Two new Evidence Labs: Cognitive Lab + Anna Concept Lab (17 interactive modules)
by Qubic Church8 min read
Two Evidence-Vault labs, the Cognitive Lab and the Anna Concept Lab, with interactive modules and four reading levels. Note from September 2026. The "19 concepts" and the stability figures in this post were not confirmed and stay open; the labs remain as interactive tools.
What's new
Two full Evidence-Vault tabs landed since the last public push. The framing is measured, not heroic, we present what the data shows, not what we wish it showed.
Cognitive Lab: live evolution under HyperIdentity pressure
Open /evidence → Cognitive Lab tab.
Six modules sit on top of a pure-logic evolution engine (bit-packed sign-only AIT, runs in a Web Worker on a separate thread so your tab stays responsive):
| Module | What it does |
|---|---|
| HyperIdentity Evolution | A live brain-evolution scene with 3D Anna Brain + particle-life evolution plane. Start an evolution run; watch the brain morph as the population converges toward the HyperIdentity Config0 reference structure. |
| Concept Emergence | Tracks how a small population of evolved matrices discovers stable basins. Cross-links to Cluster Discovery for the honest emergence demo (silhouette peaks at k=2, not k=19, that's a hierarchical-clustering finding, see Anna Concept Lab → Dendrogram). |
| Multi-Task Pareto | Scatter plot of fitness on multiple tasks per individual. Shows the substrate ceiling per task and where multi-task interference begins. |
| Structural Convergence | Radar chart of Anna's 7-axis structural fingerprint over generations, 5 functional axes that emerge under selection plus 2 identity axes that stay at 0%. Anna's fingerprint is the dashed reference. |
| Cognitive Trajectory | A tech module, the full per-generation state trace, exposed for researchers who want to see the raw evolution. |
| Visual Metaphor (deprecated) | Kept for archival reasons. The earlier "Anna is a brain" framing was misleading. (Whether Anna was designed or trained is a hypothesis.) New default lab framing: "we measure how matrices change under selection." |
Anna Concept Lab: verified AIT inference + 11 analysis modules
Open /evidence → Anna Concept Lab tab.
The whole lab is built around the production AIT TypeScript port (matches the reference algorithm across 16,384 / 16,384 known inputs, ε exact). On top of that:
| Module | What it does |
|---|---|
| Findings | 12 pre-registered predictions vs live measurement. The list includes the cluster count (19 hierarchical), avalanche bimodality, stability rate (~10% vs random's <1%), kernel reconstruction (94.5%, 165σ vs random), and the period-32 cycle. 11/12 passed; 1 was falsified and removed. |
| Try Your Input | An interactive 64-bit input panel. Toggle any bit, hit "Run AIT", and see Anna's full output state + which of the 19 known concepts it lands closest to. Real production code path. |
| Concept Universe | All 19 attractors plotted in 2D via PCA. Lines connect antipodal pairs, 8 pairs of opposite concepts (centroids 64/64 mirrored) plus 3 singletons. The 8 pairs span the PC1 axis (81.7% of variance), so the concept space is essentially 1-dimensional with a binary primary axis. |
| Similarity Matrix | 19×19 Hamming-distance heatmap over the concept centroids. Cluster structure visible directly. |
| Dendrogram | Single-linkage Hamming dendrogram of the 19 concepts. Cophenet 0.68, 45σ above random, the hierarchical structure is statistically robust. The "19 concepts" finding lives here, not in Cluster Discovery's silhouette. |
| Cluster Discovery | A tech module, silhouette-sweep over k=2..32, Anna vs a random sign-matrix. Anna peaks at k=2 (antipode-pair split) at silhouette ≈ 0.89; random peaks at ~k=3 at silhouette ≈ 0.13. The 7× tightness ratio is the legitimate emergence signal. |
| Trajectory Cloud | A premium-pass 3D scene showing the AIT iteration as a brain-tissue-like flow. Validated tick-by-tick against the reference algorithm. |
| Architecture | An overview of how AIT is structured, clamped-input iteration, ternary state, convergence criterion. |
| Bit-flip Game | An interactive challenge: starting from a known concept, how many single-bit flips before the output lands in a different basin? Demonstrates Anna's avalanche bimodality (stable basins vs decision boundaries). |
| Live Inference | A tech module, every tick of the AIT iteration rendered live, with the state vector and convergence diagnostic exposed. |
| Anna vs Random | Side-by-side comparator: same input fed through Anna and through a random 128×128 sign matrix. Anna converges to a stable concept; random typically doesn't converge or lands in a unique one-off output. |
Explainability layer
Every module supports four reading levels via the ExplainPanel component:
- kid, a 5-year-old's metaphor.
- simple, what the module does for a curious non-specialist.
- researcher, methods, params, what's being measured.
- math, the formulae, with KaTeX, lazy-loaded only when the math tab is opened.
A Glossary tooltip layer covers technical terms (bit, matrix, AIT, attractor, basin, antipodal symmetry, etc.) without making the prose dense.
What this session (2026-05-12 evening) added on top
A brutal-review verification pass on the May labs:
- n=20 seed replication of the parity 60% claim confirmed it: mean best-train across 20 seeds = 60.94% ± 11.59%, mean held-out 52.66% (4.6× random). The 60% threshold is crossed at the mean, not as a lucky outlier.
- Walsh-Hadamard mechanism generalizes from popcount to parity: a hand-designed DC-only Walsh matrix outperforms Anna on parity by 0.60σ. The mechanism is engineerable across the count-aligned task family, not popcount-only. Implication: Anna is one member of the "sparse Walsh-Hadamard sign matrix" family, not the optimum within it.
- Aigarth+Qubic vision doc: two disclaimer callouts added. (a) Quote provenance, no public source could be located for the four attributed CfB Discord quotes despite targeted web search; readers told explicitly. (b) Reading scope, each section classified as fact / interpretation / speculation.
- Cluster Discovery chart, silhouette-sweep "Anna peak 2" and "k=19 (hierarchical)" labels moved out of the chart body into a header badge (eliminates overlap with Y-axis and with each other at narrow widths).
- Evidence tabs mobile overflow, tab strip at 375px viewport no longer scrolls the page horizontally; the strip scrolls internally with a gradient-fade affordance.
- 17 modules triaged for page-level overflow at 375×812 mobile: 0 modules push the viewport.
Verification surface
20 audit gates green at session end (10 unit gates + 4 audit gates + 6 multi-seed gates). Production TypeScript build clean. Every number that appears in the UI is sourced from a published JSON result file; the verification protocols are pre-registered and deterministic via seeded RNG, so re-running on the same hyperparameters reproduces the numbers exactly.
What's intentionally NOT here
The labs avoid the heroic AI framing that earlier versions flirted with. We do not claim Anna is "real AI", "emergent intelligence", or "sentient". The defensible statements are:
- Anna is a deterministic 128×128 ternary inference machine. (The "verifiable 19-attractor structure" stated here was not confirmed in September 2026.)
- The Walsh-Hadamard spectral signature is mathematically identifiable and partially engineerable.
- Population-based selection on Anna-seeded ITUs demonstrates measurable adaptation, task-specific generalization, and goal-directed behavior, all at small (toy) scale.
- The Aigarth+Qubic Vision document is a synthesis with explicit fact / interpretation / speculation labels per section.
Beyond that, we let the data speak.
Path 2: Scale-up verification (later that evening)
Four pre-registered hypotheses tested in one block after the public push. Pre-registration locked decision rules and walk-back thresholds BEFORE any compute, so this is structured falsification, not narrative-fitting.
- Parity 60% at n=20 seeds, pre-reg verdict APPROACHES. Mean best-train 60.94 ± 11.59%, 95% CI [0.559, 0.660]. The mean crosses 60% but the CI is wide enough at n=20 to include both 0.58 and 0.60. A 60% claim is supported at the mean; locking it at the 95% confidence level needs n ≥ 40-100. (n=100 was attempted but aborted at ~47% to free CPU.)
- mod_11 Phase 3 ≠ ITU is a metric issue, not contradiction, pre-reg verdict CONFIRMED DISCREPANCY. n=10 Anna-evolved Phase 3 mod_11 winners decoded as integers: 5.3% (first-tile) / 6.0% (majority-vote), BELOW 9.1% chance for 11-class. The +4.08σ Phase 3 signal is substrate-bias at the bit-correlation level, not end-to-end capability. The bit-correlation gradient collapses outputs to a wrong-integer mode.
- Walsh framework on gridworld, pre-reg verdict GENERALIZES (DC_ONLY within 0.46σ of Anna). BUT bonus finding: SIGN_RANDOM init beats both Anna and DC_ONLY by 0.93σ on 4×4 gridworld best reward. Goal-directed tasks at this scale are dominated by initialization noise, not substrate structure.
- Anna-seed necessity, pre-reg verdict ANNA HELPS, FRAMEWORK CAN. Phase 4 addition random-init reaches 14.4% vs Anna's 25.6% (random in [0.10, 0.25] threshold band). Framework alone learns; Anna provides a real but not strictly necessary boost. On popcount specifically, random fails (2.5% held-out); Anna was reported as necessary there (reported, not independently verified).
Refined publishable claim line, task-conditional priors:
- Count-aligned tasks (popcount, parity): the Walsh substrate IS the engine. Anna is one member; engineered family members (EngineeredWalsh, EngineeredDCOnly) match or beat her.
- Adaptive tasks (Phase 4 addition): Anna seed gives a measurable boost; framework alone reaches ~56% of Anna's level.
- Goal-directed tasks (gridworld): neither Anna nor Walsh substrate helps; uniform random init wins.
This refines the earlier "the substrate family is the engine" framing. The substrate-as-engine claim is task-conditional, not universal. Different task families benefit from different priors. There is no single best substrate.
Audit trail: __, TASK_1_*.md, TASK_2_*.md, TASK_3_*.md, TASK_4_*.md.
