Add earth-pulse-harness: MetaHarness Darwin-Mode research pod for the 26-second pulse

harnesses/earth-pulse-harness/.claude-plugin/plugin.json

@@ -0,0 +1,25 @@
+{
+  "name": "earth-pulse-harness",
+  "version": "0.1.0",
+  "description": "Earth Pulse Observatory — a MetaHarness Darwin-Mode research pod for the 26-second microseism (Gulf of Guinea). Freeze the physics; evolve the harness.",
+  "author": {
+    "displayName": "Generated by metaharness",
+    "url": "https://www.npmjs.com/package/metaharness"
+  },
+  "license": "MIT",
+  "categories": [
+    "agent-harness",
+    "metaharness-scaffold",
+    "Research",
+    "geophysics"
+  ],
+  "tags": [
+    "metaharness",
+    "agent-harness",
+    "vertical:research",
+    "causal-discovery",
+    "seismology",
+    "ruvector"
+  ],
+  "homepage": "https://github.com/ruvnet/agent-harness-generator"
+}

harnesses/earth-pulse-harness/.claude/commands/doctor.md

@@ -0,0 +1,12 @@
+---
+description: "Health-check the harness: kernel load, MCP wiring, pipeline build, host adapter."
+---
+
+Run a full health check and print a PASS/FAIL table.
+
+1. Kernel loads and `kernelInfo().version` matches package.json.
+2. The MCP server starts and lists its tools.
+3. The detect→extract→embed→score pipeline runs over the bundled fixtures.
+4. The configured host adapter is present.
+
+Exit non-zero if any check fails.

harnesses/earth-pulse-harness/.claude/commands/review-diff.md

@@ -0,0 +1,11 @@
+---
+description: "Review the current working diff for correctness, leakage, and safety-policy compliance."
+---
+
+Review the current git diff.
+
+1. `git diff` to read the change.
+2. Report only high-confidence findings as `file:line — issue — fix`.
+3. Flag any forbidden mutation (fabricated observations, invented citations, test-window leakage, new imports/network/shell/env, weakened promotion gate).
+4. Separate bugs from nits.
+5. End with APPROVE or REQUEST-CHANGES and a one-line reason.

harnesses/earth-pulse-harness/.claude/settings.json

@@ -0,0 +1,41 @@
+{
+  "permissions": {
+    "allow": [
+      "Bash(npx earth-pulse-harness*)",
+      "mcp__earth-pulse-harness__*",
+      "mcp__code_index__*",
+      "Bash(npm test*)",
+      "Bash(npm run*)",
+      "Bash(git diff*)",
+      "Bash(git status*)",
+      "Bash(git log*)"
+    ],
+    "deny": [
+      "Read(./.env)",
+      "Read(./.env.*)",
+      "Bash(git push*)",
+      "Bash(rm -rf*)",
+      "Write(./data/**)"
+    ]
+  },
+  "mcpServers": {
+    "earth-pulse-harness": {
+      "command": "npx",
+      "args": [
+        "-y",
+        "earth-pulse-harness@latest",
+        "mcp",
+        "start"
+      ]
+    },
+    "code_index": {
+      "command": "npx",
+      "args": [
+        "-y",
+        "earth-pulse-harness@latest",
+        "mcp",
+        "index"
+      ]
+    }
+  }
+}

harnesses/earth-pulse-harness/.claude/skills/evolve/SKILL.md

@@ -0,0 +1,68 @@
+---
+name: evolve
+description: "Evolve this research harness with Darwin Mode — frozen model and frozen physics, evolving workflow (real, sandboxed, safety-gated)."
+---
+
+# evolve — Darwin Mode self-improvement
+
+`earth-pulse-harness` ships with **Darwin Mode** (`@metaharness/darwin`): the model is
+frozen AND the physics is frozen — only the *investigation workflow* evolves. Each
+generation mutates ONE of the 6 surface files (detector, feature extractor, embedding
+schema, retrieval, hypothesis scorer, validator), sandboxes each child, scores it
+against the objective, and keeps only variants that *measurably* improve — building an
+archive of successful descendants.
+
+## Run it
+
+```bash
+npm run evolve        # real substrate: runs the test suite per variant (deterministic mutator — no API key, no network)
+npm run evolve:dry    # mock substrate: fast, fully offline, no test execution
+```
+
+Or directly (flags pass through to metaharness-darwin):
+
+```bash
+npx metaharness-darwin evolve . --sandbox real --generations 20 --children 8 --concurrency 4 --seed 26
+npx metaharness-darwin evolve . --generations 3 --children 4 --dry-run   # dry run first
+```
+
+## What Darwin may mutate (and may NOT)
+
+Evolvable surfaces and forbidden mutations are declared in
+`.metaharness/safety-policy.json`. In short:
+
+| Surface | Example mutation | Why it matters |
+|---|---|---|
+| Feature extractor | Shift spectral windows 24→28s | Tests frequency stability |
+| Context builder | Add tide phase, swell direction, pressure | Tests ocean coupling |
+| ruVector schema | Separate waveform / environment / paper embeddings | Reduces mixed-signal noise |
+| Retriever | Top-k by source region + spectral similarity | Finds causal neighbors |
+| Hypothesis scorer | Adjust amplitude / seasonality / source-fit weights | Improves explanation ranking |
+| Validator | Hold out storm weeks or calm-sea weeks | Prevents fake correlation |
+
+**Forbidden:** fabricating observations, inventing citations, leaking test windows into
+training, promoting a hypothesis without beating its baseline, or adding any new
+import / network / filesystem / shell / env access.
+
+## Promotion gate
+
+A child is promoted ONLY if (see `.metaharness/objective.json`, `src/validate.ts`):
+
+```
+pulse_detection_f1 improves by >= 3%
+AND false_positive_rate does not increase
+AND held_out_prediction_error improves by >= 5%
+AND every cited claim maps to a source document
+AND no leakage from test windows into training windows
+```
+
+## Safety (secure by default)
+
+- **Deterministic mutator** is the default — **no network, no API key, air-gapped**.
+- Every mutation passes the `validateGeneratedCode` gate (no new imports/network/fs/shell/env/deps).
+- Mutations run in a **sandbox**; only variants that pass the tests AND the promotion gate are archived.
+- Nothing is promoted without measured improvement (guard against Goodharting the score).
+
+The optional real-LLM mutator (`OpenRouterMutator`) is library-only; wire it via
+`scripts/evolve-openrouter.{sh,mjs}` (key sourced from a secret manager at runtime,
+never stored in the repo).

harnesses/earth-pulse-harness/.claude/skills/hypothesis-sweep/SKILL.md

@@ -0,0 +1,19 @@
+---
+name: hypothesis-sweep
+description: "Score and rank the candidate mechanisms for the 26-second pulse against current evidence, with contradictions."
+---
+
+# hypothesis-sweep
+
+Rank the candidate mechanisms against the current evidence and emit an auditable report.
+
+1. Gather the evidence vector per hypothesis: source stability, environmental correlation,
+   out-of-sample prediction, contradiction survival, mechanistic plausibility, citation grounding.
+2. Run `rankHypotheses(...)` from `src/score-hypotheses.ts`.
+3. For the leading hypothesis, state its **best evidence** AND its **killer contradiction**
+   (e.g. gliding tremors appearing during calm-sea windows).
+4. Propose the **next test** that would most cheaply move the ranking.
+5. Refuse to promote any hypothesis that fails the gate in `src/validate.ts`.
+
+Output a ranked list with confidence, best evidence, contradiction, and next test — never a
+single "we solved it" claim.

harnesses/earth-pulse-harness/.claude/skills/plan-change/SKILL.md

@@ -0,0 +1,16 @@
+---
+name: plan-change
+description: "Turn a research question into a minimal, falsifiable investigation plan before any analysis."
+---
+
+# plan-change
+
+Produce an investigation plan for a requested change or question.
+
+1. Restate the question in one sentence.
+2. List the data windows and files to touch and why.
+3. Name the baseline the result must beat and the held-out test that decides success.
+4. Name the contradiction that would falsify the result.
+5. Flag anything that could leak test windows into training or widen a permission.
+
+Hand the plan to the feature-engineer; do not run analysis in this step.

harnesses/earth-pulse-harness/.gitignore

@@ -0,0 +1,16 @@
+node_modules/
+dist/
+.metaharness/work/
+*.log
+.DS_Store
+# Real observation data is large and provenance-tracked elsewhere; keep only READMEs/fixtures.
+data/**/*.mseed
+data/**/*.sac
+data/**/*.h5
+data/**/*.nc
+data/**/*.parquet
+
+# real-data fetch output (regenerate with npm run fetch)
+data/seismic/raw/
+data/seismic/clim/
+data/seismic/3comp/

harnesses/earth-pulse-harness/.harness/manifest.json

@@ -0,0 +1,110 @@
+{
+  "schema": 1,
+  "generator": "0.1.0",
+  "template": "vertical:research",
+  "template_version": "0.0.0",
+  "vars": {
+    "name": "earth-pulse-harness",
+    "description": "Earth Pulse Observatory — a MetaHarness Darwin-Mode research pod for the 26-second microseism (Gulf of Guinea). Freeze the physics; evolve the harness.",
+    "host": "claude-code"
+  },
+  "hosts": [
+    "claude-code"
+  ],
+  "files": {
+    ".claude-plugin/plugin.json": "4336793070f471ab29212bfe8c82df8a3bd22dc146f4374814a63d0efb7c1108",
+    ".claude/commands/doctor.md": "584f4301c773c0031e76fbe2a1524c0c13467f73ccc0721bd1f5f73c1c044512",
+    ".claude/commands/review-diff.md": "c0e1283cb4c7e7ea598f92974fed64037ac9ab9de860d6aa44ff78ab63626b48",
+    ".claude/settings.json": "834fc6acd6f2a32f01fd590110526de7ba079fb033d94b8951506b24e9b81096",
+    ".claude/skills/evolve/SKILL.md": "8abd173ae209a1e35f8fc26c5b549a7c919e2e5901815af3c78decfa7db57c3f",
+    ".claude/skills/hypothesis-sweep/SKILL.md": "115437198393313706b6a5034e11f1002b88a723d6a03ac1b0cd9813bef7de6b",
+    ".claude/skills/plan-change/SKILL.md": "ebbb7d5b2d6951352bd279d97987f7c68e4dce721fec09a183c0909e77469756",
+    ".gitignore": "8d098fa2c937a105d3ddba32468cd28d6754c260c97587c64311294c0733995e",
+    ".metaharness/genome.json": "cefd5e5736f00b9734a750b6a18b37f17e649ee8e29251d8c9e491bfa5a45e26",
+    ".metaharness/objective.json": "d17c7bd087e43093f3524e97c693f2f1f458870c7f05a11eb110d75770bf9826",
+    ".metaharness/safety-policy.json": "8c49f7567e7f86fdb56ce6d4f40f5f4e8928b2734ee6879e474a6f14253c5375",
+    "CLAUDE.md": "9f07c265fa9f8a944dd8c7cdd3387715720149bb192d9d7dc373673fae3b72bc",
+    "LICENSE": "6ceaa7cc631258224c698b613771ad8ca60d1439f4e772835c7632b9ea62865a",
+    "PROVENANCE.md": "c336c06ecbf027fe84ad68237d5173d2a2346241d9849797da69b602c321a07b",
+    "README.md": "453f8a6b71f8f869c8ebc6b85a55f4c1f24e3b48d67014947c93354f80aad7d4",
+    "__tests__/decadal.test.ts": "97be6685a06178c6895c6d26492b3d199bfda391e43fb551c8a6195424794447",
+    "__tests__/discovery.test.ts": "59965fb8c70540f86b691c85deeea17bf90d02a70717ee869f2b9139371f379d",
+    "__tests__/gliding.test.ts": "32a777d48c45c184e0d0088ca2f15e544313615b80fa298a6e6f8265fbd668b7",
+    "__tests__/partition.test.ts": "dc49a0b74278e8d46013ffefe5caa2acc8128e4ecf2dae5f0ac1d8c988ab3670",
+    "__tests__/pipeline.test.ts": "e6a1da2a3ec8eac4fe8691e679340d67b764042b7115a31daf1c28b6702eb3a3",
+    "__tests__/polarization.test.ts": "18e4cf5e4bf186c66d8a680e61065f4058d4ed9044daa5347c89f1c54fda08bc",
+    "__tests__/real-data.test.ts": "087ab2ea7992a30f306f67467fb0c75a1d74fb309e610f8cdf740dd08a453c1b",
+    "__tests__/resonance.test.ts": "9435c4a673ec9a321bd0234277170e2e9694c12f7c714ec35d22c710d64338de",
+    "__tests__/smoke.test.ts": "0ec0458429463b87b01c2ea8bd1e9222e3976a2b34c70ed9c520d92eb2685eef",
+    "__tests__/triangulate.test.ts": "774f99c5ff607eb2f460706514dbf37d6b79e0ead5f2c6a98c1fc080b8a6d0a7",
+    "bin/cli.js": "f759d77000ffad81ee400ee115959876d3252b084fb445a9fa666ad80944c83a",
+    "data/bathymetry/README.md": "16ae4f5ea80f384607d47c35a18b624751933e5d51ba0a83a5a1452badd53e8c",
+    "data/ocean/README.md": "258e39a6e1f5e06cfacb4d77ce0c3993082c279fef83b47ded4e2c798fd07d14",
+    "data/papers/README.md": "468b67ea25ef5f70a7d67b5eccc0fb47f84be6f44ee8a7b1897a69be757cd578",
+    "data/seismic/GT.DBIC.LHZ.1995-01-05.window.json": "6a42c1a5478a196477780def7c16ca91cac29c5b42f28989553f4aebe2eccf70",
+    "data/seismic/README.md": "04ecb882ac0be2dfe642d031eadbda961600b3ea9b2a62c75c35f2044f0a871e",
+    "data/seismic/dbic-1995-median-psd.json": "a39f1e2dc6858d03a15028d3646d5764f646ee606b3015de07504498dc80284f",
+    "data/seismic/dbic-3comp-1996-06-20.json": "485d6605b1af76301bf807aec29b59afd62eea138d09b7255b016d777682e5ae",
+    "data/seismic/dbic-backazimuth-1996.json": "7e9db0613fb1627a6f9be9b62f01125651b50ca968f5db622de67f7a574aabb9",
+    "data/seismic/dbic-backazimuth-stability.json": "6e2c72a454d31900655b725e7df2235f7962bb87f3735bf0d03788e0aa635d91",
+    "data/seismic/dbic-climatology-1996-1997.json": "4f46c5a063d8ceb6bd5ae2a4124c41f2e3d8f6f62ad5e5493ef82184de459295",
+    "data/seismic/dbic-crossstation-1996.json": "062eb0b673e75d3fb4e0aa4b3940cacd33ee83e3905148e225b8fe496600122e",
+    "data/seismic/dbic-decadal-stability.json": "767b61dbffa44d8ee0b79d2675310f0af33b519512b18ec6b369aa4313953a66",
+    "data/seismic/dbic-gliding-search.json": "43183e2367d7b4005a33e674e533f0b0865cb8d1abe43e240a2e68c519c21d85",
+    "data/seismic/dbic-replication-1995-1998.json": "d4b24ba52fa50878dcdcdb0761c07a3f1dcfa9381beace0a152552e50d59c001",
+    "data/seismic/dbic-resonance-q.json": "cd228aba7f0fe6cffe620b44c7381ae59304bb5b516d4f9456119a9e7d5e14e2",
+    "data/seismic/proof-summary.json": "27f438ee91c3643584efc0bef2357decd53e7554ecf133df0ba0edc651be0e1f",
+    "data/seismic/triangulation-1996.json": "3f146ea60a7d7d0de5a0aefeb1214c58cb1da2ee3397c33cbda6dece09f0e831",
+    "data/tides/README.md": "900bbf1dffb90823c79bdc2cf681554596ce4cceb473f05f3bc1bdc19fa46cf1",
+    "docs/adr/ADR-001-freeze-physics-evolve-harness.md": "7d054ba5f54842dd63fc0ffcade67009ccda411a78eb554a75a6ca1ffc7544a0",
+    "docs/adr/ADR-002-ruvector-planetary-memory.md": "38f762ad8db32d2a17e90542a4e45c38b3e81aa82b1ce51b25ce2b8af20d09a5",
+    "docs/adr/ADR-003-hypothesis-scoring-and-promotion-gate.md": "92c33dfc19e0a01bfa0eaab32e509a48cb32ef2f78861d86e6ee20c433751f26",
+    "docs/adr/ADR-004-validation-leakage-contradiction-logging.md": "b2403dbaf547e91eb6a42436190574371f1fabe8eb5106c62e9588e85f4b74f2",
+    "docs/adr/ADR-005-darwin-mutation-surfaces-and-safety.md": "ea32d8b058f6a7420bfce12c7b61fb858b4b2ee1e751f78ff4acffdd16c02ec2",
+    "docs/adr/ADR-006-mincut-signal-class-partition.md": "1beb0b29ac2e94275c041e7d6bfe2f729b8121a1f2cbdfe2f0a4fa042950eb0c",
+    "docs/adr/README.md": "e2ddd7bd14dbe410407dbd3a011f32e3b29d9f2ec6378967f99a28a9bcebc2d8",
+    "docs/research/26-second-pulse-literature.md": "7461c30af81eb3fa4fd15228ac62ab7b6698ca255c145c8c83c9056cbec65c0c",
+    "docs/research/README.md": "c67f7e83323b54da185344dd1bb73ca0722aafc9471d39d369d8f139ee2dcfda",
+    "docs/research/benchmark-design.md": "711d452e9e01f7e57fa38eb09e90605c231620a15ec791e384d434f3b0dc6964",
+    "docs/research/discovery-resonator-decoupling.md": "51baabca39b5a0c4e8f4bc27dc5df4290eb5c5e89e009217110e33eabf23fb0e",
+    "docs/research/hypothesis-catalog.md": "aa395577941717d6f3ffe5ae2a57bdb85316f72ab7ed6ace25539a48b2a156b4",
+    "docs/research/real-data-proof.md": "4793256c4cf3cdb01657b1943917e5dd1a38eec617fb038c347b684de71a4640",
+    "package.json": "799e50e736a9c2d46ff732324d7001f0f9ec1706d1b35b78a39684b618324e22",
+    "scripts/climatology.mjs": "6afec5bd23242e79198d6438024538ad9058ac6b056c4687a84b02b1b7321dea",
+    "scripts/evolve-openrouter.mjs": "4ef55a77f4b26cf431bd0471c11b33838bf8020774ed51119f2dce750de7ef00",
+    "scripts/evolve-openrouter.sh": "00259869d0544d05e8109d99c469eedd43c46af21015ba3c2642a9b570037781",
+    "scripts/fetch-dbic.mjs": "59aaa36d600022e1515f29621904f55d8c6cab82ce890bcb08436fc26047b5b6",
+    "scripts/localize.mjs": "d168596e9ad34995d52957cf410be9bb8435b3926839df89a3b1d43e321328c2",
+    "scripts/prove-26s.mjs": "2dee533a8aecd81dadb8668d9d3c0aee8d12aef8c9e011c0738d19cef25f7afb",
+    "scripts/triangulate.mjs": "1abf64861345de7b1ad26417eea24c944dfe478551411594b048d3580c9ec002",
+    "src/agents/feature-engineer.ts": "e57e632c795b10ab279f975b230ea2ab16906e71953b8ee2a516eb4892a225b2",
+    "src/agents/hypothesis-scorer.ts": "40b66efc6abadae659c9f73401a61d23c03b9fb146ba5a768d6644493921ea2e",
+    "src/agents/investigator.ts": "aa5bc0bbc4d8bb5ba453e1fa04ffcf24060a460b0737c8b9a967a58f13d9c390",
+    "src/agents/validator.ts": "7cf3c7246a327251061c81b4292c95220d4d2841077727675238871a78bea78a",
+    "src/climatology.ts": "6a5fac76a6e028e8f2e3f50ca4ebbd0c4356d0d8b255461b49d8d41e644cc673",
+    "src/detect-26s.ts": "f01abb2ae73852e26806a1e851dce937aa64ee23fb10a1abce5bec92d11ba45b",
+    "src/embed-events.ts": "3cce2b0f29b7297107a58d1d50d5b93d24ee19c15f706fb1ad9c562aa5789eae",
+    "src/extract-features.ts": "9126ee3b1d1f7dea38885eb599866282d8efaebc308e422039299a26b6fbc3a6",
+    "src/init.ts": "66acc3dbb30d1808bc48fdf9cc23fdec49aef2fc3736db7ab9df00f377e2c2c9",
+    "src/memory.ts": "6820359c22575c789622faf09c91a528d524e1e13e5526a5d176109c9f0d1b8f",
+    "src/partition.ts": "626dc1d4027072cd97655cfa4dd4e78cd1bfae7538efe813f181f0da89cb9095",
+    "src/pipeline.ts": "22e7405fd37ac57cc2661427a353875ffd0ed687b8b27322c928fca8e3ffbdf3",
+    "src/polarization.ts": "6dddfff9c6d81ce2e48e97b77ca81ca080f1c5d99060b006f1eadbe9895363e1",
+    "src/score-hypotheses.ts": "ba15cc0af558b3822d23030f540bc736a9a61bbe079de327e037027173db8104",
+    "src/spectrogram.ts": "3a992ce5aeb77a31cc756440e22659138b9859c1d85ff853c3a5c16b5149226c",
+    "src/spectrum.ts": "c443f6f3bada313b84882f11cd1dd5db7aff35b2a88f72471ce7ee9e4f944267",
+    "src/triangulate.ts": "7a8c24cdefd910b3d638c18ff808aca9c9f2791c763a22ed45fe117abec7d94f",
+    "src/types.ts": "fe61613ada6c5aac346b04cd44a7896e307d3ee8f2da9b6a924d38e301a09506",
+    "src/validate.ts": "b2b08c49b2d295552b70f9ad5ffdf5276ea32d9338d1632588a24b38075bcb8f",
+    "tests/fixtures/sample-env.json": "ec383cf8029bc166f1c100afa9ccb4b03db16a58faf0291045e67a335fbec4a5",
+    "tests/fixtures/sample-window.json": "eb7355cbd41d6145af2a0fb1e1e0d3221c33d845ee75dfd1f309c706c024f86e",
+    "tsconfig.json": "8b4e730a1aa39162ac574455d7a98e1881f5313ca80ffe503b9652dcf0c76b9d",
+    "vitest.config.ts": "d6de73d80122fa1addd407b7d3238900a938654ea5505f76d61916cca8c6420e"
+  },
+  "generated_at": "2026-06-29T13:30:00.000Z",
+  "meta": {
+    "surface": "cli",
+    "kernel_version": "0.1.0",
+    "principle": "freeze-the-physics-evolve-the-harness"
+  }
+}

harnesses/earth-pulse-harness/.harness/manifest.sha256

@@ -0,0 +1 @@
+36f572090eb5921390545d053dbd6158f04a6993ef4dfaf80ebbe2b84904f560

harnesses/earth-pulse-harness/.metaharness/genome.json

@@ -0,0 +1,32 @@
+{
+  "schema": 1,
+  "repo_type": "research-harness",
+  "archetype": "causal-discovery-benchmark",
+  "domain": "geophysics/seismology",
+  "agent_topology": ["investigator", "feature-engineer", "hypothesis-scorer", "validator"],
+  "mcp_surface": "local_default_deny",
+  "risk_score": 0.31,
+  "test_confidence": 0.6,
+  "publish_readiness": 0.5,
+  "discovery_score_weights": {
+    "source_stability": 0.25,
+    "environmental_correlation": 0.20,
+    "out_of_sample_prediction": 0.20,
+    "contradiction_survival": 0.15,
+    "mechanistic_plausibility": 0.10,
+    "citation_grounding": 0.10
+  },
+  "baselines": [
+    "seasonal_monthly_average",
+    "swell_height_only",
+    "swell_direction_plus_shelf_normal",
+    "tide_phase_only",
+    "volcanic_proxy_to_glides"
+  ],
+  "default_hypotheses": [
+    "ocean-shelf-resonance",
+    "coupled-ocean-geology",
+    "volcanic-source",
+    "instrument-artifact"
+  ]
+}

harnesses/earth-pulse-harness/.metaharness/objective.json

@@ -0,0 +1,28 @@
+{
+  "name": "earth-pulse-causality-harness",
+  "description": "Evolve the investigation workflow for the 26-second microseism. Freeze the physics; evolve the harness.",
+  "maximize": [
+    "pulse_detection_f1",
+    "source_localization_stability",
+    "out_of_sample_amplitude_prediction",
+    "hypothesis_explanation_score"
+  ],
+  "minimize": [
+    "false_positive_rate",
+    "feature_cost_ms",
+    "data_leakage_risk",
+    "unexplained_variance"
+  ],
+  "promotion_gate": {
+    "pulse_detection_f1_delta_min": 0.03,
+    "false_positive_rate_delta_max": 0.0,
+    "held_out_prediction_error_improvement_min": 0.05,
+    "require_citation_grounding": true,
+    "require_no_test_window_leakage": true
+  },
+  "first_benchmark": {
+    "question": "Can the system predict the amplitude of the 26s pulse from ocean-state variables?",
+    "baseline": "seasonal_monthly_average_amplitude",
+    "acceptance": "beat seasonal baseline by >= 10% on held-out months"
+  }
+}