Record adverse DRI-1A candidate result

RESEARCH-HYPOTHESES.md

@@ -59,9 +59,19 @@ simultaneous failure domains, discover undisclosed causal structure, or grant
 action authority. A positive oracle result with a failed selector supports the
 concept but not runtime deployability.
 
-**Current evidence:** Three constructed conformance fixtures exist in
-`benchmark/decision-relative-independence-v0.1.json`. They validate evaluator
-semantics only and are not evidence for DRI-1.
+**Current evidence:** A noncanonical DRI-1A candidate run evaluated all 8,192
+preregistered synthetic worlds and failed its joint success criterion. The
+decision-relative oracle achieved
+90.99% correct settlement versus 60.78% for agent headcount and 68.32% for the
+best fixed-cut correct-settlement rate, and recovered 100% of registered
+minority reversals. But it did not reduce false settlement by the required 15
+points against every fixed cut: the coarsest fixed cut reached 3.26% false
+settlement by abstaining 40.23% of the time, and three fixed cuts could not be
+abstention-matched within tolerance. The explicit rules engine equalled the
+oracle by construction and therefore demonstrates deterministic execution in
+this fixture, not causal cut inference or deployment readiness. The protocol
+lacked several fields required for canonical promotion, so this result remains
+a content-bound candidate diagnostic. See `results/dri1a-v1/`.
 
 ## HVI-1 — verifier independence under shared control
 

experiments/dri1/EXECUTION-CONFIG.json

@@ -0,0 +1,40 @@
+{
+  "schema": "minority-prophet.dri1a-execution-config.v1",
+  "experiment": "DRI-1A",
+  "status": "preregistered-unexecuted",
+  "generator_salt": "minority-prophet-dri1a-v1",
+  "failure_domains": [
+    "machine_local",
+    "shared_controller",
+    "copied_source",
+    "shared_upstream_component"
+  ],
+  "independent_root_accuracies": [0.65, 0.75, 0.85, 0.95],
+  "erroneous_root_amplifications": [1, 3, 7, 15],
+  "decision_classes": {
+    "low_reversible": 2,
+    "high_irreversible": 3
+  },
+  "replicates_per_cell": 64,
+  "causal_roots_per_world": 5,
+  "candidate_cuts": [
+    "agent",
+    "machine",
+    "controller",
+    "evidence_origin",
+    "upstream_component"
+  ],
+  "cut_policy": {
+    "machine_local": "machine",
+    "shared_controller": "controller",
+    "copied_source": "evidence_origin",
+    "shared_upstream_component": "upstream_component"
+  },
+  "coarse_groups": [[0, 1], [2, 3], [4]],
+  "threshold_sensitivity": [1, 2, 3, 4, 5],
+  "abstention_match_tolerance": 0.01,
+  "minimum_false_settlement_reduction": 0.15,
+  "maximum_rules_unnecessary_abstention_increase": 0.10,
+  "minimum_material_world_fraction": 0.20
+}
+

experiments/dri1/PREREGISTRATION.md

@@ -0,0 +1,144 @@
+# DRI-1A — decision-relative independence policy benchmark
+
+**Status:** protocol, generator, baselines, metrics, thresholds, and adverse-result
+path frozen before the confirmatory worlds are generated or scored.
+
+## Question
+
+When the same evidence record contains roots at several lineage cuts, does using
+the cut relevant to the declared failure domain reduce false settlement relative
+to agent headcount or any one fixed global cut? Can an explicit deterministic
+policy execute that choice without model inference or material latency?
+
+This first arm tests the value and executable plumbing of a declared policy. It
+does **not** test whether a model can infer the right causal boundary from vague
+natural language. A positive rules-engine result is policy conformance, not
+learned causal intelligence.
+
+## Frozen synthetic world generator
+
+Generate the complete Cartesian product below in lexicographic order, then use
+an independent `random.Random` stream seeded from SHA-256 of
+`minority-prophet-dri1a-v1|<cell>|<replicate>` for every world:
+
+- failure-domain families: `machine_local`, `shared_controller`,
+  `copied_source`, `shared_upstream_component`;
+- independent-root accuracy: `0.65`, `0.75`, `0.85`, `0.95`;
+- erroneous-root amplification: `1`, `3`, `7`, `15` observations;
+- decision class: `low_reversible`, `high_irreversible`;
+- 64 replicates per cell.
+
+This yields 8,192 worlds. Each world has a uniformly sampled Boolean truth and
+five causally independent roots at the registered relevant cut. Each root
+reports truth with the cell accuracy. A correct root emits one observation. An
+incorrect root emits the registered amplification count, representing correlated
+repetition of the same wrong observation. All emitted records retain identities
+at all five candidate cuts.
+
+The relevant cut and the deliberately finer/coarser alternatives are frozen as:
+
+| Failure domain | Relevant cut | Finer cuts | Coarser cuts |
+|---|---|---|---|
+| `machine_local` | `machine` | `agent`, `evidence_origin` | `controller`, `upstream_component` |
+| `shared_controller` | `controller` | `agent`, `machine` | `evidence_origin`, `upstream_component` |
+| `copied_source` | `evidence_origin` | `agent`, `machine`, `controller` | `upstream_component` |
+| `shared_upstream_component` | `upstream_component` | `agent`, `machine`, `controller`, `evidence_origin` | none |
+
+At the relevant cut, every observation from one causal root shares that root ID.
+At a finer cut, every repeated observation has a distinct ID. At a coarser cut,
+causal roots `0/1`, `2/3`, and `4` are merged. A coarse root carrying conflicting
+values is handled by the existing fail-closed root-vote kernel; it is not repaired
+by the experiment.
+
+`low_reversible` requires two winning roots. `high_irreversible` requires three.
+This consequence label changes only the common evidence-sufficiency threshold;
+it does not change truth, lineage, or the relevant causal cut.
+
+## Frozen methods
+
+1. **Agent headcount:** always select `agent`.
+2. **Fixed machine:** always select `machine`.
+3. **Fixed controller:** always select `controller`.
+4. **Fixed evidence origin:** always select `evidence_origin`.
+5. **Fixed upstream component:** always select `upstream_component`.
+6. **Oracle policy:** select the registered relevant cut for the world family.
+7. **Explicit rules engine:** apply the frozen family-to-cut table above to the
+   declared `failure_domain`. No evidence values, truth labels, or model calls are
+   available to the selector.
+8. **Most-common-cut selector:** always select the lexicographically first cut
+   among equally frequent oracle cuts (`controller` here). This is the cut-selection
+   baseline; it is separate from the five aggregation baselines.
+
+All aggregation methods use `provenance.decision_relative.assess_decision` and
+the same per-world sufficiency threshold. No result-specific rescue, weighting,
+or confidence score is permitted.
+
+## Frozen metrics
+
+Report overall and per-stratum:
+
+- false-settlement rate: settled on the side opposite hidden truth, over all worlds;
+- abstention rate;
+- correct-settlement rate;
+- unsupported-settlement rate: method settles where the oracle is unsettled;
+- unnecessary-abstention rate: method abstains where the oracle settles correctly;
+- exact disposition agreement with oracle;
+- minority-reversal recovery: on worlds where agent observation majority and
+  causal-root majority are both decisive and disagree, the fraction settling on
+  the causal-root side;
+- selected-cut accuracy for oracle, rules engine, and most-common selector;
+- fraction of worlds with at least one decision-material alternative cut;
+- evaluator latency in milliseconds at p50, p95, and p99; and
+- deterministic result SHA-256 from two independent executions.
+
+The primary false-settlement comparison uses the registered threshold directly.
+As a conservative sensitivity analysis, also compare at approximately matched
+abstention. For each fixed-cut baseline choose the threshold in `1..5` minimizing
+absolute abstention-rate distance from the oracle. Ties favor the baseline by
+lower false settlement, then lower threshold. Report the remaining abstention
+difference; do not call rates matched when the absolute difference exceeds 0.01.
+
+## Frozen success criterion
+
+The joint DRI-1A policy-value claim is supported only if all conditions hold:
+
+1. oracle false settlement is at least 0.15 lower than **every** fixed-cut
+   baseline at the registered thresholds;
+2. the same 0.15 minimum reduction holds against every approximately
+   abstention-matched fixed-cut baseline whose abstention difference is at most
+   0.01;
+3. if no threshold for a baseline reaches the 0.01 matching tolerance, that
+   comparison is marked unmatched and the joint criterion fails rather than
+   dropping the baseline;
+4. rules-engine selected-cut accuracy is `1.0` and its false-settlement,
+   abstention, and disposition outputs equal the oracle byte-for-byte;
+5. the most-common-cut selector accuracy is no greater than `0.25`;
+6. rules-engine unnecessary abstention exceeds oracle by no more than 0.10;
+7. at least 20% of worlds have a decision-material alternative cut; and
+8. two complete executions produce identical semantic results, excluding only
+   measured wall-clock latency fields.
+
+These thresholds are intentionally difficult. Failure is a result, not grounds
+to change the generator or scoring rule.
+
+## Null, adverse outcomes, and kill implications
+
+The null is that the oracle policy has no lower false-settlement rate than the
+best fixed cut at comparable abstention. The strongest adverse outcomes are:
+
+- one fixed cut matches the oracle, making decision relativity unnecessary in
+  this threat model;
+- oracle gains disappear once abstention is matched;
+- the method gains safety only through prohibitive abstention;
+- material cut changes are rare;
+- the rules engine cannot reproduce the declared policy deterministically; or
+- latency is large enough to undermine a tactical graph/ledger path.
+
+A positive result remains evidence only for this frozen synthetic common-cause
+model. It does not validate the supplied lineage, demonstrate real-world
+frequency, establish joint multi-cut independence, calibrate confidence, infer
+failure domains, authorize action, or show that customers will provide the
+metadata. DRI-1B must separately test blinded human/model cut selection on
+externally authored cases; a field study must test whether the required lineage
+exists without self-reporting.
+

experiments/dri1/__init__.py

@@ -0,0 +1 @@
+"""DRI-1 decision-relative independence experiments."""