feat(observations)!: dated events never collapse across dates (D106)

benchmarks/locomo/README.md

@@ -1,4 +1,4 @@
-# RS-LoCoMo-Full-v20 setup
+# RS-LoCoMo-Full-v21 setup
 
 This directory contains the unshipped full-system LoCoMo adapter. It does not vendor or
 auto-download LoCoMo. Supply the exact pinned `locomo10.json` only after confirming its
@@ -16,12 +16,12 @@ The safe first command is local and makes no API or model call:
 uv run --extra benchmark python -m benchmarks.locomo prepare \
   --dataset /absolute/path/locomo10.json \
   --tier smoke \
-  --protocol full-v20 \
+  --protocol full-v21 \
   --output .benchmark-runs/locomo-smoke
 ```
 
 The harness validates the pinned bytes, renders session documents, and fingerprints the
-eight-question smoke plan. `--protocol` is prepare-only; `full-v20` is the one
+eight-question smoke plan. `--protocol` is prepare-only; `full-v21` is the one
 current-system protocol, and every later stage reads that immutable choice from
 `run.json`. Do not run remote stages until reviewing
 [`locomo_benchmark_design.md`](../../plan/designs/locomo_benchmark_design.md).
@@ -64,6 +64,17 @@ call sees the complete bounded candidate snapshot and returns a supplied
 candidate id or `new`. There is no insufficient-evidence result or
 confidence-routed frontier call.
 
+V21 fingerprints the D106 observation adjudicator: dated events with disjoint
+resolved windows never collapse onto or supersede each other (they may only
+contradict or stay distinct), a dated event is never `evidence` for an undated
+statement (nor the reverse), open-ended windows stay unbounded, and the verdict
+prompt shows when each statement was said and what time it is about. The
+`adjudicate_observations` component version pins
+that generation; the dataset, rendered documents, retrieval surface, answer
+and judge prompts, budgets, and scoring are those of v20. V20 and v21 scores
+are directional, not a one-variable comparison — the fact layer a v21 store
+serves differs from a v20 store's.
+
 Build the image from the revision under test — Compose otherwise serves the
 published release image, and the harness refuses to run against an engine whose
 stamped revision does not match the prepared run:

benchmarks/locomo/__init__.py

@@ -1 +1 @@
-"""The pinned RS-LoCoMo-Full-v20 benchmark adapter."""
+"""The pinned RS-LoCoMo-Full-v21 benchmark adapter."""

benchmarks/locomo/cli.py

@@ -118,7 +118,7 @@ def _parser() -> argparse.ArgumentParser:
     parser = argparse.ArgumentParser(
         prog="python -m benchmarks.locomo",
         description=(
-            "RS-LoCoMo-Full-v20: prepare is local; ingest/answer/judge require "
+            "RS-LoCoMo-Full-v21: prepare is local; ingest/answer/judge require "
             "explicit execution acknowledgements"
         ),
     )

benchmarks/locomo/dataset.py

@@ -140,10 +140,10 @@ def load_manifest(tier: str) -> QuestionManifest:
         )
     if manifest.dataset_commit != DATASET_COMMIT:
         raise DatasetValidationError(
-            "manifest dataset commit is not RS-LoCoMo-Full-v20"
+            "manifest dataset commit is not RS-LoCoMo-Full-v21"
         )
     if manifest.dataset_sha256 != DATASET_SHA256:
-        raise DatasetValidationError("manifest dataset hash is not RS-LoCoMo-Full-v20")
+        raise DatasetValidationError("manifest dataset hash is not RS-LoCoMo-Full-v21")
     actual = item_ids_hash(item_ids=manifest.item_ids)
     if actual != manifest.item_ids_sha256:
         raise DatasetValidationError(

benchmarks/locomo/model.py

@@ -1,4 +1,4 @@
-"""Typed values for the full-system RS-LoCoMo-Full-v20 protocol."""
+"""Typed values for the full-system RS-LoCoMo-Full-v21 protocol."""
 
 from __future__ import annotations
 
@@ -28,8 +28,8 @@
 Category = Literal[1, 2, 3, 4, 5]
 RetainedCategory = Literal[1, 2, 3, 4]
 Tier = Literal["smoke", "development", "publication"]
-ProtocolKey = Literal["full-v20"]
-ProtocolName = Literal["RS-LoCoMo-Full-v20"]
+ProtocolKey = Literal["full-v21"]
+ProtocolName = Literal["RS-LoCoMo-Full-v21"]
 SourceTimezoneBasis = Literal["assumed_utc"]
 AnswerAgentModel = Literal["openai/gpt-5.6-luna"]
 JudgeModel = Literal["openai/gpt-5.6-luna"]
@@ -120,7 +120,7 @@ class QuestionManifest(FrozenModel):
 class RunConfiguration(FrozenModel):
     """Immutable identity of one prepared benchmark run."""
 
-    protocol_name: ProtocolName = "RS-LoCoMo-Full-v20"
+    protocol_name: ProtocolName = "RS-LoCoMo-Full-v21"
     adapter_version: NonEmpty
     prepared_at: datetime
     repository_revision: NonEmpty
@@ -393,7 +393,7 @@ class SessionDiagnosticSummary(FrozenModel):
 class RunSummary(FrozenModel):
     """Publication-ready local aggregate with no hidden denominator."""
 
-    protocol_name: ProtocolName = "RS-LoCoMo-Full-v20"
+    protocol_name: ProtocolName = "RS-LoCoMo-Full-v21"
     protocol_fingerprint: NonEmpty
     tier: Tier
     questions: int = Field(ge=1)

benchmarks/locomo/protocol.py

@@ -29,9 +29,9 @@
 from rememberstack.model import Envelope
 from rememberstack.model import ToolDescriptor
 
-PROTOCOL_NAME: Final = "RS-LoCoMo-Full-v20"
-DEFAULT_PROTOCOL_KEY: Final = "full-v20"
-ADAPTER_VERSION: Final = "locomo-full-adapter-2026.09-complete-values-v20"
+PROTOCOL_NAME: Final = "RS-LoCoMo-Full-v21"
+DEFAULT_PROTOCOL_KEY: Final = "full-v21"
+ADAPTER_VERSION: Final = "locomo-full-adapter-2026.09-temporal-gate-v21"
 MAX_TOOL_CALLS: Final = 8
 MAX_AGENT_CALLS: Final = 9
 ANSWER_READER_RETRY_BUDGET: Final = 2
@@ -74,7 +74,7 @@
             "no-types-1:binary-t4-1:document-t0-1"
         ),
         "adjudicate_observations": (
-            "e3-obs-flush-2026.08a:claim-fanout-1:entity-fanout-1"
+            "e3-obs-flush-2026.09a:temporal-gate-1:claim-fanout-1:entity-fanout-1"
         ),
         "adjudicate_supersession": "adjudicator-2026.07b:temp0-1",
         "embed_claim": "p1-embed-claims-2026.07",
@@ -211,8 +211,8 @@ class LoCoMoProtocol:
     answer_word_cap: int | None = None
 
 
-_FULL_V20 = LoCoMoProtocol(
-    key="full-v20",
+_FULL_V21 = LoCoMoProtocol(
+    key="full-v21",
     name=PROTOCOL_NAME,
     answer_agent_model=ANSWER_AGENT_MODEL,
     judge_model=JUDGE_MODEL,
@@ -234,7 +234,7 @@ class LoCoMoProtocol:
 )
 
 PROTOCOL_REGISTRY: Final[Mapping[ProtocolKey, LoCoMoProtocol]] = MappingProxyType(
-    {_FULL_V20.key: _FULL_V20}
+    {_FULL_V21.key: _FULL_V21}
 )
 
 

benchmarks/locomo/runner.py

@@ -411,7 +411,7 @@ def ingest_sample(
         _require_current_ingest_bindings(model_bindings=build.model_bindings)
         if build.document_binding_generation != EXPECTED_DOCUMENT_BINDING_GENERATION:
             raise ExecutionGuardError(
-                "deployment document binding generation differs from RS-LoCoMo-Full-v20"
+                "deployment document binding generation differs from RS-LoCoMo-Full-v21"
             )
         _require_current_query_surface(context=context, client=client)
         _require_exact_live_ingests(
@@ -570,7 +570,7 @@ def answer_sample(
     ):
         raise ExecutionGuardError(
             "the deployment did not report the exact completed"
-            " RS-LoCoMo-Full-v20 pipeline, live graph, and fresh P3 projection"
+            " RS-LoCoMo-Full-v21 pipeline, live graph, and fresh P3 projection"
         )
     _require_serving_revision(context=context, readiness=readiness)
     prior_readiness = context.state.readiness.get(sample_id)
@@ -1177,7 +1177,7 @@ def _validate_run(
     """Recompute immutable run identity before any local or remote stage."""
     selected_protocol = protocol_for_name(configuration.protocol_name)
     if configuration.dataset_sha256 != DATASET_SHA256:
-        raise BenchmarkRunError("run dataset hash is not RS-LoCoMo-Full-v20")
+        raise BenchmarkRunError("run dataset hash is not RS-LoCoMo-Full-v21")
     if item_ids_hash(item_ids=manifest.item_ids) != manifest.item_ids_sha256:
         raise BenchmarkRunError("run manifest item hash changed")
     if manifest_bytes_hash(manifest=manifest) != configuration.manifest_sha256:
@@ -1187,7 +1187,7 @@ def _validate_run(
     if manifest.tier != configuration.tier:
         raise BenchmarkRunError("run manifest tier changed")
     if configuration.dataset_commit != DATASET_COMMIT:
-        raise BenchmarkRunError("run dataset commit is not RS-LoCoMo-Full-v20")
+        raise BenchmarkRunError("run dataset commit is not RS-LoCoMo-Full-v21")
     if configuration.adapter_version != ADAPTER_VERSION:
         raise BenchmarkRunError("run adapter version differs from current code")
     if _models_hash(values=documents) != configuration.documents_sha256:
@@ -1445,7 +1445,7 @@ def _require_current_ingest_bindings(*, model_bindings: dict[str, str]) -> None:
             if model_bindings.get(name) != expected.get(name)
         )
         raise ExecutionGuardError(
-            "deployment ingest model bindings differ from RS-LoCoMo-Full-v20: "
+            "deployment ingest model bindings differ from RS-LoCoMo-Full-v21: "
             + ", ".join(mismatches)
         )
 

benchmarks/locomo/sharding/README.md

@@ -178,7 +178,7 @@ as source-derived data.
 | --- | ---: | --- |
 | `LOCOMO_PYTHON` | `.venv/bin/python` | repository virtual-environment Python |
 | `LOCOMO_TIER` | `publication` | prepared manifest tier |
-| `LOCOMO_PROTOCOL` | `full-v20` | prepare-time protocol key |
+| `LOCOMO_PROTOCOL` | `full-v21` | prepare-time protocol key |
 | `LOCOMO_MOUNT_ROOT` | `$RUN_DIR/.mounts` | host/container-identical P3 mount root |
 | `LOCOMO_MAX_DOCUMENTS` | `100` | per-sample ingest authorization |
 | `LOCOMO_MAX_QUESTIONS` | `1540` | run-absolute answer item authorization |

benchmarks/locomo/sharding/run_shard.sh

@@ -25,7 +25,7 @@ run_dir=$2
 dataset_path=$3
 python_bin=${LOCOMO_PYTHON:-.venv/bin/python}
 tier=${LOCOMO_TIER:-publication}
-protocol=${LOCOMO_PROTOCOL:-full-v20}
+protocol=${LOCOMO_PROTOCOL:-full-v21}
 mount_root=${LOCOMO_MOUNT_ROOT:-$run_dir/.mounts}
 max_documents=${LOCOMO_MAX_DOCUMENTS:-100}
 max_questions=${LOCOMO_MAX_QUESTIONS:-1540}
@@ -50,7 +50,7 @@ export GOOGLE_APPLICATION_CREDENTIALS=${LOCOMO_GCP_CREDENTIALS_FILE:-/etc/rememb
 export GOOGLE_API_CERTIFICATE_CONFIG=${LOCOMO_GCP_CERTIFICATE_CONFIG_FILE:-/etc/rememberstack/locomo-gcs/certificate-config.json}
 export GOOGLE_API_USE_CLIENT_CERTIFICATE=true
 
-# RS-LoCoMo-Full-v20's non-secret ingest identity. Override ambient self-host
+# RS-LoCoMo-Full-v21's non-secret ingest identity. Override ambient self-host
 # defaults so every shard runs the exact Luna/Qwen pipeline the protocol checks.
 export REMEMBERSTACK_STRUCTURER_MODEL=openai/gpt-5.6-luna
 export REMEMBERSTACK_SKELETON_CHECK_MODEL=openai/gpt-5.6-luna

decisions.md

@@ -2910,7 +2910,11 @@ may proceed to its first tagged artifact proof after CLA activation.
 
 ## D78. LoCoMo measures the ordinary OSS query system, not a claims-only shortcut
 
-> **D105 amendment.** The current protocol is `RS-LoCoMo-Full-v20`. It retains
+> **D106 amendment.** The current protocol is `RS-LoCoMo-Full-v21`. It retains
+> v20's dataset, rendered documents, models, tools, budgets, prompts, and
+> scoring and pins the D106 observation adjudicator generation.
+
+> **Historical D105 amendment.** The D105 protocol was `RS-LoCoMo-Full-v20`. It retains
 > v19's surface, dataset, ingestion, models, tools, and budgets while adding the
 > D105 complete-direct-values answer instruction.
 >
@@ -4889,3 +4893,99 @@ retry; mention research or adoption in the prompt.
 **Amends.** Advances D104's Full-v19 benchmark identity only. It preserves D98
 retrieval, D99's objective content-before-`Unknown` guard, and D104's bounded
 counterfactual instruction.
+
+## D106. Dated events never collapse across dates; evidence requires temporal compatibility
+
+**Decision (2026-09-03).** The observation adjudicator (D43) gains a
+deterministic temporal-compatibility rung that runs before any model call,
+reading the D41 valid-time each claim already carries:
+
+1. Two statements that both describe **datable events**
+   (`claim_valid_kind = 'event_time'`) whose resolved windows are **disjoint**
+   are different occurrences unless the date-aware model finds they name the
+   *same* occurrence with disputed dates. Such a pair may therefore only
+   `contradict` (both stand, grouped) or stay `new`; an `evidence` or
+   `supersede` verdict is coerced to `new` and recorded. This holds even when
+   the wording is byte-identical — "won a tournament last week" said in
+   January and again in October is two wins — and identical text is then
+   kept apart without any model call.
+2. A dated event paired with an **undated** statement (a state, a period
+   figure, a summary such as "has been winning a few tournaments", or
+   testimony the extractor could not anchor) may still be judged for
+   `supersede`/`contradict` — a dated resignation can end an undated "is
+   CEO" state — but an `evidence` verdict for such a pair is coerced to `new`
+   and recorded, and identical text is kept apart without a model call. A
+   specific dated event is never a re-assertion of a vaguer statement, and a
+   summary never re-asserts a specific event.
+3. Two undated statements are adjudicated exactly as before; two dated
+   events with overlapping windows — including a year-level and a day-level
+   window for one occurrence — go to the ladder as before, and a re-mention
+   of the same event still collapses.
+4. A D41 open interval (a resolved start with no end) is unbounded, never a
+   point: it overlaps every later window.
+
+The verdict prompt now shows, for each statement, when the source *said it*
+and what world-time it *is about* (the resolved D41 window of any kind), and
+states the rule, so the model judges with both timelines in front of it. An
+observation's windows are the span of its supporting current-testimony
+claims, derived at block time and kept current within a batch as the row
+absorbs evidence; nothing is stored. Every adjudication record carries the
+coercions that preceded it. The adjudicator generation, the
+`adjudicate_observations` flush component version, and the LoCoMo protocol
+(Full-v20 → Full-v21) roll; nothing else about D43 — the entity block, the
+novelty gate, the no-cap rule, the fail-safe-to-coexist contract — changes.
+
+**Context.** The v0.11.0 LoCoMo `conv-42` run (protocol Full-v18,
+`plan/analysis/locomo_conv42_recurring_event_adjudication.md`) extracted all
+seven of Nate's tournament wins as clean, dated claims and resolved one Nate
+entity, yet the fact layer kept only four distinct win facts. The
+`observation_adjudications` transcript shows why: the small model saw two
+bare strings, judged "won a really big video game tournament last week"
+(October) to be `evidence` for "won his first video game tournament last
+week" (January) because both say "last week", and folded the international
+and Valorant wins into "has been winning a few gaming tournaments". Ten
+participation lineages collapsed into the header boilerplate "Nate is a
+participant.", swallowing two real tournament entries. The claims carried
+resolved event dates the whole time; the adjudicator never read them. The
+reader answered the counting question "At least five" against a gold of
+seven with all seven wins in its testimony context — the fact envelope's
+undercount is the plausible anchor.
+
+**Consequences.** Recurring same-shaped events survive as distinct facts, so
+entity-anchored counting and enumeration have a complete fact layer to read,
+while a genuine same-occurrence date dispute can still surface as a
+contradiction. Verdict spend is unchanged in shape (similar pairs are still
+judged; only identical-text decisions become model-free). The fail-safe
+direction is preserved and slightly widened: an undated re-mention of a
+dated event now stays a separate observation rather than collapsing — a
+duplicate, never a loss. Observation `valid_from` continues to be the
+claim's `asserted_at`; using the resolved window as the observation's own
+validity is a separate change this decision does not make.
+
+Rollout is stop-drain-rebuild. The `adjudicate_observations` flush handler
+reports the claimed unit's *own* component generation to its barrier, so
+entity units enqueued before the roll complete under the generation their
+barrier counts and drain cleanly; readiness then reports the stage at the old
+generation until the deployment is rebuilt under the new one (D7/D12).
+Existing stores re-adjudicate observations under the new generation on that
+rebuild.
+
+**Rejected.** Fixing only the verdict prompt (a model may still merge across
+dates; the coercions must be deterministic); skipping disjoint-date pairs
+without any verdict (would silently double-count one occurrence whose date
+two sources dispute, and suppress the contradiction group); treating every
+mixed-dating pair as non-interacting (would stop a dated event from
+superseding a state it ends); dropping vague summaries at extraction
+(source-faithful testimony is kept — D32 — the defect was absorption, not
+extraction); a typed period/value column (D43 §4 rejects the typed schema;
+the event window is already carried by the claim).
+
+**Design.** `plan/designs/observations_design.md` §3 (temporal compatibility
+rung) and `plan/designs/locomo_benchmark_design.md` Full-v21 protocol.
+
+**Analysis.** `plan/analysis/locomo_conv42_recurring_event_adjudication.md`.
+
+**Amends.** D43's adjudication cascade (adds the deterministic rung its
+design named but never implemented) and D105's Full-v20 benchmark identity.
+Preserves D41 (claims stay the one home of asserted validity), D98
+retrieval, D100/D102 identity, and D104/D105 answer-prompt contracts.