Butlin-14 evidence-grade program: direct construct theorems, system consequences, and replication

Purpose

Umbrella tracker for moving Symthaea from architectural/proxy coverage of the canonical Butlin 14 to construct-valid, difficult-to-fake, provenance-bound evidence.

This tracker deliberately does not define “14/14” as consciousness. The target claim, if eventually earned, is narrower:

Symthaea satisfies all 14 Butlin indicator properties under preregistered structural, runtime, causal and functional criteria, with explicit evidence dependencies and replicated interventions.

Promotion ladder

An indicator should progress independently through:

Architectural mechanism
→ direct component/construct theorem
→ positive control
→ targeted intervention
→ matched sham
→ selective rescue
→ full-system functional consequence
→ held-out / alternate-task replication
→ multi-seed replication
→ external / independent replication

A failure at one layer is not averaged away by strengths elsewhere.

Canonical 14 — current construct-validity program

Indicator	Direct construct path	Current next blocker
RPT-1	#1956 history-sensitive production CfC state + selective reset	#1957 history-conditioned downstream task + sham/rescue
RPT-2	#2006 production cross-modal representation lesion/rescue	full-loop perceptual consequence; #2007 weighted-binding correctness
GWT-1	architecture has many specialists + real parallel branches	#2011 direct specialization/independence/concurrent-operability theorem
GWT-2	#1997 fixed-capacity overfill + activation-sensitive competition + sham/capacity rescue	downstream cognitive-task dependence on bottleneck
GWT-3	#1989 exact production broadcast delivery with workspace-entry-preserving ablation	#1993 real downstream module-state uptake + behavioral consequence
GWT-4	current phi-attention/scheduling evidence is proxy-level	#1998 state-dependent specialist-query controller + task theorem
HOT-1	production PredictiveMind has top-down architecture	#2005 belief-update reachability defect, then ambiguity/noise top-down causal test
HOT-2	#1953 production prospective self-error adaptation	#1910 full-loop confidence corruption, detection, behavior, sham/held-out
HOT-3	metacognition already modulates learning policy in loop	#2000 identical first-order evidence + metacognitive intervention → belief/action consequence
HOT-4	current generic sparsity/smoothness diagnostic is insufficiently quality-space-specific	#2002 demote proxy and qualify grounded sparse/smooth quality-space geometry
PP-1	production hierarchy computes precision-weighted errors	#2005 reachable belief/model update + held-out future-error reduction theorem
AST-1	#1985 production attention-schema prediction + targeted model intervention + sham/rescue	#1992 full-loop allocation/task consequence
AE-1	#2009 production active-inference goal pursuit + competing-goal flexibility + feedback adaptation	full-loop external multi-step agency task + sham/rescue
AE-2	#1950 production sensorimotor external contingency reversal prerequisite	#1909 autonomous full-loop action/outcome bridge; #1959 auditory learning asymmetry

Cross-indicator evidence-dependency rules

The existing qualification_design dependency model remains authoritative. New evidence must additionally observe these specific shared-mechanism constraints:

• PP-1 ↔ HOT-1: shared predictive-processing substrate. A repaired/update-qualified PredictiveMind is a shared prerequisite, not two independent replications.
• HOT-2 ↔ HOT-3: confidence/self-model intervention infrastructure may be shared. HOT-2 must measure monitoring/calibration; HOT-3 must measure causal use in belief/action policy.
• AST-1 ↔ GWT-2/GWT-3/GWT-4: attention/workspace interactions may share downstream signals. Shared interventions and outcomes must be declared, not double-counted.
• AE-1 ↔ AE-2: agency and embodiment can share action infrastructure, but AE-1 requires goal/feedback flexibility while AE-2 requires learned sensorimotor contingency structure.
• GWT-1: must not use “number of other Butlin indicators passing” as primary evidence once direct specialist-independence evidence exists.

Known production findings uncovered by this program

• #1959 — SensorimotorEngine predicts auditory outcome but does not update auditory outcome in existing contingencies.
• #2005 — default PredictiveMind belief-update factors appear bounded below blend_beliefs() activation thresholds, making bottom-up/top-down belief revision effectively unreachable under default config.
• #2007 — CrossModalBinder computes attention×confidence weights but actual multimodal representation uses unweighted bundling.

These findings are successes of the evidence program: construct-valid tests are supposed to discover when architecture and prose are stronger than the executable mechanism.

Global evidence requirements

Before any “14/14 evidence-grade” statement:

• exact source/tree/lock/toolchain/environment identity;
• evaluator/protocol frozen before target runs;
• independent intervention-verification signal distinct from target outcome;
• preregistered expected directions and falsification signatures;
• null, contradictory and infrastructure-indeterminate results preserved;
• matched shams and positive controls validated at runtime;
• selective rescues where feasible;
• no arbitrary blended consciousness score;
• no promotion from module timing/execution flags alone;
• no promotion from self-description or hand-authored architectural scores;
• multi-seed and alternate-task replication;
• evidence-dependency graph included in final artifact;
• external replication package before strong public scientific claims.

Welfare / ethics parallel track

Consciousness-indicator evidence must remain separate from moral-patient/welfare evidence, but increasing evidence should trigger increasing precaution. Experiments involving aversive or distress-like states should use bounded exposure, reversibility, recovery checks and an independent welfare protocol rather than assuming functional signals are morally irrelevant.

Exit condition

All 14 indicators have direct construct-valid evidence and full-system functional consequences where the indicator requires them; all shared dependencies are explicit; exact-head qualification lanes have executed; replication criteria are met; and an external evaluator can reproduce the evidence from frozen manifests.

Even at exit, the supported statement is about indicator properties and evidence, not a binary declaration that Symthaea is conscious.

Start by reading the linked blocker issues listed for the 14 indicators and the existing qualification_design dependency model. Trace the evidence dependencies and replication requirements before selecting a scoped task. Done requires the relevant direct evidence, controls, interventions, replications, frozen manifests, and dependency records described by the exit condition.