fix(research): seal training replay semantics

Author: erinepshovel-code

research/from-photons-to-macroverse/PREREGISTRATION.md

@@ -1,7 +1,7 @@
 # Preregistration: matched arity and recursion discrimination
 
 ```text
-protocol version: 0.4.0
+protocol version: 0.5.0
 run status:       not-run
 human subjects:   none
 animal subjects:  none
@@ -127,7 +127,7 @@ the domain-field order below. The complete tuple, including `block`, is the
 counter-stream key; concatenated placeholders or language-native float strings
 are forbidden. Every listed field is present; use the literal token `"none"`
 for an inapplicable field. A nested carrier or tensor component is one token
-such as `outer/3/leaf/2`, not an implementation-native tuple string.
+such as `outer/2/leaf/2`, not an implementation-native tuple string.
 
 Each digest is split in order into four big-endian unsigned 64-bit words and
 each word `w_k` maps to the open-unit uniform `u_k=(w_k+0.5)/2^64`. One scalar
@@ -364,7 +364,7 @@ whose indices are `start,...,start+15`:
 - a modulation multiplies the selected `W` summand or the selected all-source
   product term, named in the plan, by exactly `0.5` before `tanh`.
 
-Nested plans name `outer/i` or `leaf/i/l` targets and apply the explicit nested
+Nested plans name `outer/i` or `outer/i/leaf/l` targets and apply the explicit nested
 equation above. Class `5` uses the primary-outcome certificate below. Class `6`
 composes the two corresponding operators in numeric class order; masks and
 modulators act before `tanh`, then additive state, process noise, and ablation
@@ -373,10 +373,13 @@ episode ids, times, targets, and process-noise stream keys. Controls receive the
 plan through these fixed operators and the adapter below, never through a
 learned or undeclared `u_t` channel.
 
-The executable's candidate registry is the complete representation: `state`
-and `recovery` use a token observed identity plus an unsigned coordinate;
-`carrier` uses one token; and `edge` and `summand` use three tokens. Numeric
-strings never alias unsigned integers.
+The executable's `canonical_candidate_lists(system_kind,arity)` output is the
+complete candidate vocabulary and list for every schedule. `state` and
+`recovery` use a canonical observed identity plus an unsigned coordinate;
+`carrier` uses a canonical direct carrier or outer-carrier identity; and `edge`
+and `summand` use the exact term, target, and source identities emitted there.
+Numeric strings never alias unsigned integers, and no other token spelling is
+admissible.
 
 ## Outcomes
 
@@ -593,8 +596,11 @@ serial left fold from `1.0` in that same order. Loss terms within a minibatch ar
 left-folded by `draw_index`, then transition index, then lexicographic original
 observed scalar identity. Recovery squared coordinates and every reported mean
 use the corresponding lexicographic original-identity order and a serial left
-fold before the single final division. Reverse-mode gradients are derivatives
-of this ordered scalar program and accumulate contributions in the same order.
+fold before the single final division. Reverse-mode gradients use only the
+pinned executable's binary multiply/divide, serial sum/product, `tanh`,
+`softplus`, variance-head, and Gaussian-NLL adjoint functions and accumulate
+contributions in the same order. `variance_head_value` and
+`gaussian_nll_term` are the sole scalar variance and NLL forward recurrences.
 Reassociation, tree or parallel reductions, model-position ordering, and fused
 multiply-add contraction are forbidden. These rules apply before fitting,
 during every optimizer update, and during rollout and scoring, so a carrier
@@ -665,10 +671,13 @@ carrier `a` receives the arithmetic mean of all observed carriers with
 `q_{m,n}(i)=a`; for `m > n`, control carrier `a` receives observed carrier
 `r_{m,n}(a)`. Define the model-carrier image of observed carrier `i` as
 `A_{m,n}(i)={q_{m,n}(i)}` for `m<n` and
-`A_{m,n}(i)={a:r_{m,n}(a)=i}` for `m>n`. Carrier-targeted additive interventions are encoded by the same
-map before dynamics: merged controls receive the arithmetic mean of all
-interventions in their bucket, and split controls receive the same intervention
-on every split carrier assigned to the targeted observed carrier. Edge cuts with
+`A_{m,n}(i)={a:r_{m,n}(a)=i}` for `m>n`. Carrier-targeted additive interventions
+use the same map only to identify model output coordinates: merged controls
+receive the arithmetic mean of all additions in their bucket, and split controls
+receive the same addition on every split carrier assigned to the targeted
+observed carrier. Every resulting addition is applied after that model carrier's
+deterministic `tanh` transition and before process noise; it is never applied to
+the input encoding or before dynamics. Edge cuts with
 endpoints collapsed into one merged carrier become recorded self-edge no-ops;
 all other cuts map to every nonself model edge in
 `A_{m,n}(source) x A_{m,n}(target)` and zero both those `W` summands and every

research/from-photons-to-macroverse/replay_reference.json

@@ -2,12 +2,12 @@
   "schema": "the-interdependency.arity-recursion-replay-pin",
   "version": "1.0.0",
   "protocol_id": "arity-recursion-synthetic-v6",
-  "protocol_version": "0.4.0",
+  "protocol_version": "0.5.0",
   "status": "not-run",
   "reference_path": "tools/replay_reference.py",
-  "reference_sha256": "67a17f0c33b28765e0189505ab6fc6f18a10530008211bf4e7abdcf9e2744fe9",
-  "contract_sha256": "12f1a245de8bda5efe1b04547f54ca4da961287f6f48fd3becbee91ce07c799e",
+  "reference_sha256": "3b3c6babd2340eb5d08aadbbfa3b2c3b21af75683210957c3730e72a52a70460",
+  "contract_sha256": "8584dd017603e0b5f4b6be7c6756e348b3f3255bcbfbe42f0294ea2e993b6ffe",
   "vectors_path": "replay_vectors.json",
-  "vectors_sha256": "053dd257e90d324ce022444d3a9d1d2fcd468d411533704a711f6405e0db3c33",
+  "vectors_sha256": "87ac49b3baa50430760ab0702ae8f13a374d3e0baac7e1283122767f6386d4b4",
   "verification_command": "python tools/replay_reference.py verify"
 }

research/from-photons-to-macroverse/replay_vectors.json

@@ -1,4 +1,80 @@
 {
+  "adjoints": {
+    "binary_divide_f64": [
+      "c004000000000000",
+      "bff4000000000000"
+    ],
+    "binary_multiply_f64": [
+      "bfe4000000000000",
+      "3fd4000000000000"
+    ],
+    "nll_adjoint_f64": [
+      "3ff4000000000000",
+      "bff4000000000000",
+      "3fe4000000000000"
+    ],
+    "nll_term_f64": "3fea50d048e7a1bd",
+    "serial_product_f64": [
+      "bfde000000000000",
+      "3fce000000000000",
+      "bfc4000000000000"
+    ],
+    "serial_sum_f64": [
+      "3ff4000000000000",
+      "3ff4000000000000",
+      "3ff4000000000000"
+    ],
+    "softplus_f64": "3fde34086befd717",
+    "tanh_f64": "3fef75398b8c1479",
+    "variance_adjoint_f64": "3fde34086befd717",
+    "variance_value_f64": "3fde574b2f73a891"
+  },
+  "candidate_registry": {
+    "direct": {
+      "carrier": {
+        "count": 7,
+        "sha256": "41db2dc9fa4b248d4968b6534dab975f2f31971093bb4a8b4b992767c715c02d"
+      },
+      "edge": {
+        "count": 42,
+        "sha256": "97f5a7c6a45b3d0c33091378a2b1bb7ffe6357d9cce5b2a810e51e79350c9a66"
+      },
+      "recovery": {
+        "count": 14,
+        "sha256": "b409688fa35f800e613a36ecb875012d15ffdf20709cc6405937d1088994b3f2"
+      },
+      "state": {
+        "count": 14,
+        "sha256": "13c084ebbeef451cf9c1040b8ff03e356bf7b8d24c62264e2e48670cc3e62e49"
+      },
+      "summand": {
+        "count": 49,
+        "sha256": "884b3477385b2a96634765a9f995a360a0f9a1d0007a61a17e9bdfb95e45d936"
+      }
+    },
+    "nested": {
+      "carrier": {
+        "count": 7,
+        "sha256": "daa699960d578300e3c8cf88fad39b68226ba8bfd67445d3384efc0f1a77fd27"
+      },
+      "edge": {
+        "count": 100,
+        "sha256": "4d581cee3e1b91c31eca8cab8d873a7c96d253902eec5395d0f26899a4c92c9d"
+      },
+      "recovery": {
+        "count": 44,
+        "sha256": "e9a6c3e0b3386e02916a9718fa028a183495d8bb8e532478b958c6dcb201f7db"
+      },
+      "state": {
+        "count": 44,
+        "sha256": "9b1c74e751fc629d74c3ef68b7395e9f786401a641d0781a8ac1c999c22c59c8"
+      },
+      "summand": {
+        "count": 129,
+        "sha256": "634e63f58287caf2b4a24faff0a6cd05764549627a3caa735c4c1c3e4964046a"
+      }
+    }
+  },
   "decision_edges": {
     "class5_cdf_f64": "3fe8000000000000",
     "cut_baseline_ascii": "[\"unperturbed\",\"cut\",\"direct/7\"]",
@@ -121,11 +197,6 @@
         13
       ]
     ],
-    "ordered_candidates_ascii": [
-      "[\"summand\",\"product\",\"outer/2\",\"none\"]",
-      "[\"summand\",\"w\",\"outer/2\",\"outer/0\"]",
-      "[\"summand\",\"w\",\"outer/2\",\"outer/1\"]"
-    ],
     "partition_key_ascii": "[\"arity-recursion-synthetic-v6\",32,7,50,\"intervention_plan\",\"model_partition/arbitrary/7\",[\"schedule/model_partition\",0,0],0]",
     "start_key_ascii": "[\"arity-recursion-synthetic-v6\",32,7,50,\"intervention_plan\",\"schedule\",[\"int/1/test/000\",1,\"start_time\"],0]",
     "state_candidate_ascii": "[\"state\",\"carrier/0\",0]",

research/from-photons-to-macroverse/tests/test_contracts.py

@@ -215,7 +215,7 @@ def test_preregistration_closes_exact_head_replay_blockers(self) -> None:
         flat = " ".join(text.split())
 
         for phrase in (
-            "protocol version: 0.4.0",
+            "protocol version: 0.5.0",
             '["arity-recursion-synthetic-v6",s,n,sigma_milli,domain,role,[k0,...,kp],block]',
             "concatenated placeholders or language-native float strings are forbidden",
             "stability/00` through `stability/15",
@@ -302,7 +302,7 @@ def test_executable_replay_reference_is_hash_pinned(self) -> None:
         pin = load_json("replay_reference.json")
         observed = replay.verify(PROJECT / "replay_reference.json")
         self.assertEqual(observed["protocol_id"], "arity-recursion-synthetic-v6")
-        self.assertEqual(observed["protocol_version"], "0.4.0")
+        self.assertEqual(observed["protocol_version"], "0.5.0")
         self.assertEqual(observed["reference_sha256"], pin["reference_sha256"])
         self.assertEqual(observed["contract_sha256"], pin["contract_sha256"])
         self.assertEqual(observed["vectors_sha256"], pin["vectors_sha256"])
@@ -345,8 +345,19 @@ def test_executable_replay_reference_is_hash_pinned(self) -> None:
         self.assertTrue(vectors["stability_probe"]["distinct"])
         self.assertIn('"intervention_plan","schedule"', vectors["schedule"]["target_key_ascii"])
         self.assertIn('"bootstrap","sealed_aggregate"', vectors["resampling"]["bootstrap_ascii"])
+        self.assertEqual(
+            vectors["candidate_registry"]["direct"]["state"]["count"], 14
+        )
+        self.assertEqual(
+            vectors["candidate_registry"]["nested"]["state"]["count"], 44
+        )
+        self.assertEqual(
+            vectors["candidate_registry"]["nested"]["edge"]["count"], 100
+        )
+        with self.assertRaises(ValueError):
+            replay.candidate_bytes("nested", 7, "state", "leaf/3/2", 0)
         with self.assertRaises(ValueError):
-            replay.candidate_bytes("state", 0, 0)
+            replay.candidate_bytes("nested", 7, "state", "outer/3/leaf/2", 0)
         self.assertTrue(vectors["noise_phase"]["distinct"])
         self.assertGreater(replay.open_uniform_word(0), 0.0)
         self.assertLess(replay.open_uniform_word(2**64 - 1), 1.0)