Document dense-max-100m release

README.md

@@ -96,18 +96,40 @@ The default configuration includes a ready-to-use text column. dense-100m avoids
 
     dense_pretrain = dense.map(to_text)
 
-## What models practice
+## Coverage map
 
-| Capability | Examples |
+### 16 primitive domains
+
+| | | | |
+|:--|:--|:--|:--|
+| Retention & lifecycle | Capacity & eviction | Addressing & aliasing | Editing & transactions |
+| Selectivity & access | Interference & concurrency | Automata & control | Algebra & reversibility |
+| Data structures | Graphs & spatial state | Temporal streaming | Online adaptation |
+| Distributed replication | Agents & messaging | Integrity & error control | Runtime state |
+
+### 8 composition domains
+
+| | | | |
+|:--|:--|:--|:--|
+| Transactional + temporal | Adaptive + control | Concurrent + distributed | Graph + structures |
+| Robust streaming | Multi-agent state | Algebraic automata | Memory pressure |
+
+default implements 36 explicit behavioral families. dense-100m provides a compact shared register VM. dense-max-100m expands the typed VM to 40 opcodes, 256 primitive families, 128 compositions, 32 structural renderers, and 64 semantic-preserving transformations.
+
+## Expected model capabilities
+
+| Capability | Training pressure |
 |:--|:--|
-| Retention and capacity | Delayed recall, bounded buffers, long-lived registers |
-| Addressing and editing | Multi-query recall, collisions, overwrite, erase/reuse |
-| Selection and interference | Gated writes, scoped memory, proactive and retroactive interference |
-| State dynamics | Automata, counters, permutations, monoids, finite groups |
-| Online adaptation | One-shot binding, remapping, rule switching |
-| Long execution stability | Active recurrence, checkpoints, read and generation pollution |
+| Working-memory utilization | Retain independent values across long active traces |
+| Precise state mutation | Write, overwrite, delete, patch, rollback, and restore without collateral damage |
+| Associative addressability | Resolve exact, hierarchical, indirect, and collision-prone keys |
+| Algorithmic recurrence | Execute counters, automata, permutations, graphs, and bounded data structures |
+| Online adaptation | Bind and remap rules inside context without updating model weights |
+| Interference control | Protect persistent state from noise, churn, concurrent writes, and repeated reads |
+| Transactional and distributed reasoning | Track snapshots, messages, replicas, clocks, and conflict resolution |
+| Streaming stability | Preserve behavior across checkpoints, chunks, retries, and long horizons |
 
-default implements 36 explicit behavioral families. dense-100m provides a compact shared register VM. dense-max-100m expands the typed VM to 40 opcodes, 384 compiled mechanisms, 32 structural renderers, and 64 semantic-preserving transformations.
+StateBench trains state handling, not factual knowledge. It does not by itself expand a model's physical context window, add world knowledge, or guarantee general reasoning improvements outside state-intensive tasks.
 
 ## Data contract
 
@@ -121,11 +143,12 @@ default implements 36 explicit behavioral families. dense-100m provides a compac
 | family / macro / level | scalar | four-item lists | Curriculum filtering |
 | density_units_per_1k_chars | no | yes | Semantic density control |
 
-## dense-100m at a glance
+## Dense configurations at a glance
 
 | Episodes | Physical rows | Density | Token estimate | Generator |
 |--:|--:|--:|--:|:--|
-| 100,000,000 | 25,000,000 | 162.01 units / 1k chars | 107–120B tokens | C++20 + Arrow |
+| 100,000,000 | 25,000,000 | 162.01 units / 1k chars | 107–120B tokens | dense-100m |
+| 100,000,000 | 25,000,000 | 162.88 units / 1k chars | ~107–120B tokens | dense-max-100m |
 
 The token range was measured on 1,000 real packed rows with GPT-2 and Qwen3 tokenizers. Actual training cost depends on chat templates, truncation, packing, and the SDFT generation policy.