Skip to content

SHOWCASE.md

Latest commit

 

History

History
76 lines (59 loc) · 3.58 KB

SHOWCASE.md

File metadata and controls

76 lines (59 loc) · 3.58 KB
Raw

Showcase: numbers, demo route, cases

中文导读: 本页集中展示 DeepGraph 的运行规模、现场演示路线和代表性案例。 所有数字均标注采集日期; 规模数字属于运行数据。

Scale (operational snapshot, 2026-08-06)

Collected live from /api/stats on the running deployment (port 8080):

  • Papers: 21,151 collected, 6,729 fully processed
  • Structured evidence: 28,700 claims, 156,194 results, 238 contradiction clusters detected
  • Knowledge graph: 248,441 entities, 735,913 relations, 5,051 taxonomy nodes, 32,252 paper-taxonomy assignments
  • Discovery: 11,936 insights, 110 deep insights (39 paradigm-level, 71 paper-idea-level), 25,652 mapped research opportunities
  • Execution: 113 experiment runs, 99 GPU jobs across 17 registered workers
  • Investment: 921M LLM tokens

These numbers grow continuously while ingestion and discovery run.

Demo route (15 minutes)

Open the dashboard (http://<host>:8080) and walk the loop end to end:

  1. Overview -- papers, graph, insights, runs and token totals at a glance.
  2. Explore Map -- pick an ML subfield node: node summary, its papers, and generated insights side by side.
  3. Evidence -- the method x dataset matrix; strong cells, weak cells, and white space (research opportunity) are immediately visible.
  4. Paper Ideas -- Tier 1 / Tier 2 deep insights with their gate status and experiment linkage.
  5. Experiments -- runs with their two status badges (operational and scientific), artifacts, and verdicts.
  6. Manuscripts -- generated PDFs/TeX, quality reports, and bundle status.
  7. Agent Office -- the big agents (extraction, graph, ideas, planning, execution, manuscripts, orchestration) and what each is doing right now.

Case: the accounting loop, closed end to end

On 2026-08-04 the full authorization -> execution -> settlement chain ran autonomously for the first time: a candidate passed the topic gate, received a scoped resource grant, executed, was settled token-by-token, and the unused reservation was refunded to its agenda budget automatically -- with every step leaving an auditable record (OutcomeRecord id=1). The pipeline that spends resources and the ledger that accounts for them agree by construction, not by convention.

Case: a paper-grade benchmark harness (CGGR plugin)

plugins/examples/cggr/ is a complete, self-contained experiment harness showing what "contracted experiment" means in practice:

  • A frozen benchmark contract pins datasets (MuSiQue-Ans, StrategyQA, 2WikiMultihopQA and a counterfactual stress split), a pinned Qwen instruct checkpoint, and seeds [0-4].
  • The runner emits machine-readable artifacts for every run: per-seed results, ablation tables, failure_cases.jsonl, and raw generations (plugins/examples/cggr/experiment_runner.py).
  • Statistics are computed, not asserted: paired permutation tests and 2,000-round bootstrap confidence intervals ship with the harness.

Case: a system that can say no

Manuscript generation is gated, and the gates are load-bearing:

  • A claim without quantitative evidence cannot appear in the abstract, introduction, or conclusion (agents/paper_completeness.py).
  • Fewer than three seeds, a missing baseline, or an absent significance test blocks the evidence audit (contracts/scientific_evidence.py).
  • A bundle that fails its quality gates is stamped DO_NOT_SUBMIT.md with a concrete repair list instead of being shipped (orchestrator/manuscript_watchdog.py).

For a research-automation system, knowing when not to publish is the feature that makes everything else trustworthy.