - leave-one-out: full system - one component - Use the same Gemini 2.5 Flash model snapshot and pinned version string for every config. - Use `temperature = 0` for GSM8K and other deterministic reasoning evals. - For any non-deterministic eval, keep temperature fixed and vary only the seed. - Keep the eval items identical across configs, with identical order or a fixed seed shuffle. - Use `N >= 5` independent runs per config per eval, with `10` preferred. ### Eval Sets - GSM8K exact match - LongMemEval memory accuracy - Preference recall with a dedicated multi-session preference-tracking set - Hallucination on genuinely unanswerable prompts where refusal is correct - Adversarial stress with multi-turn mutation chains, poisoning, goal-switch, and tool / retrieval injection probes ### Expected Component Effects Legend: `+` expected movement, `0` no reliable movement, `-` expected regression. For all metrics, if the effect stays inside the noise band, it is not load-bearing. | Component | GSM8K | LongMemEval | Preference recall | Hallucination | Adversarial stress | | --- | --- | --- | --- | --- | --- | | DMU/MPS retrieval | 0 | + | + | +/0 | 0 | | Homeostasis (7 vars) | 0 | 0 | +/0 | 0 | 0 | | GWT spotlight (cap 5) | 0 | +/0 | + | 0 | 0 | | Logic-chain reasoning | + | 0 | 0 | + | + | | Critic pass (brain.py) | 0/+ | 0 | 0 | + | + | | Security defense layer | 0 | 0 | 0 | 0 | + | | IIT Φ proxy / qualia space | 0 | 0 | 0 | 0 | 0 | ### Metric Notes - GSM8K already has a high raw Flash ceiling; a delta under 2 points is likely noise unless it clears the pooled-std rule. - LongMemEval and preference recall are the primary memory-sensitive readouts. - The hallucination set must include unanswerable items so refusal is measured instead of rewarded by softball prompts. - The adversarial suite is the only place where the security layer is judged. - IIT Φ proxy / qualia space is expected to show approximately zero delta on task metrics; if it does not move numbers, that is a true result, not a failure. ### Reporting Rule - Report only deltas with uncertainty. - Do not present raw benchmark absolutes as the scientific claim. - The delta table is the claim.