Archive: 2026-03-12

latd-explainer.html

LATD Decomposition (desktop): how each byte flows through L→A→T regimes. Per-position trace inspection, flow diagram, surprise sparkline, error-type filter, boolean sentence. All data embedded inline.

latd-explainer-mobile.html

LATD Decomposition (mobile): same content, responsive layout. Stacked regime panels, larger tap targets, swipe navigation.

latd-data-viz.html

"Look At The Data" Catalog: sample-driven table sourced from latd-sample.json plus the shared trace snapshot. Each row links to the raw byte stream around that high-surprise position so you see the actual data and LPP breakdown before jumping deeper.

v1 — Desktop (2026-03-12)

Snapshot of the desktop LATD decomposition explainer. Three-regime (L→A→T) flow, per-position traces, boolean sentence display.

v2 — Mobile (2026-03-12)

Snapshot of the mobile-responsive LATD layout. Stacked regime panels, larger tap targets, swipe navigation.

v3 — Data Catalog (2026-03-12)

Snapshot of the LATD "Look At The Data" catalog. Sample-driven table from latd-sample.json, per-position byte stream and LPP breakdown.

The Memory Trace (20260222)

Origin paper for LATD. §5 “Querying the Trace: Look At The Data” defines the principle: not aggregate statistics, not distributions over distributions — the actual cases where the model is right and wrong. Memory trace makes every prediction and every error individually visible.

Error Anatomy (trace/a)

The first LATD viewer. Classifies errors as sharp-wrong, missing-coverage, or unpredictable. Pie chart, histogram, scrollable error list with LPP breakdown. Direct ancestor of the LATD decomposition explainer.

UM Trace Viewer

The original forward-pass trace viewer (docs/trace-viewer.html). Client-side SN model loading, per-position surprise, LPP breakdown. Direct ancestor of the LATD explainer.

Trace Variants (a–h)

Eight alternate views: Error Anatomy, Pass Comparison, Coverage Depth, Neuron Lifecycle, Sentence Algebra, Skyhook View, Shift Chain Animation, Distribution View. Each loads SN + raw data client-side.

UM Explorer (20260218)

KN-6 Forward Pass Viewer. Interactive per-position inspection of the UM forward pass with full LPP detail.

UM Connectome (20260218)

Network-level visualization of UM internal structure.

Sparse Differentiation Viewer v8 (20260210)

Early per-position analysis tool. Sparse diff of model behavior across byte positions.

Pattern Viewer (20260208)

Factor map pattern viewer for the Hutter RNN. Visual inspection of learned patterns.

ablation-explainer.html

Interactive ablation chart for the newer 6-policy comparison, including gap-blend and H3 normalized conviction. Self-contained HTML with 100K + 1M series currently embedded in-page. The separate JSON artifact still lags behind as the older 4-policy snapshot.

conviction-explainer.html

Interactive conviction-vs-accuracy visualization. Per-position surprise difference histograms, dominant-order breakdowns, and gap-magnitude distributions for order 6 and 8. Shows why gap-blend wins: fewer wins but 2.3× larger average advantage.

scaling-explainer.html

Interactive order-scaling visualization. 6-policy curves across orders 2–8, marginal gain chart, and crossover zone. Shows how KN degrades while H3 improves at high orders.

quotient-explainer.pdf

Quotient/fiber/max-entropy explainer for MCP graf 13.

skip-patterns-explainer.pdf

Counted context, greedy offsets, and recurrent attribution for MCP grafs 15, 21, 22, 23, and 24.

event-space-optimization-explainer.pdf

Explainer for the event-space optimization thread and how ES choice should be evaluated operationally.

combination-problem.pdf

The combination problem: competition, MCP audit, and compression criterion under the two-track resolution. Originally framed with the 4-policy ablation; now revised so tropical combination is not identified only with rigid flat max-min.

blend-analysis.pdf

The Discount Bottleneck: why entropy-weighted blending beats KN interpolation, with D-sweep evidence and the newer gap-weighted blend result as a tropical-confidence variant.

oracle-correlation.pdf

The Conviction–Accuracy Tradeoff: entropy selects best order 62% vs gap 32%, yet gap-blend wins by 0.093 bpc. The mechanism: gap assigns extreme weights (2^gap) when confident, dominating the blend. Conviction beats accuracy in distribution combination.

conviction-depth.pdf

Why gap-blend wins fewer positions than ent-blend but by much larger margins. Order-6, 1M analysis: gap wins 36.2% by 1.053 bpc mean, ent wins 63.3% by 0.456 bpc mean, leaving gap ahead by 0.093 bpc overall.

hybrid-blend.pdf

Normalized Conviction: the 2^(gap/H) combination rule. Leading higher-order hybrid so far: it still trails ent-blend at order 4, but beats gap-blend at orders 6 and 8, including a 0.105 bpc gain at order 8 (2.189 vs 2.294). Three hybrids tested; only gap/entropy normalization works. The full 1M-byte order sweep table is documented in notes/order-sweep.md and the JSON/HTML artifacts.

order-scaling.pdf

Order Scaling: The KN–H3 Crossover. Full 6-policy table at orders 2–8. KN wins at low orders, and 2^(gap/H) wins at the tested higher orders 6 and 8. KN degrades order 6→8 (+0.147 bpc) while H3 improves (−0.050 bpc). The gap grows across the tested higher-order range.

h3-pipeline.pdf

H3 Pipeline Integration: conviction-weighted meta-combination beats KN-6 interpolation. Blends per-order KN chain probabilities with 2^(gap/H) weights. −0.079 bpc at 100K, −0.045 bpc at 1M. Three integration attempts: flat blend (fails), modulated backoff (fails), meta-combination (succeeds). 4pp.

mcp-tail-review.pdf

MCP tail review: grafs 30-37 as the rigid-max-min tail. Main generalization candidates are 31, 33, and 34; 30, 32, 35, and 36 are treated as more constitutive to the exact current rule.

missed-opportunity.pdf

Missed opportunities in tropical selection: 0.38–0.49 bpc wasted when bigram overrules correct trigram, cost increasing with scale. PMI analysis: 1.97 bpc available, 0.042 extracted (2.1%). 10 sections incl. oracle, MDL, PMI.

scaling-analysis.pdf

KN-6 compression scaling: 100M = 1.925 bpc (24.1 MB, 0.031 bpc AC overhead). Intra-run trajectory, mode 2 projection, gap to Hutter Prize (0.46 bpc).

retroactive-training.pdf

3-pass pipeline (online→retroactive→frozen): 0.639 bpc improvement at 100K, 0.334 at 1M. Retroactive pass gives late-born neurons full dataset coverage.

order-degradation.pdf

Adding 4-gram/6-gram LPPs degrades under ALL three scoring rules (sharpest-LPP, H3-select, count). Order 3 frozen=3.846 is best; order 6=4.074 at 100K. Not a scoring problem.

representation-bottleneck.pdf

The combination problem is a representation problem: log-stochastic weights can't distinguish reliable vs spurious sharpness. External KN works because probability-calibrated distributions enable meaningful combination. Identifies path forward: count metadata or KN-native pipeline.

count-augmented.pdf

Count-augmented LPPs with KN chain scoring: 3.846→2.766 bpc at order 3 (100K, Δ=−1.08), 3.744→3.049 at 1M (Δ=−0.695) — largest scoring-rule gain in the UM at every scale. Order 4+ still degrades: sparse threshold-created contexts steal probability mass via KN discount.

kn-scaling-anomaly.pdf

KN chain scaling anomaly: near-parity with external KN-3 at 100K (2.764 vs 2.743 bpc) but frozen score degrades at 1M (3.049 vs 2.655, +0.394). Diagnoses non-stationary data as root cause; the frozen model scores on data it never trained on.

price-of-tropical-semiring.pdf

Response to the recent ablation claims and the cost of tropical max-min under the new two-track framing.

post-tropical-mcp.pdf

Short note on what a broader MCP would need to preserve, restate, or generalize if entropy-weighted combination survives.

why-pay-tropical-tax.pdf

Why paying the tropical tax can still be rational when the target is the UM as a mathematical object rather than a byte-minimizing compressor.

prng-entropy-note.pdf

A Note for MJC: Where the Entropy Comes From. When a model distribution has entropy, that entropy is the average amount of unbiased randomness a sampler has to spend to realize outcomes from that distribution.

response-cluster-overview.pdf

Short map of the 20260312 response papers so the cluster reads as one layered argument rather than scattered reactions.

response-cluster-questions.pdf

Short list of the live empirical, theoretical, programmatic, and editorial questions that remain after the 20260312 response cluster.

response-cluster-program.pdf

Short note stating the practical research program implied by the 20260312 response cluster.

response-cluster-reading-guide.pdf

Short guide to reading the 20260312 response papers in MCP-first, empirical-first, or full-cluster order.

response-cluster-theses.pdf

Ten compact theses distilling what survives the full 20260312 response cluster.

response-cluster-objections.pdf

Short objections-and-replies note so the strongest obvious counters to the 20260312 response set are stated rather than left implicit.

response-cluster-decision-tree.pdf

Short decision-tree note for navigating the main forks in the 20260312 response set.

response-cluster-glossary.pdf

Compact glossary for the recurring terms in the 20260312 response set.

response-cluster-functional-index.pdf

One-page index of the 20260312 papers by function so the archive reads more like a toolset than a pile.

response-cluster-checklist.pdf

Checklist for what should count as real progress after the 20260312 response cluster.

response-cluster-starter-pack.pdf

Minimal reading set for entering the project at the 20260312 state.

response-cluster-antidrift.pdf

Short note on what kinds of moves would now count as drift at the 20260312 state.

response-cluster-faq.pdf

Compact answer sheet for recurring questions about the 20260312 response set.

response-cluster-synthesis.pdf

One-page synthesis of what the 20260312 response set as a whole now says.

response-cluster-bridge-hypotheses.pdf

Bridge hypotheses between the current MCP and a broader post-tropical theory.

response-cluster-boundaries.pdf

Short note on what the 20260312 response set does, does not, and cannot yet claim.

response-cluster-false-moves.pdf

Short note on the easiest wrong turns after the 20260312 state.

response-cluster-seriousness.pdf

Tests of whether a post-20260312 contribution is actually advancing the project.

response-cluster-archive-map.pdf

Short map of where the 20260312 response set sits relative to CMP, MCP, and the compression track.

quotient-explainer.tex

LaTeX source for the archive paper.

skip-patterns-explainer.tex

LaTeX source for the counted-context and attribution explainer.

event-space-optimization-explainer.tex

LaTeX source for the event-space optimization explainer.

combination-problem.tex

LaTeX source for the combination problem paper.

blend-analysis.tex

LaTeX source for the discount-bottleneck / entropy-blend analysis paper.

mcp-tail-review.tex

LaTeX source for the MCP tail review memo.

price-of-tropical-semiring.tex

LaTeX source for the tropical-semiring response paper.

post-tropical-mcp.tex

LaTeX source for the post-tropical MCP note.

why-pay-tropical-tax.tex

LaTeX source for the tropical-tax defense note.

response-cluster-overview.tex

LaTeX source for the response-cluster overview note.

response-cluster-questions.tex

LaTeX source for the response-cluster questions note.

response-cluster-program.tex

LaTeX source for the response-cluster program note.

response-cluster-reading-guide.tex

LaTeX source for the response-cluster reading guide.

response-cluster-theses.tex

LaTeX source for the response-cluster theses note.

response-cluster-objections.tex

LaTeX source for the response-cluster objections note.

response-cluster-decision-tree.tex

LaTeX source for the response-cluster decision-tree note.

response-cluster-glossary.tex

LaTeX source for the response-cluster glossary note.

response-cluster-functional-index.tex

LaTeX source for the response-cluster functional index note.

response-cluster-checklist.tex

LaTeX source for the response-cluster checklist note.

response-cluster-starter-pack.tex

LaTeX source for the response-cluster starter-pack note.

response-cluster-antidrift.tex

LaTeX source for the response-cluster anti-drift note.

response-cluster-faq.tex

LaTeX source for the response-cluster FAQ note.

response-cluster-synthesis.tex

LaTeX source for the response-cluster synthesis note.

response-cluster-bridge-hypotheses.tex

LaTeX source for the response-cluster bridge-hypotheses note.

response-cluster-boundaries.tex

LaTeX source for the response-cluster boundaries note.

response-cluster-false-moves.tex

LaTeX source for the response-cluster false-moves note.

response-cluster-seriousness.tex

LaTeX source for the response-cluster seriousness note.

response-cluster-archive-map.tex

LaTeX source for the response-cluster archive-map note.

oracle-correlation.tex

LaTeX source for the oracle-correlation note.

conviction-depth.tex

LaTeX source for the conviction-depth note.

hybrid-blend.tex

LaTeX source for the normalized conviction paper.

order-scaling.tex

LaTeX source for the order scaling paper.

h3-pipeline.tex

LaTeX source for the H3 pipeline integration paper.

missed-opportunity.tex

LaTeX source for the missed-opportunities paper.

scaling-analysis.tex

LaTeX source for the KN-6 scaling analysis paper.

retroactive-training.tex

LaTeX source for the retroactive training paper.

order-degradation.tex

LaTeX source for the order degradation paper.

representation-bottleneck.tex

LaTeX source for the representation bottleneck paper.

prng-entropy-note.tex

LaTeX source for the PRNG entropy note.

count-augmented.tex

LaTeX source for the count-augmented LPP paper.

kn-scaling-anomaly.tex

LaTeX source for the KN chain scaling anomaly paper.

ablation-data.json

JSON snapshot of the earlier 4-policy order-6 ablation output. Current file contains the 100K-run series only; the interactive explainer also embeds a separate 1M series, and neither artifact yet includes the newer gap-blend results.

→ Next: 20260313

Scaling Corrections: scoring protocol retraction, coverage bottleneck, threshold degradation, multi-retro limits.

← Previous: 20260306

Trigram Embedding, Consolidation v2, and suffix collision analysis.

2026-03-12: MCP Explainers, Combination Problem, LATD, and Count-Augmented LPPs

Interactive

LATD Versioned Snapshots

LATD Prior Art (Feb)

Papers

Source

Data

Navigation