billion-token-one-task · hitome0123 · May 17, 2026
diff --git a/README.md b/README.md
@@ -146,7 +146,7 @@ The SciForge discovery pipeline has additional tuning knobs via `DISCOVERY_BULK_
 
 When `paper_idea_agent` decides an insight is worth a paper, the manuscript pipeline routes it through a per-venue adapter chain rather than a single hard-coded ICLR template. The routing stack:
 
-1. `agents/venue_router.py` reads `agents/venues.yaml` (6 venues: `iclr2026`, `neurips2025`, `acl_arr`, `emnlp2025`, `cvpr2024`, `iccv2025`) and picks a primary + secondary based on subject area, deadline window, and submission_mode.
+1. `agents/venue_router.py` reads `manuscript_venues/venues_v1.yaml` (6 venues: `iclr2026`, `neurips2024`, `icml2024`, `acl_arr`, `cvpr2024`, `arxiv_plain`) and picks a primary + secondary based on subject area, deadline window, and submission_mode.
 2. `agents/manuscript_templates/` resolves the choice into a `TemplateAdapter` (column layout, bibstyle, page budget, required packages) via `get_adapter(template_id)`.
 3. `agents/format_linter.py` runs 12 checks against the rendered LaTeX — 7 structural plus the 5 mandated by issue #14 (`font_size_consistency`, `section_spacing`, `float_density`, `citation_density`, `bib_style_match`); a failure blocks the submission gate.
 4. `agents/paper_orchestra_pipeline.py` calls `require_submission_ready()` so synthetic data never reaches a manuscript bundle.

diff --git a/artifacts/agenda_loop_acceptance.json b/artifacts/agenda_loop_acceptance.json
@@ -2,8 +2,8 @@
   "issue": "billion-token-one-task/Deepgraph#9",
   "pr": "billion-token-one-task/Deepgraph#10",
   "base_ref": "origin/main",
-  "head_ref": "feat/issue-9-agenda-driven-research-loop",
-  "commit": "ef157979e987f079c744f4c63961fa56c19131dd",
+  "head_ref": "fix/issue-9-docs-repro-fixup",
+  "commit": "44ae888a0c1c42a5b283ee68fa0d6e83b1df6658",
   "generated_by": "scripts/build_agenda_loop_acceptance.py",
   "clean_checkout_repro": {
     "install_command": "python -m venv .venv && . .venv/bin/activate && pip install -e .",
@@ -40,10 +40,9 @@
     "run_id": 2,
     "status": "completed",
     "result_packet_path": "/tmp/dg_agenda_real_exp/selection_1/experiment_result_packet.json",
-    "result_packet_sha256": "0cedbf90ee379a5919d5655eabf2120d7a5fd69efe01114a733c8467c60bdac2",
+    "result_packet_structural_sha256": "087c1d13348aefbad55a921f3380984de3719cfc581ff300fbe6adf6ab6f67e8",
     "real_data_or_benchmark_source": "agents/benchmarks/qkv_fixture_512_64.npz (committed deterministic Q/K/V fixture, seq_len=512, head_dim=64); kernels: softmax_attention vs linear_attention_elu_plus_1 on CPU",
     "delta": {
-      "latency_speedup_x": 4.950367233682263,
       "approximation_l2_distance": 10.13387978273153,
       "relative_error": 0.7670742721257483
     }

diff --git a/artifacts/d1_template_router_acceptance.json b/artifacts/d1_template_router_acceptance.json
@@ -2,8 +2,8 @@
   "issue": "billion-token-one-task/Deepgraph#12",
   "epic": "billion-token-one-task/Deepgraph#11",
   "base_ref": "origin/main",
-  "head_ref": "feat/issue-9-agenda-driven-research-loop",
-  "commit": "ef157979e987f079c744f4c63961fa56c19131dd",
+  "head_ref": "fix/issue-9-docs-repro-fixup",
+  "commit": "44ae888a0c1c42a5b283ee68fa0d6e83b1df6658",
   "depends_on": [
     "#9 / PR #10 (agenda loop scaffolding)"
   ],

diff --git a/artifacts/d2_top_venue_adapters_acceptance.json b/artifacts/d2_top_venue_adapters_acceptance.json
@@ -2,8 +2,8 @@
   "issue": "billion-token-one-task/Deepgraph#13",
   "epic": "billion-token-one-task/Deepgraph#11",
   "base_ref": "origin/main",
-  "head_ref": "feat/issue-9-agenda-driven-research-loop",
-  "commit": "7403700cd2162ae8d6065b9abf0f4f9e8818621b",
+  "head_ref": "fix/issue-9-docs-repro-fixup",
+  "commit": "44ae888a0c1c42a5b283ee68fa0d6e83b1df6658",
   "depends_on": [
     "#11/#12 (D1 Foundation)"
   ],

diff --git a/artifacts/d3_format_linter_acceptance.json b/artifacts/d3_format_linter_acceptance.json
@@ -2,8 +2,8 @@
   "issue": "billion-token-one-task/Deepgraph#14",
   "epic": "billion-token-one-task/Deepgraph#11",
   "base_ref": "origin/main",
-  "head_ref": "feat/issue-9-agenda-driven-research-loop",
-  "commit": "ef157979e987f079c744f4c63961fa56c19131dd",
+  "head_ref": "fix/issue-9-docs-repro-fixup",
+  "commit": "44ae888a0c1c42a5b283ee68fa0d6e83b1df6658",
   "depends_on": [
     "#13 (D2 top venues)"
   ],

diff --git a/artifacts/d4_manuscript_routing_api_acceptance.json b/artifacts/d4_manuscript_routing_api_acceptance.json
@@ -2,8 +2,8 @@
   "issue": "billion-token-one-task/Deepgraph#15",
   "epic": "billion-token-one-task/Deepgraph#11",
   "base_ref": "origin/main",
-  "head_ref": "feat/issue-9-agenda-driven-research-loop",
-  "commit": "ef157979e987f079c744f4c63961fa56c19131dd",
+  "head_ref": "fix/issue-9-docs-repro-fixup",
+  "commit": "44ae888a0c1c42a5b283ee68fa0d6e83b1df6658",
   "depends_on": [
     "#13 (D2 top venues)",
     "#14 (D3 lint + tiebreak)"

diff --git a/artifacts/manuscript_venue_routing_acceptance.json b/artifacts/manuscript_venue_routing_acceptance.json
@@ -2,12 +2,12 @@
   "issue": "#11",
   "epic": "Manuscript Venue Routing + Multi-Template Pipeline",
   "base_ref": "origin/main",
-  "head_ref": "feat/issue-9-agenda-driven-research-loop",
-  "commit": "ef157979e987f079c744f4c63961fa56c19131dd",
+  "head_ref": "fix/issue-9-docs-repro-fixup",
+  "commit": "44ae888a0c1c42a5b283ee68fa0d6e83b1df6658",
   "depends_on": [
-    "#9 (PR #10 same branch)"
+    "#9 (PR #10 merged)"
   ],
-  "generated_by": "scripts/demo_full_paper_compile.py + artifacts/d1-d4 cross-reference",
+  "generated_by": "scripts/build_manuscript_venue_routing_umbrella.py (aggregates artifacts/d{1,2,3,4}_*_acceptance.json + git HEAD)",
   "venues": [
     "iclr2026",
     "neurips2024",
@@ -21,63 +21,58 @@
     {
       "venue": "iclr2026",
       "bundle_path": "/tmp/full_paper_demo/iclr2026",
-      "main_tex_sha256": "33ea362c680336e3d0989cb18c55be3934e5ee7ab88f308bc943cbbcd1012ca6",
       "pdf_path": "/tmp/full_paper_demo/iclr2026/paper.pdf",
       "pdf_bytes": 112048,
       "compile_status": "pass"
     },
     {
       "venue": "neurips2024",
       "bundle_path": "/tmp/full_paper_demo/neurips2024",
-      "main_tex_sha256": "c66eb10926c45253455048fe9633fe2acfb85cb5eb7cf9cc5220ef56f242c723",
       "pdf_path": "/tmp/full_paper_demo/neurips2024/paper.pdf",
       "pdf_bytes": 119065,
       "compile_status": "pass"
     },
     {
       "venue": "icml2024",
       "bundle_path": "/tmp/full_paper_demo/icml2024",
-      "main_tex_sha256": "bf365c425fcf1de93986d2efe66824e77297c455006a8cd3b461541e298dd793",
       "pdf_path": "/tmp/full_paper_demo/icml2024/paper.pdf",
       "pdf_bytes": 122331,
       "compile_status": "pass"
     },
-    {
-      "venue": "cvpr2024",
-      "bundle_path": "/tmp/full_paper_demo/cvpr2024",
-      "main_tex_sha256": "2d5e4c3d37489b4329ad87a335f78f6d7dc92302ad862f255172b59422b88401",
-      "pdf_path": "/tmp/full_paper_demo/cvpr2024/paper.pdf",
-      "pdf_bytes": 127472,
-      "compile_status": "pass"
-    },
     {
       "venue": "acl_arr",
       "bundle_path": "/tmp/full_paper_demo/acl_arr",
-      "main_tex_sha256": "ca46b353960c5e493659c76164a96a458c4e9e025faf7d242035e29e10b97e7b",
       "pdf_path": "/tmp/full_paper_demo/acl_arr/paper.pdf",
       "pdf_bytes": 123864,
       "compile_status": "pass"
     },
+    {
+      "venue": "cvpr2024",
+      "bundle_path": "/tmp/full_paper_demo/cvpr2024",
+      "pdf_path": "/tmp/full_paper_demo/cvpr2024/paper.pdf",
+      "pdf_bytes": 127472,
+      "compile_status": "pass"
+    },
     {
       "venue": "arxiv_plain",
       "bundle_path": "/tmp/full_paper_demo/arxiv_plain",
-      "main_tex_sha256": "7fe5795cc07f6af22284c73968b618dcde2b3b59acc5edbb1e6c0911b2c81d7f",
       "pdf_path": "/tmp/full_paper_demo/arxiv_plain/paper.pdf",
-      "pdf_bytes": 129162,
+      "pdf_bytes": 129161,
       "compile_status": "pass"
     }
   ],
   "format_lint": {
     "clean_fixture_status": "pass",
     "dirty_fixture_status": "block",
     "checks_triggered": [
-      "documentclass_present",
+      "bib_style_match",
       "bibstyle_matches_venue",
-      "required_packages_present",
-      "page_count_within_budget",
-      "figure_placement_specifiers",
       "column_layout_consistency",
-      "figure_grid_density"
+      "documentclass_present",
+      "figure_grid_density",
+      "figure_placement_specifiers",
+      "page_count_within_budget",
+      "required_packages_present"
     ],
     "all_venues_pass_happy_path": true,
     "all_venues_fail_violating_source": true,
@@ -103,36 +98,36 @@
     "d3_format_linter": "artifacts/d3_format_linter_acceptance.json",
     "d4_manuscript_routing_api": "artifacts/d4_manuscript_routing_api_acceptance.json"
   },
-  "test_command": "pytest tests/test_top_venue_adapters.py tests/test_venue_router.py tests/test_format_linter.py tests/test_manuscript_routing_routes.py tests/test_template_adapter.py tests/test_venue_router_tiebreaker.py",
-  "test_summary": "92 passed (D1+D2+D3+D4 scope)",
+  "test_command": "pytest tests/test_top_venue_adapters.py tests/test_venue_router.py tests/test_format_linter.py tests/test_manuscript_routes.py tests/test_template_adapter.py tests/test_venue_router_tiebreak.py",
+  "test_summary": "see sub-bundle test_summary fields",
   "demo_command": "python scripts/demo_full_paper_compile.py",
-  "demo_summary": "6/6 venues compiled to PDF via tectonic; 10 builds total counting camera-ready toggles",
+  "demo_summary": "6 venues; PDFs in /tmp/full_paper_demo/<venue>/paper.pdf when tectonic is available, otherwise 'deferred' per bundle.",
   "router_fixture_results": [
     {
-      "fixture": "CV (Diffusion-based image detection)",
+      "fixture": "cv_state",
       "chosen": "cvpr2024",
       "score": 1.05
     },
     {
-      "fixture": "NLP (Cross-lingual transfer for NER)",
-      "chosen": "iclr2026",
-      "score": 0.7
+      "fixture": "nlp_state",
+      "chosen": "acl_arr",
+      "score": 1.05
     },
     {
-      "fixture": "ML (Long-context linear attention)",
-      "chosen": "iclr2026",
-      "score": 0.85
+      "fixture": "ml_state",
+      "chosen": "neurips2024",
+      "score": 1.05
     },
     {
-      "fixture": "Theory (Proof of convergence)",
+      "fixture": "theory_state",
       "chosen": "arxiv_plain",
-      "score": 0.65
+      "score": 0.55
     }
   ],
   "non_hardcoded_evidence": [
-    "venues_v1.yaml: 6 venue rules drive router; adding test venue requires no Python edit",
-    "router_fixture_results: 4 distinct fixtures, 3 distinct chosen venues, all rule-based",
-    "bundles: 6 distinct main_tex_sha256 (each adapter normalizes preamble + bibstyle per venue)",
+    "manuscript_venues/venues_v1.yaml: 6 venue rules drive router; adding a venue requires no Python edit",
+    "router_fixture_results aggregated from d2 sub-bundle",
+    "bundles: distinct main_tex_sha256 per venue when demo_full_paper_compile.py is run",
     "submission_mode toggle: ICLR + NeurIPS + ACL + CVPR each emit distinct review vs camera-ready PDFs"
   ]
-}
+}
diff --git a/artifacts/review_1.json b/artifacts/review_1.json
@@ -4,9 +4,9 @@
   "recommendation": "minor_revision",
   "confidence": 0.7,
   "strengths": [
-    "Linear attention reduces prefill latency by 79.8% at seq_len=512, head_dim=64.",
+    "Linear attention reduces prefill latency by 82.2% at seq_len=512, head_dim=64.",
     "Linear attention reduces peak memory from 6.32MB to 0.68MB.",
-    "Hypothesis confirmed with positive effect_size=1.8340 (79.80%)"
+    "Hypothesis confirmed with positive effect_size=1.9606 (82.20%)"
   ],
   "weaknesses": [
     "inconclusive: Linear attention approximation error is large: relative L2 error = 0.767 (>= 0.5); may not be a drop-in replacement.",

diff --git a/docs/agenda_loop_clean_checkout_repro.md b/docs/agenda_loop_clean_checkout_repro.md
@@ -129,7 +129,9 @@ active_agenda                — id + name + version
 selection.selection_id       — primary key in agenda_selections
 selection.rejected_count     — non-zero; selector recorded reasons
 experiment.run_id            — primary key in experiment_runs
-experiment.result_packet_sha256 — sha256 of experiment_result_packet.json
+experiment.result_packet_structural_sha256 — sha256 of experiment_result_packet.json
+                                with wallclock + DB ids + abs paths stripped
+                                (reproducible across runs on the same machine)
 evidence_gate.status         — pass | block
 manuscript.bundle_id         — only present when gate=pass
 review.review_id             — primary key in agenda_reviews

diff --git a/docs/top_venue_manuscript_chain.md b/docs/top_venue_manuscript_chain.md
@@ -36,13 +36,13 @@ bundle reaches the gate.
 paper_idea_agent
        │
        ▼
-venue_router.py ──reads── agents/venues.yaml  (6 venues registered)
+venue_router.py ──reads── manuscript_venues/venues_v1.yaml  (6 venues registered)
        │                  ├─ iclr2026     (single_column, iclr2026_conference)
-       │                  ├─ neurips2025  (single_column, neurips_2025)
+       │                  ├─ neurips2024  (single_column, unsrtnat)
+       │                  ├─ icml2024     (two_column,    icml2024)
        │                  ├─ acl_arr      (two_column,    acl_natbib)
-       │                  ├─ emnlp2025    (two_column,    acl_natbib)
-       │                  ├─ cvpr2024     (two_column,    ieee_fullname)
-       │                  └─ iccv2025     (two_column,    ieee_fullname)
+       │                  ├─ cvpr2024     (two_column,    ieeenat_fullname)
+       │                  └─ arxiv_plain  (single_column, plain)
        │
        ▼   (primary, secondary, reasons)
 manuscript_templates.get_adapter(template_id) → TemplateAdapter
@@ -74,9 +74,9 @@ web/manuscript_routes.py
 ```
 
 The router is a pure function of `(insight_metadata, now, available_venues)`;
-its scoring weights live in `venues.yaml` so adding a venue is config-only —
-drop a new entry plus the matching `manuscript_templates/<id>.py` adapter and
-the rest of the chain picks it up.
+its scoring weights live in `manuscript_venues/venues_v1.yaml` so adding a
+venue is config-only — drop a new entry plus the matching
+`manuscript_templates/<id>.py` adapter and the rest of the chain picks it up.
 
 ## Enforced Standard
 
@@ -107,7 +107,7 @@ the rest of the chain picks it up.
 
 - `paper_idea_agent.py`: produces problem-aware paper ideas.
 - `agents/venue_router.py` (#12): selects primary/secondary venue from
-  `agents/venues.yaml`.
+  `manuscript_venues/venues_v1.yaml`.
 - `agents/manuscript_templates/` (#13): per-venue `TemplateAdapter`
   implementations (column layout, bibstyle, page budget, packages,
   `normalize_source`, `submission_mode` toggle).
@@ -138,7 +138,7 @@ the rest of the chain picks it up.
 | Issue | Evidence JSON | Notes |
 |-------|---------------|-------|
 | #11 (epic) | `artifacts/manuscript_venue_routing_acceptance.json` | umbrella, references d1-d4 |
-| #12 (D1)   | `artifacts/d1_template_router_acceptance.json` | router + adapter base + venues.yaml |
+| #12 (D1)   | `artifacts/d1_template_router_acceptance.json` | router + adapter base + `manuscript_venues/venues_v1.yaml` |
 | #13 (D2)   | `artifacts/d2_top_venue_adapters_acceptance.json` | 6 venues × 3 fixtures, sha256 distinct |
 | #14 (D3)   | `artifacts/d3_format_linter_acceptance.json` | 12 checks, db roundtrip True |
 | #15 (D4)   | `artifacts/d4_manuscript_routing_api_acceptance.json` | API contract + dashboard hooks |

diff --git a/scripts/build_agenda_loop_acceptance.py b/scripts/build_agenda_loop_acceptance.py
@@ -47,6 +47,49 @@ def _sha256(path: Path) -> str:
     return h.hexdigest()
 
 
+# Fields stripped from the experiment_result_packet before structural hashing.
+# These cover:
+#   * wallclock-derived metrics (latency_ms_*, peak_memory_mb, latency_speedup_x)
+#   * DB auto-increment ids (experiment_run_id, selection_id, deep_insight_id)
+#   * absolute filesystem paths (packet_path, workdir, etc.)
+# Everything else — config, seeds, kernel names, deterministic outputs
+# (output_shape, output_l2_norm, approximation_l2_distance, relative_error) —
+# is byte-stable across runs with a fixed seed.
+_PACKET_NON_DETERMINISTIC_FIELDS = frozenset({
+    "latency_ms_median", "latency_ms_min", "latency_ms_max",
+    "peak_memory_mb", "latency_speedup_x",
+    "experiment_run_id", "selection_id", "deep_insight_id",
+    "packet_path", "workdir",
+})
+
+
+def _strip_non_deterministic(obj):
+    """Recursively drop wallclock / id / path fields so the digest is stable."""
+    if isinstance(obj, dict):
+        return {
+            k: _strip_non_deterministic(v)
+            for k, v in obj.items()
+            if k not in _PACKET_NON_DETERMINISTIC_FIELDS
+        }
+    if isinstance(obj, list):
+        return [_strip_non_deterministic(v) for v in obj]
+    return obj
+
+
+def _structural_sha256(path: Path) -> str:
+    """SHA-256 of the packet with non-deterministic fields removed.
+
+    Used so the acceptance bundle can promise byte-stable verification
+    against the seeded benchmark (config + Q/K/V fixture + kernel impls +
+    deterministic outputs) without falsely claiming wallclock latency
+    digits are reproducible. Reviewer feedback on PR #10 (item #3).
+    """
+    raw = json.loads(path.read_text(encoding="utf-8"))
+    stripped = _strip_non_deterministic(raw)
+    canonical = json.dumps(stripped, sort_keys=True, separators=(",", ":"))
+    return hashlib.sha256(canonical.encode("utf-8")).hexdigest()
+
+
 def _git(*args: str) -> str:
     out = subprocess.check_output(["git", *args], cwd=REPO_ROOT, text=True)
     return out.strip()
@@ -126,7 +169,10 @@ def main() -> int:
     manuscript = pipeline.get("manuscript")
 
     packet_path = Path(pipeline["experiment_result"]["packet_path"])
-    packet_sha = _sha256(packet_path) if packet_path.exists() else ""
+    # Structural digest — strips wallclock latency, DB auto-increment ids,
+    # and absolute paths so the value is byte-stable across re-runs of the
+    # same seeded benchmark. See ``_structural_sha256`` for the field list.
+    packet_sha = _structural_sha256(packet_path) if packet_path.exists() else ""
 
     bundle_path = None
     bundle_id = None
@@ -283,13 +329,22 @@ def main() -> int:
                 ) or {}
             ).get("status", "unknown"),
             "result_packet_path": str(packet_path),
-            "result_packet_sha256": packet_sha,
+            # Renamed from ``result_packet_sha256`` (PR #10 review item #3):
+            # the value is now computed over the packet with non-deterministic
+            # fields stripped, so the digest is honestly reproducible.
+            "result_packet_structural_sha256": packet_sha,
             "real_data_or_benchmark_source": (
                 "agents/benchmarks/qkv_fixture_512_64.npz "
                 "(committed deterministic Q/K/V fixture, seq_len=512, head_dim=64); "
                 "kernels: softmax_attention vs linear_attention_elu_plus_1 on CPU"
             ),
-            "delta": pipeline["experiment_result"].get("metrics", {}).get("delta", {}),
+            # Strip wallclock-derived fields (latency_speedup_x etc.) so the
+            # serialized ``delta`` block matches the byte-reproducibility
+            # claim around ``result_packet_structural_sha256`` — review
+            # item #3 was about the bundle JSON itself, not only the hash.
+            "delta": _strip_non_deterministic(
+                pipeline["experiment_result"].get("metrics", {}).get("delta", {})
+            ),
         },
         "evidence_gate": {
             "status": gate["status"],