Add numerical and logical-state conformance support

benchmarks/conformance/.gitignore

@@ -0,0 +1,7 @@
+.venv/
+__pycache__/
+.pytest_cache/
+*.egg-info/
+private/
+*.pt
+*.so

benchmarks/conformance/README.md

@@ -0,0 +1,51 @@
+# Numerical and state conformance support
+
+This is the portable support used in the
+[M1/M8 and eager/compiled investigation](https://github.com/Terrydaktal/d7-rdna4-report/blob/main/reports/d7-rdna4-2026-09-17/REPORT.md).
+It compares implementations against a declared reference; passing finite samples
+does not prove arbitrary-input equivalence or model task accuracy.
+
+```text
+conformance/
+├── src/qwen_r9700_lab/   # preserved module names and source identities
+├── tests/               # synthetic CPU regressions and negative controls
+├── configs/profiles/    # explicit serial GDN arithmetic contract
+├── SOURCE_PROVENANCE.json
+├── pyproject.toml
+└── uv.lock
+```
+
+Install and run the CPU suite without loading a model:
+
+```sh
+uv sync --project benchmarks/conformance --group dev
+uv run --project benchmarks/conformance pytest benchmarks/conformance/tests -q
+```
+
+The top-k comparator consumes aligned full-logit summaries and reports top-1,
+top-10 and top-20 set agreement, order agreement, overlap, retained-score equality,
+boundary ties and full-row digests separately. Empty domains, mismatched positions,
+changed manifests and incomplete evidence are errors. It does not decode tokens.
+
+The other modules provide logical-state comparisons, transition invariants,
+tentative-output checking, fault injection, source/binary binding, exclusive GPU
+leases and checkpointed worker supervision. Abstract proofs and CPU simulations
+remain distinct from native GPU qualification. Private raw rows and public
+aggregate receipts have different storage rules.
+
+For native replay, the companion `benchmarks/d7-repair` submission supplies the
+actual source-bound workers and numerical repair adapters. Build and qualification
+receipts must match before those adapters can be installed. Production defaults
+and container builds are unaffected by this support package.
+
+The Python namespace is retained to preserve compatibility with existing sealed
+workers. Renaming and narrowing this dependency closure can be reviewed separately
+from the initial import. The standalone CPU suite is the publication-time check;
+the report describes the separate completed pinned-stack GPU campaign.
+
+The companion D7 repair bundle also supplies the source-preserving native call
+tape and four-column isolated-stage study. Its evidence auditor rejects missing
+layers, changed fixtures, unchecked reference output and failed state recovery.
+The combined support/repair CPU suite passes **412 tests**, with one retained
+compiler-artifact check skipped. Native qualification is reported separately in
+the [public report](https://github.com/Terrydaktal/d7-rdna4-report/blob/main/reports/d7-rdna4-2026-09-17/REPORT.md).

benchmarks/conformance/SOURCE_PROVENANCE.json

@@ -0,0 +1,58 @@
+{
+  "files": {
+    "src/qwen_r9700_lab/config.py": "799d3612a07a431463abaac657fcdcaabf749bd034d2767a212b98fef134c9ed",
+    "src/qwen_r9700_lab/conformance_artifacts.py": "4067a3648ab3a7cf08179d9848304f05f019c09a49654f6306a9020d48e0b704",
+    "src/qwen_r9700_lab/conformance_attention_cut.py": "74f810ca277fad47a7d8ca3877d5104368e01ca8493eaa216b6b9ab07694a0fb",
+    "src/qwen_r9700_lab/conformance_boundaries.py": "4548064e635243f945e443222809ef7f858b366cd545809bc2672c433c4c9c39",
+    "src/qwen_r9700_lab/conformance_campaign.py": "d8d8fbd45c9767f8bec07a96e65453e9e6c938289720be03587a60f8a011169d",
+    "src/qwen_r9700_lab/conformance_cli.py": "7228102934f5364dad27a639c8e3e860fbb07d2bc574db878e6849a15656d212",
+    "src/qwen_r9700_lab/conformance_control.py": "0ff173eeb849d6fa6f61960c3a08aee06508329cb4f1068a9ba7cc427da43ad3",
+    "src/qwen_r9700_lab/conformance_dispatch.py": "9e3ecdbf1a81e9123c7edabce425e442545725b312b5a2282573cbef14ee30f8",
+    "src/qwen_r9700_lab/conformance_execution_modes.py": "f1f651829bb1177e6acc6695396615ac17f3d91e59fe4f058af8cdd119c7817a",
+    "src/qwen_r9700_lab/conformance_faults.py": "22b348e8ca1577b5e66debbcde09890d6432f0230d56252d3e1b6dcb10a22366",
+    "src/qwen_r9700_lab/conformance_gate.py": "495e5f04a756c479131f15db11d9300a59ceda75cdfe332b688a668b69ffd998",
+    "src/qwen_r9700_lab/conformance_gdn_contract.py": "7c84ce706a87f7bb6264c3d57d1e005f9c07bf57f53bdba97ea5e1d102a406ed",
+    "src/qwen_r9700_lab/conformance_gpu_lease.py": "d224177846aab2d068de543ca35ae3473c752f0329a5490a9f9d9ef0a97c9622",
+    "src/qwen_r9700_lab/conformance_instrumentation.py": "e1f1c7d4ab4d7b73f535e177809ffe25546209352af6b67e76c21a35a55f96eb",
+    "src/qwen_r9700_lab/conformance_invariants.py": "ec146c9785835d244eefd8562fc23aff1f8899676351515f7921dbe517561394",
+    "src/qwen_r9700_lab/conformance_lifecycle.py": "5c38d401e2e01102d2194ad8e1427eb39b72759e3ea2346fdf9dd76d1e3f5727",
+    "src/qwen_r9700_lab/conformance_mode_boundaries.py": "b50dcf80a9a6c6de5cae6d30779d6feefe4edeed24b4825e20b74baadac24275",
+    "src/qwen_r9700_lab/conformance_model.py": "3e0d2cca336b3f86b9f3a552a716519a359995a99e28e5c9658b1732b3210f44",
+    "src/qwen_r9700_lab/conformance_native_reference.py": "e1e3b4cebe62e46a398cd6070656c45ea756335121941a6759711677195746a3",
+    "src/qwen_r9700_lab/conformance_obligations.py": "ba3f91e6c473302998eec3b087daeaef9578f6395286c0c91dc6cd1915af58df",
+    "src/qwen_r9700_lab/conformance_observer.py": "d3a73a96e0871c53d7a198e36925c90ef4095cdfe6feda69b14a8549b100ad26",
+    "src/qwen_r9700_lab/conformance_parser.py": "6964d502c8b3cc51fc910237f6b5a8d476cab8de9ac9b40b4dbb13942820048a",
+    "src/qwen_r9700_lab/conformance_precision_intervention.py": "1212c666c14a72982f948f7a145074150117fc8f2315e3289ad40171892bbafe",
+    "src/qwen_r9700_lab/conformance_priority.py": "1cf992191a1736ee4848252091a8dbba56bbdb6a9c35cd90e14317dab7ec3926",
+    "src/qwen_r9700_lab/conformance_proofs.py": "979a10bd8203f4c3142f36ef9a40bf1532c185e474d3adbf677fdb49cef05428",
+    "src/qwen_r9700_lab/conformance_protocol.py": "c290f71871d34dcce9fe277da682a849c6c24f2358944852cdc39a48711dc89f",
+    "src/qwen_r9700_lab/conformance_queue.py": "d1ccf1d9d73cf703caf218324ca5f03767910e25ef1e73e903f23884426cdf95",
+    "src/qwen_r9700_lab/conformance_radiance.py": "d4980810ce03b2aa451738f321ba4346e5b8192f3034605b08674ab3888568a3",
+    "src/qwen_r9700_lab/conformance_reference.py": "0bc2e74d1295d53d839c822661e42ce8bddd36e63e36853418e9544187aa1e04",
+    "src/qwen_r9700_lab/conformance_reference_store.py": "720258558c8152c2f8e99035926a3abee95c111359c1a323e7da9098ba3d10b1",
+    "src/qwen_r9700_lab/conformance_replay.py": "178fcf75d2e88979c34085fba110df1df25277bbd30af543c905e109ce3a3834",
+    "src/qwen_r9700_lab/conformance_rotary_intervention.py": "a4ec7ce8cfebb0d715c5c259f6eb092dcf8854aeeee42e01a7f54c5988a19c57",
+    "src/qwen_r9700_lab/conformance_rotary_repair.py": "98f464b0dd53c2bbb44a4cc9e9307d520202a81a9a3175a3c091c8d95da050a0",
+    "src/qwen_r9700_lab/conformance_runtime.py": "830e09e70d609c28abe3e23fa13b6cb96176c70cb5aeba0da419029338d967ef",
+    "src/qwen_r9700_lab/conformance_scenarios.py": "57d5f84bd1db875a13f3c61fe65f5118de388e2736bb0e865558ee624472351f",
+    "src/qwen_r9700_lab/conformance_session.py": "3069b56e1af18dbc5db365968807459e45936f67ae0d1f8731250b02450b81b6",
+    "src/qwen_r9700_lab/conformance_shm.py": "eab5f933b6ce3be6f0942713230699d31ecfb4a73101d5776a523aa774b47e8c",
+    "src/qwen_r9700_lab/conformance_stage_isolation.py": "312a7e8353530e2807b89ecc5111a7de5bce3cc2dd56f850dd662490fd860149",
+    "src/qwen_r9700_lab/conformance_state.py": "42ade68e94bf53594b53b17ee7e916a77c443c18fe4b90a795e7ce510ac38c69",
+    "src/qwen_r9700_lab/conformance_topk.py": "940edebf21f2e8dd335ad52d37d7507d86a41e7fafc8884a5499e704069f30ce",
+    "src/qwen_r9700_lab/conformance_transport.py": "2bcf4eba295ea245c68d4f8bb8af28b68e60397086df72435ebd11973079cdef",
+    "src/qwen_r9700_lab/diagnostic_contract.py": "603680020b6a69d728d615eef9815b424610b4979410f223ab76e28dc4ee779e",
+    "src/qwen_r9700_lab/exact_fp8_metrics.py": "6c7aacb3a50e0f17c3f4eafdcd90b64b5efab8a63c9e0e8e57be98431d94a2bd",
+    "src/qwen_r9700_lab/manifests.py": "1eaea2743bcd851d54e6fc4809b5a26e9ab33a755fdbf547264183fc70ec6485",
+    "src/qwen_r9700_lab/ordered_reference_linear.c": "314ff8a91cd06dd14eafd5ee5695a2d90f77c5186820797a4460cb182fb3d272",
+    "src/qwen_r9700_lab/ordered_reference_linear.py": "7a8d5559918a3a21f2f0491f26942455f14aa0fe71b8a6029d4fcb80c4319bc7",
+    "src/qwen_r9700_lab/radiance_cache.py": "04adc11af696ab006570ee481b9e0b04d673e4bc6282469ea03061e6567abe03",
+    "tests/test_conformance_execution_modes.py": "cf4d3d64365c7797460f02a152bb43f50875a9ef3f79a1a4bed1d4f8d02c98b6",
+    "tests/test_conformance_rotary_repair.py": "0d5b771d005e805f4f3f0251ed82aab6ad1c39c0182f0811c7614de3f69e312b",
+    "tests/test_conformance_stage_isolation.py": "08e448a1620959e153907ad4b9d8d119a65b50990430946aebabab34c9c9df63",
+    "tests/test_d7_precision_comparison.py": "b40d22a9f2f27c63d0933b2f7a7e26f5804a3f9a406a2e28c5743c2032972902",
+    "tests/test_d7_rotary_comparison.py": "c78dd032b9aa9173da36fe5e16841f454207a8ca806065e6225d36d8c3a62e03"
+  },
+  "origin": "https://github.com/Terrydaktal/qwen-r9700-lab",
+  "report": "https://github.com/Terrydaktal/d7-rdna4-report/blob/main/reports/d7-rdna4-2026-09-17/REPORT.md"
+}

benchmarks/conformance/configs/profiles/gdn-stock-m1-arithmetic-v1.json

@@ -0,0 +1,152 @@
+{
+  "schema": "urn:qwen:gdn-arithmetic-contract:v1",
+  "id": "stock-packed-gdn-m1-gfx1201-v1",
+  "scope": "GDN core from raw convolution Q/K/V and gate projections to output plus recurrent state; other operators have separate contracts.",
+  "status": "DECLARED; native cross-mode equivalence UNPROVED",
+  "selection": "Pinned stock vLLM packed single-token GDN, applied causally to each materialized token. R4D and NumPy remain separately identified implementations.",
+  "domain": {
+    "query_heads": 16,
+    "value_heads": 48,
+    "key_width": 128,
+    "value_width": 128,
+    "batch_sequences": 1,
+    "qkv_width": 10240,
+    "state_order": [
+      "head",
+      "value",
+      "key"
+    ],
+    "state_elements": 786432,
+    "inputs": "Finite BF16 Q/K/V, a, b, dt_bias; finite FP32 A_log and initial state. Invalid geometry, aliasing, nonfinite output/state or incomplete evidence fails closed."
+  },
+  "precision": {
+    "qkv": "BF16 widened exactly to FP32",
+    "a_b_dt_bias": "BF16 widened exactly to FP32",
+    "A_log": "FP32",
+    "qk_normalization": "FP32 throughout; no intermediate BF16 cast",
+    "beta": "sigmoid in the pinned FP32 kernel, then BF16 round-to-nearest ties-even, then exact FP32 widening",
+    "decay_gate": "FP32",
+    "recurrent_state": "FP32 on every read, update and committed write",
+    "output": "BF16 round-to-nearest ties-even"
+  },
+  "operations": [
+    {
+      "name": "normalize",
+      "formula": "n(x) = x / sqrt(sum(x_i*x_i) + float32(1e-6))",
+      "arithmetic": "Pinned compiled reduction, square root and division. No replacement with a differently rounded reciprocal-square-root multiply."
+    },
+    {
+      "name": "scale_query",
+      "formula": "q = float32(n(Q) * float32(128**-0.5)); k = n(K)",
+      "arithmetic": "Scale q before the output dot product; not after it."
+    },
+    {
+      "name": "gates",
+      "formula": "x = float32(a + dt_bias); softplus = x if x > 20 else log(1 + exp(x)); g = -exp(A_log)*softplus; beta = widen(BF16_RNE(sigmoid(b)))",
+      "arithmetic": "The pinned kernel intrinsics and their compiled implementations define rounding. No cumulative-gate reconstruction or future-dependent fallback."
+    },
+    {
+      "name": "decay",
+      "formula": "D = float32(exp(g)*S)"
+    },
+    {
+      "name": "prediction",
+      "formula": "p = dot_stock(D,k)"
+    },
+    {
+      "name": "residual",
+      "formula": "r = float32(float32(v-p)*beta)"
+    },
+    {
+      "name": "state",
+      "formula": "S_next = update_stock(D,r,k)",
+      "arithmetic": "Retain the compiled multiply-add contraction; separate multiply/add is a different numerical path."
+    },
+    {
+      "name": "output",
+      "formula": "o = BF16_RNE(dot_stock(S_next,q))"
+    }
+  ],
+  "arithmetic_authority": {
+    "kind": "pinned-native-executable",
+    "note": "Real-number formulae describe dataflow. Exact finite-precision results are defined by the retained kernel artifact and packing below, under the declared compiler/hardware assumptions. Source inspection or matching hashes is not proof of runtime dispatch or functional correctness.",
+    "kernel": "fused_recurrent_gated_delta_rule_packed_decode_kernel",
+    "source": {
+      "fused_recurrent.py": "00a3b971b0dbb6ed26e246970a0e1a21a9a174030974685bdd3f0c8ab5fb4bfe",
+      "fla_op.py": "456da84fd12411cf53c96a90dd4f78f49afb454c3d51d2c524ea8e7106fe8ae5"
+    },
+    "cache_entry": "XW6RXQHLHKDBX2FF2LE7HICYZGQCSR5L4HJZ7USZF4L2FZQULJTQ",
+    "files": {
+      "fused_recurrent_gated_delta_rule_packed_decode_kernel.source": "f47615b4e3c0a4136cb3d5f7e526f222c8609e73dff2d89111c34ba1745ac1a7",
+      "fused_recurrent_gated_delta_rule_packed_decode_kernel.ttir": "0333349d4ff544078a5d35354a5ca42d7fe37360af7ce188d9778dba7887aabf",
+      "fused_recurrent_gated_delta_rule_packed_decode_kernel.ttgir": "889064969acdafd8931d3d7ae02e5b0ce98fdb80ac1f9df9fb919b5259c2645c",
+      "fused_recurrent_gated_delta_rule_packed_decode_kernel.llir": "eb6f3975ed77e9f0b53975d1d1c0739abdba06813e64bea8aa6f66baeadd19f2",
+      "fused_recurrent_gated_delta_rule_packed_decode_kernel.amdgcn": "f07d71ae362c0e87633253a85c2571e0ac626ff8b520034e16b554687ec4e1a1",
+      "fused_recurrent_gated_delta_rule_packed_decode_kernel.hsaco": "9feeb930b8635122c552ad727fbe897d0854332f9a7880468662b9ed83350135",
+      "fused_recurrent_gated_delta_rule_packed_decode_kernel.json": "de1084d3b9e844ded01cd70f866eda11618a6e95e1471b0ccc0f295921c96fad"
+    },
+    "target": {
+      "backend": "hip",
+      "arch": "gfx1201",
+      "warp_size": 32
+    },
+    "compiler": {
+      "triton_version": "3.7.1",
+      "num_warps": 1,
+      "num_stages": 3,
+      "enable_fp_fusion": true,
+      "allow_flush_denorm": false,
+      "sanitize_overflow": true
+    },
+    "environment": {
+      "FLA_USE_FAST_OPS": "0"
+    },
+    "packing": {
+      "query_heads": 16,
+      "value_heads": 48,
+      "K": 128,
+      "V": 128,
+      "BK": 128,
+      "BV": 32,
+      "state_token_stride": 802816,
+      "a_token_stride": 96,
+      "b_token_stride": 96,
+      "index_stride": 1,
+      "state_slot": 1,
+      "grid": [
+        4,
+        48
+      ],
+      "scale_fp32_hex": "f304b53d",
+      "note": "Single row makes the packed-QKV row stride irrelevant to this invocation. Reject or separately qualify other physical packing; physical block numbers are not semantic state."
+    }
+  },
+  "execution_modes": {
+    "m1": "One invocation F(S,u) with exactly one materialized input token.",
+    "prefill": "Ordered fold of that same F over tokens. Empty segment is the identity. Every intermediate prefix must match.",
+    "d7": "Verification rows are u[0]=previously emitted pending token, u[1:8]=seven proposals. For accepted proposal count k in 0..7, commit the state after rows 0..k (k+1 processed rows); the newly emitted correction/bonus token stays pending.",
+    "rollback": "Rejected rows cannot affect committed GDN state, convolution history, KV, positions or events. The GDN-only contract does not certify those other components.",
+    "snapshot": "Persist the contract identity with state; do not treat states created under a different arithmetic contract as interchangeable without checking or rebuilding."
+  },
+  "obligations": {
+    "partition": "fold(F,S,A+B) == fold(F,fold(F,S,A).state,B) for outputs and state, including empty segments",
+    "causal_prefix": "Changing inputs after prefix p leaves every output/state through p byte-identical.",
+    "accepted_prefix": "D7_commit(S,u,k) == fold(F,S,u[:k+1]).state with a separate pending-token field.",
+    "exactness": "Compare BF16 output bits and every FP32 logical-state bit; hashes and error tolerances alone do not establish equality."
+  },
+  "separate_targets": {
+    "numpy-serial-fp32-v1": "Independent canonical model. It rounds normalized Q/K to BF16, uses ordered non-fused reductions and scales the final output dot. It is not relabelled as this stock kernel.",
+    "r4d-unfused-m1": "Retains FP32 beta, uses its own reduction/normalization intrinsics, disables FMA and rounds output ties away. Existing prefill candidate targets this path, not this contract.",
+    "huggingface-fallback": "The pinned Transformers fallback also returns BF16 beta, but its normalization and framework reductions differ. Full bit equality with it is UNPROVED."
+  },
+  "required_qualification": [
+    "Reproduce the preserved stock first transition exactly with the selected invocation packing.",
+    "Compare repaired prefill/M1/D7 against this oracle from equal independent initial state.",
+    "Exercise every accepted width, all-prefix equality, future-suffix mutations, padding, nonzero state, extreme gates and BF16 halfway boundaries.",
+    "Then qualify graph execution, full-model logits/state, snapshots and long contexts; benchmark separately."
+  ],
+  "production_changed": false,
+  "existing_matrix_contract_changed": false,
+  "native_repair_qualified": false,
+  "sha256": "d8149b00210c94acfca4f1f66030de302dcaf8155894de6a23d3f4b76da61e00"
+}

benchmarks/conformance/pyproject.toml

@@ -0,0 +1,38 @@
+[project]
+name = "radiance-conformance"
+version = "0.1.0"
+requires-python = ">=3.12"
+dependencies = [
+    "numpy>=2",
+    "jsonschema>=4.23,<5",
+    "PyYAML>=6",
+    "psutil>=6",
+    "requests>=2",
+    "safetensors>=0.5",
+]
+
+[project.optional-dependencies]
+cpu-tests = [
+    "torch>=2.14.0",
+]
+
+[dependency-groups]
+dev = ["pytest>=8", "z3-solver>=4.13"]
+
+[build-system]
+requires = ["hatchling"]
+build-backend = "hatchling.build"
+
+[tool.hatch.build.targets.wheel]
+packages = ["src/qwen_r9700_lab"]
+
+[[tool.uv.index]]
+name = "pytorch-cpu"
+url = "https://download.pytorch.org/whl/cpu"
+explicit = true
+
+[tool.pytest.ini_options]
+pythonpath = ["src", "tests"]
+
+[tool.uv.sources]
+torch = { index = "pytorch-cpu" }

benchmarks/conformance/src/qwen_r9700_lab/__init__.py

@@ -0,0 +1 @@
+"""Source-bound conformance helpers; importing them does not activate inference."""