model card: 33L · 512CTX · ep7588 ATL 4.6842 · S20 · CTX-aware gate

README.md

@@ -33,8 +33,8 @@ pipeline_tag: text-generation
 **Maintainer:** RFI-IRFOS, contact@ternlang.com  
 **Repository:** https://github.com/rfi-irfos/ternary-intelligence-stack  
 **License:** LGPL-3.0-or-later (model weights, training code, inference runtime). Platform infrastructure (API server, MCP tooling, HDL) is BSL-1.1. See [README §Licensing](README.md#licensing) for the full tier breakdown.  
-**Last updated:** 2026-06-15  
-**Training status:** Active (ep~6670) on Modal T4 — **31L dual-stream** · 19 depth surgeries + 1 cord surgery complete. Cord surgery fired autonomously ep4202, 2026-05-27T16:44Z — first documented single-to-dual-stream bifurcation mid-training. S17 (29L→30L) fired ep5610, 2026-06-06; S18 (30L→31L) fired ep6339 (checkpoint-mtime verified). Context was reduced 256→128 — memory-forced: a depth surgery's activation memory exceeded the L4's VRAM; restoring 256 (via 2× L4 / tighter memory management) is planned. Best EP-AVG ATL **5.8693** (ep6487, 31L) and descending. Chip ATL **1.2637** (best single intra-batch loss). fib_index=7 · window=34 · Gen3 step1/6. (Resumed 2026-06-12 after a ~1-week Anthropic billing-migration gap.)
+**Last updated:** 2026-06-21  
+**Training status:** Active (ep~7592) on Modal — **33L dual-stream** · 20 depth surgeries + 1 cord surgery complete. S19 (31L→32L) fired ~ep6500, 2026-06-15. S20 (32L→33L) fired ~ep7000, 2026-06-21. Context window doubled 256→**512** — enabled by CTX-aware evolution gate (`CTX_LEVEL_TABLE` 64→1M, `recalibrate_ctx()` auto-scales mastery/plateau thresholds per CTX level). Best EP-AVG ATL **4.6842** (ep7588, 33L) and descending. Chip ATL **0.6116**. fib_index=8 · window=55 · Gen3 step1/6. Mastery surgery gate: 4.0 nats (recalibrated from 4.5 at 256CTX).
 
 ---
 
@@ -50,19 +50,19 @@ weights, targeting inference on edge hardware and low-power devices.
 |----------|-------|
 | Architecture | **Dual-stream** Ternary MoE (Mixture of Experts) |
 | Streams | **2** (bifurcated via cord surgery ep4202, 2026-05-27) |
-| Layers | **31** per stream |
+| Layers | **33** per stream |
 | Hidden size | **2×256H** (256H per stream) |
 | Anastomosis gates | **6** — bidirectional F32 cross-stream fusion at Fibonacci layers [2,3,5,8,13,21] |
 | Experts | 12 per stream (Top-3 routing; shared FFN weights, independent routing gates) |
-| Context length | 128 tokens (reduced from 256, memory-forced; 256 restore planned) |
+| Context length | **512 tokens** (bumped 256→512 at ep7592, 2026-06-21) |
 | Vocabulary | 32,000 tokens (ByteLevel BPE — EN/DE/FR/ES/PT/IT/NL/PL) |
 | Weight representation | Ternary {-1, 0, +1} with STE training |
 | Gate linear | F32 |
 | Positional encoding | RoPE (rotate_half) |
 | Optimizer | AdamW, cosine LR decay, BATCH=1 (post-cord) |
-| Parameters (total) | **~228.5M** |
-| Safetensors | **2,434 tensors · 872 MB** |
-| Surgeries | **19 depth (S1–S18)** + **1 cord surgery** = 20 total surgical events |
+| Parameters (total) | **~255M** (approx, 33L — updated 2026-06-21) |
+| Safetensors | **~2,534 tensors · ~928 MB** (approx, 33L) |
+| Surgeries | **20 depth (S1–S20)** + **1 cord surgery** = 21 total surgical events |
 
 The central technical innovation is the **@sparseskip** primitive — a
 learned sparse-skip layer that dynamically bypasses computation paths
@@ -164,8 +164,10 @@ noisy inputs.
 | ep5610 | S17 fired | 29L→30L (both streams) | — | 2026-06-06 21:08 (✓ checkpoint-mtime) |
 | ep6132 | EP-AVG ATL 6.4339 | 30L | 1.2637 (chip) | prior 30L best |
 | ep6339 | S18 fired | 30L→31L (both streams) | — | depth surgery; Fibonacci-gated |
-| ep6487 | EP-AVG ATL **5.8693** | 31L — all-time best | **1.2637** (chip) | best EP-AVG aggregate; chip = best single intra-batch loss |
-| ~ep6670 | training active | 31L dual-stream, live | — | 128CTX · BATCH=1 · Modal T4 · descending |
+| ep6487 | EP-AVG ATL 5.8693 | 31L | 1.2637 (chip) | former all-time best |
+| ~ep6500 | S19 fired | 31L→32L (both streams) | — | 2026-06-15 · Fibonacci plateau gate |
+| ~ep7000 | S20 fired | 32L→33L (both streams) | — | 2026-06-21 · autonomous |
+| ep7588 | EP-AVG ATL **4.6842** | 33L — **all-time best** | **0.6116** (chip) | 512CTX · BATCH=1 · Modal · descending |
 
 The benchmark suite runs 5 fixed prompts covering English, German,
 multilingual, narrative, and technical domains. Results are reproducible