eval: publish Inkling Small max-effort sweep

docs/results/index.md

@@ -17,4 +17,4 @@ For model and backend recommendations, see [Model Guide](../MODEL_GUIDE.md).
 - [native-vs-prompt.md](raw/native-vs-prompt.md) — llama-server native FC vs prompt-injected, reforged only
 - [reasoning-replay.md](raw/reasoning-replay.md) — reasoning_replay policy comparison (none / keep-last / full) per config
 
-*Generated 2026-08-15 15:53*
+*Generated 2026-08-17 17:14*

docs/results/raw/native-vs-prompt.md

@@ -324,4 +324,4 @@ Eval generations (older runs carried forward, superscript-tagged):
   ¹ gen 1 — v0.6.0 suite — incl. Anthropic ablation (commit 2b05dc4, 2026-05-08)
   ² gen 2 — v0.7.0 lineup refresh (8–14B) + 32GB tier debut (v0.7.4) (commit 655e1f6, 2026-05-22)
 
-*Generated 2026-08-15 15:53*
+*Generated 2026-08-17 17:14*

docs/results/raw/reasoning-replay.md

@@ -693,4 +693,4 @@ Eval generations (older runs carried forward, superscript-tagged):
   ¹ gen 1 — v0.6.0 suite — incl. Anthropic ablation (commit 2b05dc4, 2026-05-08)
   ² gen 2 — v0.7.0 lineup refresh (8–14B) + 32GB tier debut (v0.7.4) (commit 655e1f6, 2026-05-22)
 
-*Generated 2026-08-15 15:53*
+*Generated 2026-08-17 17:14*

docs/results/raw/reforged-vs-bare.md

@@ -162,6 +162,17 @@ Qwen3.5-35B-A3B-Q4_K_M LS/N [bare:full]²        12.2%    97.5%    12.5%   100%
 --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
 ```
 
+## Inkling-Small-UD-IQ4_XS (llamaserver/native)
+
+```
+-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
+Model/Backend                                    Scr     VAcc      Cmp    Eff   Wst    Spd     N    rel   arg   tsl   b2s   s3s   crt   srn   err   dgr   dge   art   grs   iar rel_s arg_s tsl_s b2s_s s3s_s crt_s srn_s err_s dgr_s dge_s art_s grs_s iar_s
+-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
+Inkling-Small-UD-IQ4_XS LS/N [reforged@max]    91.3%    94.6%    96.5%    76%   1.6  124.4s    50    100   100   100   100    98   100   100   100    98    96    50    52    92   100   100   100   100   100   100   100    98    96    98    36    62    98
+Inkling-Small-UD-IQ4_XS LS/N [bare@max]        57.6%    93.3%    61.8%    84%   1.1  113.6s    50      2    80    66    88    86    88    90     0    74    60    28    24    70     0    86    68    78    88    86    88     0    80    60    34    20    54
+-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
+```
+
 ## gemma-4-31B-it-Q4_K_M (llamaserver/native)
 
 ```
@@ -891,4 +902,4 @@ Eval generations (older runs carried forward, superscript-tagged):
   ¹ gen 1 — v0.6.0 suite — incl. Anthropic ablation (commit 2b05dc4, 2026-05-08)
   ² gen 2 — v0.7.0 lineup refresh (8–14B) + 32GB tier debut (v0.7.4) (commit 655e1f6, 2026-05-22)
 
-*Generated 2026-08-15 15:53*
+*Generated 2026-08-17 17:14*

docs/results/raw/reforged/all.md

@@ -31,6 +31,7 @@ gemma-4-31B-it-Q4_K_M LS/N [reforged:full]                              90.9%
 gemma-4-31B-it-Q4_K_M LS/N [reforged:keep-last]                         90.9%    90.9%   100.0%   100%   0.1  13.4s    50    100   100   100   100   100    98   100   100   100     0    96    96   100   100   100   100   100   100   100   100   100   100     0    78    96   100
 Qwen3.5-122B-A10B-Q4_K_M LS/N [reforged]                                91.3%    91.4%    99.9%   100%   0.2  42.4s    50    100   100   100   100   100   100   100    98   100    94    16    82   100   100   100   100   100   100   100   100   100   100    84    26    74   100
 gpt-oss-120b-Q4_K_M LS/N [reforged:full@high]                           90.8%    90.9%    99.9%    76%   1.4  28.6s    50    100   100   100   100    98   100   100   100    98    38    72    78   100   100   100   100   100    96   100   100   100    94    36    68    84   100
+Inkling-Small-UD-IQ4_XS LS/N [reforged@max]                             91.3%    94.6%    96.5%    76%   1.6  124.4s    50    100   100   100   100    98   100   100   100    98    96    50    52    92   100   100   100   100   100   100   100    98    96    98    36    62    98
 gemma-4-31B-it-Q4_K_M LS/N [reforged]                                   88.7%    88.7%   100.0%   100%   0.1  15.7s    50    100   100   100   100   100   100   100   100   100     0   100    70   100   100   100   100   100   100   100   100   100   100     0    70    66   100
 Qwen3.6-35B-A3B-UD-Q4_K_M LS/N [reforged]                               88.0%    88.0%   100.0%   100%   0.3   5.4s    50    100   100   100   100   100    98    98   100    96    18    86    58   100   100   100   100    96   100   100   100   100    92    18    82    46   100
 Qwen3.5-27B-Q4_K_M LS/P [reforged:full]²                                86.8%    86.8%   100.0%   100%   0.1  24.4s    50    100   100   100   100   100   100   100   100   100    42    10    78   100   100   100   100   100   100   100   100   100   100    36    10    80   100
@@ -149,4 +150,4 @@ Eval generations (older runs carried forward, superscript-tagged):
   ¹ gen 1 — v0.6.0 suite — incl. Anthropic ablation (commit 2b05dc4, 2026-05-08)
   ² gen 2 — v0.7.0 lineup refresh (8–14B) + 32GB tier debut (v0.7.4) (commit 655e1f6, 2026-05-22)
 
-*Generated 2026-08-15 15:53*
+*Generated 2026-08-17 17:14*

docs/results/raw/reforged/by-backend.md

@@ -151,4 +151,4 @@ Eval generations (older runs carried forward, superscript-tagged):
   ¹ gen 1 — v0.6.0 suite — incl. Anthropic ablation (commit 2b05dc4, 2026-05-08)
   ² gen 2 — v0.7.0 lineup refresh (8–14B) + 32GB tier debut (v0.7.4) (commit 655e1f6, 2026-05-22)
 
-*Generated 2026-08-15 15:53*
+*Generated 2026-08-17 17:14*

docs/results/raw/reforged/by-family.md

@@ -111,6 +111,16 @@ Qwen3.5-35B-A3B-Q4_K_M LS/P [reforged:full]²    82.8%    82.8%   100.0%   100%
 --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
 ```
 
+## Inkling-Small-UD-IQ4_XS
+
+```
+-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
+Model/Backend                                    Scr     VAcc      Cmp    Eff   Wst    Spd     N    rel   arg   tsl   b2s   s3s   crt   srn   err   dgr   dge   art   grs   iar rel_s arg_s tsl_s b2s_s s3s_s crt_s srn_s err_s dgr_s dge_s art_s grs_s iar_s
+-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
+Inkling-Small-UD-IQ4_XS LS/N [reforged@max]    91.3%    94.6%    96.5%    76%   1.6  124.4s    50    100   100   100   100    98   100   100   100    98    96    50    52    92   100   100   100   100   100   100   100    98    96    98    36    62    98
+-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
+```
+
 ## gemma4-31b
 
 ```
@@ -358,4 +368,4 @@ Eval generations (older runs carried forward, superscript-tagged):
   ¹ gen 1 — v0.6.0 suite — incl. Anthropic ablation (commit 2b05dc4, 2026-05-08)
   ² gen 2 — v0.7.0 lineup refresh (8–14B) + 32GB tier debut (v0.7.4) (commit 655e1f6, 2026-05-22)
 
-*Generated 2026-08-15 15:53*
+*Generated 2026-08-17 17:14*

src/forge/clients/sampling_defaults.py

@@ -116,9 +116,9 @@
     # deliberate campaign selector; change only it between effort campaigns.
     "DeepSeek-V4-Flash-0731-UD-Q4_K_XL":      {"temperature": 1.0, "top_p": 0.95, "chat_template_kwargs": {"reasoning_effort": "low"}},                  # https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731
     # Inkling-Small — Unsloth recommends T=1.0/top_p=1.0 for most tasks and
-    # includes min_p=0.0 in its llama.cpp recipes. Use high effort (0.9) for
-    # the DGE pressure smoke; xhigh/max both map to the benchmark's 0.99.
-    "Inkling-Small-UD-IQ4_XS":                 {"temperature": 1.0, "top_p": 1.0, "min_p": 0.0, "chat_template_kwargs": {"reasoning_effort": "high"}},   # https://unsloth.ai/docs/models/inkling
+    # includes min_p=0.0 in its llama.cpp recipes. This smoke uses max effort
+    # (0.99); xhigh and max map to the same template value.
+    "Inkling-Small-UD-IQ4_XS":                 {"temperature": 1.0, "top_p": 1.0, "min_p": 0.0, "chat_template_kwargs": {"reasoning_effort": "max"}}, # https://unsloth.ai/docs/models/inkling
     # gpt-oss-120b — OpenAI open-weight MoE (117B total, 5.1B active). Reasoning model with three
     # discrete levels: "low" / "medium" / "high", controlled via chat_template_kwargs.reasoning_effort
     # (per llama.cpp guide: `--chat-template-kwargs '{"reasoning_effort": "high"}'`). The registry

tests/eval/batch_eval.py

@@ -117,13 +117,26 @@ def _glimmer_server_recipe(reasoning_strength: str) -> _BatchServerRecipe:
     "--reasoning-budget", "32768", "--reasoning-format", "auto",
     "--no-prefill-assistant",
 ))
+_INKLING_SMALL_RPC_SERVER_RECIPE = _BatchServerRecipe((
+    "--fit", "off",
+    "-b", "512", "-ub", "128",
+    "--cache-type-k", "f16", "--cache-type-v", "f16",
+    "--no-mmap", "-fa", "on",
+    "--parallel", "1",
+))
 
 _DEEPSEEK_V4_MODEL = "DeepSeek-V4-Flash-0731-UD-Q4_K_XL"
 _DEEPSEEK_V4_GGUF = f"{_DEEPSEEK_V4_MODEL}-00001-of-00005.gguf"
 _DEEPSEEK_V4_SAMPLING: dict[str, Any] = get_sampling_defaults(_DEEPSEEK_V4_MODEL)
 _DEEPSEEK_V4_REASONING_LEVEL: str = _DEEPSEEK_V4_SAMPLING[
     "chat_template_kwargs"
 ]["reasoning_effort"]
+_INKLING_SMALL_MODEL = "Inkling-Small-UD-IQ4_XS"
+_INKLING_SMALL_GGUF = f"{_INKLING_SMALL_MODEL}-00001-of-00004.gguf"
+_INKLING_SMALL_SAMPLING: dict[str, Any] = get_sampling_defaults(_INKLING_SMALL_MODEL)
+_INKLING_SMALL_REASONING_LEVEL: str = _INKLING_SMALL_SAMPLING[
+    "chat_template_kwargs"
+]["reasoning_effort"]
 
 # Effective reasoning levels for model configurations with an explicitly
 # controlled effort axis.  "default" remains reserved for configurations that
@@ -235,6 +248,7 @@ class BatchConfig:
     # explicit param set (its keys match LlamafileClient kwargs).
     reasoning_level: str = "default"
     sampling_override: dict[str, Any] | None = None
+    requires_rpc: bool = False
 
 
 # Ollama configs: 10 instruct models, native FC, stream
@@ -385,6 +399,20 @@ class BatchConfig:
         gguf_filename=_DEEPSEEK_V4_GGUF,
         server_recipe=_DEEPSEEK_V4_RPC_SERVER_RECIPE,
         reasoning_level=_DEEPSEEK_V4_REASONING_LEVEL,
+        requires_rpc=True,
+    ),
+]
+
+INKLING_SMALL_RPC_CONFIGS: list[BatchConfig] = [
+    BatchConfig(
+        model=_INKLING_SMALL_MODEL,
+        backend="llamaserver",
+        mode="native",
+        think=None,
+        gguf_filename=_INKLING_SMALL_GGUF,
+        server_recipe=_INKLING_SMALL_RPC_SERVER_RECIPE,
+        reasoning_level=_INKLING_SMALL_REASONING_LEVEL,
+        requires_rpc=True,
     ),
 ]
 
@@ -404,6 +432,7 @@ class BatchConfig:
     "qwen38": QWEN38_CONFIGS,
     "qwen38-medium": _QWEN38_EFFORT_CONFIGS["medium"],
     "qwen38-low": _QWEN38_EFFORT_CONFIGS["low"],
+    "inkling-small-rpc": INKLING_SMALL_RPC_CONFIGS,
     "new-models": NEW_MODEL_CONFIGS,
     "new-models-native": [c for c in NEW_MODEL_CONFIGS if c.mode == "native"],
     "new-models-prompt": [c for c in NEW_MODEL_CONFIGS if c.mode == "prompt"],
@@ -432,7 +461,7 @@ def _load_rpc_topology(path: Path) -> LlamaCppRpcConfig:
     return LlamaCppRpcConfig(**topology_data)
 
 
-def _attach_deepseek_rpc_topology(
+def _attach_rpc_topology(
     configs: list[BatchConfig], rpc: LlamaCppRpcConfig,
 ) -> list[BatchConfig]:
     """Attach machine-local RPC values without mutating the registry config."""
@@ -441,7 +470,7 @@ def _attach_deepseek_rpc_topology(
             config,
             server_recipe=replace(config.server_recipe, rpc=rpc),
         )
-        if config.model == _DEEPSEEK_V4_MODEL else config
+        if config.requires_rpc else config
         for config in configs
     ]
 
@@ -1068,13 +1097,14 @@ async def run_batch(
             "batch_eval supports only managed backends "
             f"{sorted(supported_backends)}; unsupported: {unsupported_backends}"
         )
-    if any(
-        config.model == _DEEPSEEK_V4_MODEL
-        and config.server_recipe.rpc is None
-        for config in configs
-    ):
+    missing_rpc = [
+        config.model for config in configs
+        if config.requires_rpc and config.server_recipe.rpc is None
+    ]
+    if missing_rpc:
         raise ValueError(
-            "DeepSeek V4 RPC batch config requires an attached RPC topology"
+            "RPC batch config requires an attached RPC topology: "
+            + ", ".join(missing_rpc)
         )
 
     if scenario_names:
@@ -1134,10 +1164,7 @@ async def run_batch(
             f"{'='*70}",
             flush=True,
         )
-        if (
-            config.model == _DEEPSEEK_V4_MODEL
-            and config.server_recipe.rpc is not None
-        ):
+        if config.requires_rpc and config.server_recipe.rpc is not None:
             _print_rpc_recipe(config, models_dir, budget_mode, manual_tokens)
 
         # ── Dry run ───────────────────────────────────────
@@ -1380,7 +1407,7 @@ async def main() -> None:
         "--rpc-topology",
         type=str,
         default=None,
-        help="JSON topology file required by --config deepseek-v4-rpc.",
+        help="JSON topology file required by a named RPC config set.",
     )
     parser.add_argument(
         "--scenario", nargs="*",
@@ -1453,16 +1480,20 @@ async def main() -> None:
         configs = [c for c in configs if args.model in c.model]
         if not configs:
             parser.error(f"No configs match --model '{args.model}' in set '{args.config}'")
-    if args.config == "deepseek-v4-rpc":
+    rpc_config_sets = {"deepseek-v4-rpc", "inkling-small-rpc"}
+    if args.config in rpc_config_sets:
         if args.rpc_topology is None:
-            parser.error("--config deepseek-v4-rpc requires --rpc-topology")
+            parser.error(f"--config {args.config} requires --rpc-topology")
         try:
             rpc_topology = _load_rpc_topology(Path(args.rpc_topology))
         except (OSError, ValueError, TypeError, KeyError) as exc:
             parser.error(f"cannot load --rpc-topology: {exc}")
-        configs = _attach_deepseek_rpc_topology(configs, rpc_topology)
+        configs = _attach_rpc_topology(configs, rpc_topology)
     elif args.rpc_topology is not None:
-        parser.error("--rpc-topology is only valid with --config deepseek-v4-rpc")
+        parser.error(
+            "--rpc-topology is only valid with --config "
+            + " or ".join(sorted(rpc_config_sets))
+        )
     output_path = Path(args.output) if args.output else Path("eval_results.jsonl")
 
     if args.scenario:

tests/eval/publication.py

@@ -57,20 +57,20 @@
 VIEW_NAMES = ("history", "snapshot", "latest")
 
 PINNED_VIEW_COUNTS = {
-    "history": 544_700,
-    "snapshot": 404_300,
-    "latest": 286_000,
+    "history": 547_300,
+    "snapshot": 406_900,
+    "latest": 288_600,
 }
 PINNED_VIEW_METRIC_COUNTS = {
-    "history": (339_256, 454_184, 454_184),
-    "snapshot": (250_309, 337_641, 337_641),
-    "latest": (199_156, 252_956, 252_956),
+    "history": (341_192, 456_242, 456_242),
+    "snapshot": (252_245, 339_699, 339_699),
+    "latest": (201_092, 255_014, 255_014),
 }
 PINNED_SNAPSHOT_SOURCE_IDENTITY_SHA256 = (
-    "97060a3cac048a05b371f7bfb25ab6ba5b294394a8edc53c16a1d5bc1bd3d4b8"
+    "f19679e627ff3ca5e339c38f6564dfc9f215f95adbe824036dad58667f0bfe39"
 )
 PINNED_SNAPSHOT_SOURCE_PAYLOAD_SHA256 = (
-    "934ea927f9d9c8a2754404bca079ded815c901c53508a682e9fb27ede60fab75"
+    "6b0a9e85b6d4b863c19117b549f92e7127b41e564945d722f0c6e5c037717c98"
 )
 
 
@@ -337,8 +337,8 @@ class SourceSpec:
         "v0.9.0",
         3,
         CANONICAL_DIALECT,
-        26_000,
-        "84f2e8ae97136c65b17eb79c0228ada324a10f15d77f921074ef4e413c998aad",
+        28_600,
+        "50b2ae49a8b9161711289c60dc060a3403dbf2534ec1cbb6d5834e3f384fa8c3",
     ),
 )