Add performance evidence lab PoC

README.md

@@ -134,9 +134,10 @@ uv run blueprinting-workbench
 blueprinting --help
 ```
 
-It provides a shared configuration surface for single-point analysis, canonical IR derivation audit, and bounded
-TP/PP/DP strategy exploration. Analysis runs outside the UI event loop, and failed candidates remain visible as
-structured diagnostics.
+It provides a shared configuration surface for single-point analysis, canonical IR derivation audit, bounded
+TP/PP/DP strategy exploration, and a read-only performance-evidence lab. The evidence PoC catalogs the pinned
+Vidur Phi-2/A100 records and compares exact GEMM samples with the analytical roofline on identical workload facts.
+Analysis runs outside the UI event loop, and failed candidates remain visible as structured diagnostics.
 
 The existing Calculon Streamlit tools remain isolated as an optional legacy interface. Floating-point analysis is available in the primary NiceGUI workbench:
 

docs/design/performance/providers.en.md

@@ -172,6 +172,12 @@ The application boundary also accepts `cost_resolver` and `cost_context`, so evi
 
 Inference task queries are derived from canonical `PlanTask.workload`; GEMM dimensions and local attention-head dimensions are derived from model and TP facts. Tensor-parallel collectives and pipeline P2P use the same resolver. All tasks must be covered by an installed provider—normally an exact database followed by roofline—so an unknown task never becomes zero.
 
+## Performance-evidence UI PoC
+
+The NiceGUI workbench's Performance Evidence entry provides a read-only evidence catalog, operation coverage, GEMM characteristic curves, and a query/record inspector. Its first pinned dataset is the Vidur Phi-2/A100 profile, shipped in the base wheel together with its source revision, manifest, and license.
+
+The PoC compares only four GEMM semantics because their selectors can reconstruct `M/N/K`, FLOPs, and bytes without ambiguity. The measured series contains exact sample points and makes no interpolation claim; the analytical series is a compute/memory roofline over the same workload facts, not an event-level simulation. Attention and other operations currently appear only in the coverage catalog rather than receiving a fabricated generic analytical formula.
+
 ## Implemented boundary and next steps
 
 Implemented now:
@@ -183,6 +189,7 @@ Implemented now:
 - generic simulator/profiler table ingestion;
 - explicit Vidur and four-family AIConfigurator ingestion;
 - inference task and pipeline integration with regression tests.
+- a read-only evidence catalog and GEMM curve-comparison PoC over the pinned Vidur profile.
 
 Still missing:
 
@@ -193,5 +200,6 @@ Still missing:
 - contention, overlap, queueing, and plan-level discrete-event simulation;
 - energy/power metrics as normalized planner objectives;
 - observation ingestion and calibration revisions.
+- multi-database repository/registry, user import, and explicit evidence-promotion UI.
 
 These omissions are important: task latency resolution is an evidence layer, not a completed hardware or serving simulator.

docs/design/performance/providers.zh.md

@@ -172,6 +172,12 @@ Application 入口同样接受 `cost_resolver` 与 `cost_context`，因此导入
 
 Inference task query 直接从 canonical `PlanTask.workload` 推导；GEMM dimension 与 local attention-head dimension 来自 model 和 TP facts。Tensor-parallel collective 与 pipeline P2P 使用同一个 resolver。所有 task 都必须被已安装的 provider 覆盖——通常是 exact database 后接 roofline——unknown task 不会被静默变成零。
 
+## 性能证据界面 PoC
+
+NiceGUI workbench 的“性能证据”入口提供只读 evidence catalog、operation coverage、GEMM 特性曲线和 query/record inspector。首个 pinned dataset 是 Vidur Phi-2/A100 profile；它连同 source revision、manifest 与 license 一起进入 base wheel。
+
+PoC 只对四类 GEMM semantic 绘制 database-versus-roofline 对比，因为这些 selector 可以无歧义地重建 `M/N/K`、FLOPs 与 bytes。Measured series 只显示 exact sample point，不声明插值；analytical series 是相同 workload facts 下的 compute/memory roofline，不是事件级仿真。Attention 与其他 operation 当前只进入 coverage catalog，尚不伪造通用解析对比公式。
+
 ## 已实现边界与后续工作
 
 当前已经实现：
@@ -183,6 +189,7 @@ Inference task query 直接从 canonical `PlanTask.workload` 推导；GEMM dimen
 - 通用 simulator/profiler table ingestion；
 - 显式 Vidur 与四类 AIConfigurator ingestion；
 - inference task/pipeline 接入与回归测试。
+- pinned Vidur profile 的只读 evidence catalog 与 GEMM curve comparison PoC。
 
 仍未实现：
 
@@ -193,5 +200,6 @@ Inference task query 直接从 canonical `PlanTask.workload` 推导；GEMM dimen
 - contention、overlap、queueing 与 plan-level discrete-event simulation；
 - 作为 normalized planner objective 的 energy/power metric；
 - observation ingestion 与 calibration revision。
+- 多数据库 repository/registry、用户导入与显式 evidence promotion UI。
 
 这些边界非常重要：task latency resolution 是 evidence layer，不是已经完成的 hardware simulator 或 serving simulator。

pyproject.toml

@@ -123,6 +123,7 @@ packages = ["src/blueprinting", "src/calculon"]
 [tool.hatch.build.targets.wheel.force-include]
 "data/models" = "blueprinting/presets/models"
 "data/systems" = "blueprinting/presets/systems"
+"data/validation/vidur/phi2_a100_tp1" = "blueprinting/presets/evidence/vidur/phi2_a100_tp1"
 
 [tool.hatch.envs.default]
 features = ["dev", "full"]

scripts/check_wheel_contract.py

@@ -1,4 +1,4 @@
-"""Verify that the base wheel contains presets but excludes optional evidence."""
+"""Verify that the base wheel contains presets and only the pinned PoC evidence slice."""
 
 from __future__ import annotations
 
@@ -23,11 +23,22 @@ def main(argv: list[str]) -> int:
     for prefix in required_prefixes:
         if not any(name.startswith(prefix) and name.endswith(".json") for name in names):
             raise SystemExit(f"wheel is missing JSON presets under {prefix}")
-    forbidden_prefixes = (
-        "blueprinting/systems/",
-        "data/evidence/",
-        "blueprinting/presets/evidence/",
+    evidence_prefix = "blueprinting/presets/evidence/vidur/phi2_a100_tp1/"
+    required_evidence = {
+        f"{evidence_prefix}LICENSE.vidur",
+        f"{evidence_prefix}attention.csv",
+        f"{evidence_prefix}manifest.json",
+        f"{evidence_prefix}mlp.csv",
+    }
+    missing_evidence = required_evidence - set(names)
+    if missing_evidence:
+        raise SystemExit(f"wheel is missing pinned PoC evidence: {sorted(missing_evidence)!r}")
+    unexpected_evidence = tuple(
+        name for name in names if name.startswith("blueprinting/presets/evidence/") and name not in required_evidence
     )
+    if unexpected_evidence:
+        raise SystemExit(f"wheel contains unapproved evidence: {unexpected_evidence[:3]!r}")
+    forbidden_prefixes = ("blueprinting/systems/", "data/evidence/")
     leaked = tuple(name for name in names if name.startswith(forbidden_prefixes))
     if leaked:
         raise SystemExit(f"wheel contains optional evidence: {leaked[:3]!r}")

src/blueprinting/analysis/__init__.py

@@ -3,6 +3,8 @@
 from .cost import (
     AIConfiguratorPerformanceImporter,
     AIConfiguratorTable,
+    CostCurvePoint,
+    CostCurveReport,
     CostEstimate,
     CostNotAvailableError,
     CostProvider,
@@ -14,6 +16,8 @@
     EstimateMatch,
     EstimateMethod,
     EstimateUncertainty,
+    EvidenceCoverage,
+    EvidenceDatabaseSummary,
     EvidenceProvenance,
     LatencyUnit,
     PerformanceDatabase,
@@ -23,6 +27,9 @@
     SimulatorPerformanceImporter,
     TabularImportSpec,
     TabularPerformanceImporter,
+    build_gemm_comparison_curve,
+    comparable_gemm_semantics,
+    summarize_performance_database,
 )
 from .cost_model import (
     BlockEstimate,
@@ -54,6 +61,8 @@
     "AIConfiguratorPerformanceImporter",
     "AIConfiguratorTable",
     "CostEstimate",
+    "CostCurvePoint",
+    "CostCurveReport",
     "CostNotAvailableError",
     "CostProvider",
     "CostQuery",
@@ -65,6 +74,8 @@
     "EstimateMethod",
     "EstimateUncertainty",
     "EvidenceProvenance",
+    "EvidenceCoverage",
+    "EvidenceDatabaseSummary",
     "InferenceBaseline",
     "InferenceCostProvider",
     "InferenceEvidenceQuery",
@@ -82,6 +93,9 @@
     "SimulatorPerformanceImporter",
     "TabularImportSpec",
     "TabularPerformanceImporter",
+    "build_gemm_comparison_curve",
+    "comparable_gemm_semantics",
+    "summarize_performance_database",
     "cost_query_for_inference_task",
     "estimate_block",
     "estimate_inference_phase",

src/blueprinting/analysis/cost/__init__.py

@@ -2,6 +2,15 @@
 
 from .aiconfigurator import AIConfiguratorPerformanceImporter, AIConfiguratorTable
 from .database import EvidenceProvenance, PerformanceDatabase, PerformanceDatabaseProvider, PerformanceRecord
+from .explorer import (
+    CostCurvePoint,
+    CostCurveReport,
+    EvidenceCoverage,
+    EvidenceDatabaseSummary,
+    build_gemm_comparison_curve,
+    comparable_gemm_semantics,
+    summarize_performance_database,
+)
 from .importers import (
     LatencyUnit,
     SimulatorPerformanceImporter,
@@ -37,6 +46,8 @@
     "CostProvider",
     "CostQuery",
     "CostQueryContext",
+    "CostCurvePoint",
+    "CostCurveReport",
     "CostResolution",
     "CostResolver",
     "CostSubject",
@@ -45,6 +56,8 @@
     "EstimateMethod",
     "EstimateUncertainty",
     "EvidenceProvenance",
+    "EvidenceCoverage",
+    "EvidenceDatabaseSummary",
     "InvalidCostEvidenceError",
     "LatencyUnit",
     "PerformanceDatabase",
@@ -56,4 +69,7 @@
     "SupportStatus",
     "TabularImportSpec",
     "TabularPerformanceImporter",
+    "build_gemm_comparison_curve",
+    "comparable_gemm_semantics",
+    "summarize_performance_database",
 ]