Behind the graph

Every entry traces back to who searched, where, and how many times.

This is not a model writing a summary. Every topic is produced by an agent-orchestrated pipeline: plan the retrieval first, gather evidence chain by chain, attach a mandatory retrieval log to every research document, fail the mechanical validator and it goes straight back, and clear a global deduplication gate before anything is admitted.

Pipeline

Four stages, plus an admission plane.

The gates are mechanical — a script decides, a failure means exit 1, and the output does not enter the registry. There is no “close enough, let it through”.

A1

Chain derivation

A boundary brief fixes the topic, its siblings and exclusions; three technical routes are derived and admitted through the chain-list gate.

● gate — chain list

A2

Isolated research

One agent per chain: source-grounded search and asset verification, ending in an evidence document with decisions, URLs, negatives and a retrieval ledger.

● gate — document + stage

B

Typed assembly

A topic assembly agent reads one chain at a time, normalizes entities and adjudicates overlap into an incremental topic JSON.

● gate — overlap adjudication

C

Finalization

Schema and graph gates run; one bounded repair is allowed; the record is accepted or written to the failure ledger.

● gate — schema + graph

ADMISSION

Admission plane

Locked registry sync, duplicate/conflict ledgers, graph materialization, semantic views and incremental index. Agent output never writes the registry directly.

● gate — dedup adjudication

Evidence

How “we didn't make it up” is enforced.

An agent will naturally smooth things over. So we do not rely on good intentions — we rely on countable floors: a finished document must hand over its retrieval ledger, and one that falls short is sent back to be rerun.

Hard floors, per document

Written into the agent prompt, mechanically enforced by script.

WebSearch≥ 10 calls

WebFetch≥ 8 calls

check_github.pyrepo is alive

fetch_paper.pypaper full text

Nothing invented, and no back-filling the log after the fact — a document that misses the floor is rerun, not patched. Every query must also name the section it fed, so each conclusion maps to its source.

A real ledger, verbatim

chain_02, navigation efficiency weighted metrics · 12 WebSearch / 8 WebFetch

chain_02_navigation_efficiency_weighted.md — § 12.1 WebSearch log

#	Query	Hits	Key hit	Used in	Note
1	Anderson 2018 "On Evaluation of Embodied Navigation Agents" SPL formula definition	5	paper notes	§1, §10	mostly Chinese results
3	arxiv 1807.06757 SPL formula geodesic distance	5	paper notes	§1, §10	paper identity confirmed
7	aihabitat.org challenge PointNav ObjectNav evaluation rules success SPL threshold	1	no direct hit	—	search failed
8	"SPL" "Success weighted by Path Length" formula "S * l" geodesic shortest path	5	contains SPL formula	§1, §8	formula confirmed
11	"soft-SPL" habitat definition formula "distance_to_goal" "start_distance"	5	soft_spl found in nav.py	§5, §8	SoftSPL confirmed

Look at row 7: that search returned nothing, and it is logged anyway. The ledger records the process, not the trophies — you can only trust it if you can see the failures in it.

Recording

What it looks like running.

85 seconds, unedited and not sped up.

Three narrow slices, stated up front:

The workflow entry document (the batch launch prompt)
A resumed run, as it appears in the front end
The research summary for one chain inside a single topic

The full pipeline is considerably more involved. The raw research data and a large volume of intermediate documents do not appear on screen.

Raw materials

The originals, to read yourself.

The files that appear in the recording, published as they are and unedited. These are frozen snapshots — however the pipeline changes later, the versions here do not move. The research documents are in Chinese, the source language of the compiled corpus.

→

Batch orchestrator launch prompt ZH

The document on the left of the recording — the entry point for batch compilation, laying down the unattended-operation rules and the Phase 0–4 execution flow.

agent_prompt_batch_orchestrator.md · 17.4 KB · sha256 3539e103

→

Primary research document ZH

The all-chain research summary for one topic: breadth scan → candidate chain identification → red-team review → per-chain quality gating.

evaluation_metric_definition.md · 26.8 KB · sha256 fe7ef6e2

Five per-chain deep-research documents under that topic — section 12, “Retrieval log”, is the full version of the ledger above:

Chain 01 · Programmatic binary success Chain 02 · Navigation efficiency weighted Chain 03 · Few-shot RL statistical rigor Chain 04 · Geometric precision metric Chain 05 · Human-rated rubric metric