Arctic Tracker / Pipeline

How a number gets into this paper

Seven steps in two stages. Stage 1 is deterministic: every value is a query result and no model touches it. Stage 2 is LLM-assisted, and every output it produces is checked back against Stage 1 before it is published or cited. The boundary between them is the point at which the corpus freezes, and it has a name: G3.

How this page was produced

Read from deposit/pipeline.json, which is derived from the manuscript's locked step table and from Supplementary Table 4 as deposited. Nothing on this page is written here; the step names, the gates, the tools and the audit-question bindings are all fields in that file, and the two diagrams are generated from it by deposit/build_pipeline_diagram.py.

The record carries the SHA-256 of the Supplementary Table 4 it was built from, so the mapping drawn below can be checked against the mapping deposited rather than trusted: 02dba91223918b7c…

Stage 1: deterministic

The LLM TOOL lane runs the width of this diagram and is empty. That is the claim, drawn rather than asserted: no model produces any reported value in Stage 1.

Stage 1: deterministic 01 IDENTIFY 02 COMPILE 03 INTEGRATE EXTERNAL SOURCES DETERMINISTIC PIPELINE LLM TOOL HUMAN GATE corpus frozen SRC CITES + CMS external register custodian record WB PIPE S1 Species list Q1, Q2 deterministic query WB DB GATE G1 author-verified reconciled to source DB SRC IUCN Red List external register custodian record DB PIPE S2 Status compile Q7 deterministic query DB DB GATE G2 author-verified reconciled to source DB SRC CITES Trade DB external register custodian record DB PIPE S3 Trade integrate Q3, Q5, Q6 deterministic query DB TB GATE G3 author-verified reconciled to source TB STEPS 01 Identify 02 Compile 03 Integrate DATA DB Source dataset TB Frozen table WB Public register FLOW Data hand-off Corpus frozen, Stage 1 ends

Counted from the same record the diagram is drawn from: all 3 Stage 1 steps name no model. Every row of the LLM tool column below reads None. That is the empty lane, written as a column.

StepWhat it doesLLM toolHuman gateAudit question
S1Species & listing identificationNone (deterministic query)G1Q1, Q2
S2Conservation-status compilationNone (deterministic query)G2Q7
S3Trade-record integrationNone (deterministic query)G3Q3, Q5, Q6

Stage 2: LLM-assisted

Each step hands its output to a human gate. G6 is the one that carries the weight: the model summarises verified metrics, and the status-trajectory scoring is applied by the authors under the Table 3 rule, not by the model.

Here it is the EXTERNAL SOURCES lane that runs the width of the diagram and is empty, and that is a claim too: the corpus freezes at G3, so nothing enters Stage 2 from outside. Every value it works on is already in the frozen corpus, which is why the only box in the deterministic lane is that corpus itself.

Stage 2: LLM-assisted 04 RETRIEVE 05 SYNTHESISE 06 CLASSIFY 07 PUBLISH EXTERNAL SOURCES DETERMINISTIC PIPELINE LLM TOOL HUMAN GATE scored by authors LLM Scite model-assisted LLM tool TB FL GATE G4 Q3, Q4 author-checked FL FL LLM Gemini model-assisted LLM tool FL FL GATE G5 Q3, Q4 author-checked FL FL LLM Summary only summary only LLM tool FL FL GATE G6 Q3, Q4, Q7 author-scored FL TB LLM None in figure path summary only walled off FL FL GATE G7 no audit question author-checked FL FL PIPE Frozen corpus Stage 1 output corpus frozen at G3 TB STEPS 04 Retrieve 05 Synthesise 06 Classify 07 Publish DATA FL Prose / report TB Frozen table FLOW Data hand-off Scored by the authors, not the model Published to the platform (G6 to G7)
StepWhat it doesLLM toolHuman gateAudit question
S4Literature retrievalScite — source discovery/summarisationG4Q3, Q4
S5Knowledge synthesisGemini — structured extraction/synthesis to proseG5Q3, Q4
S6Status-trajectory classification & implementation auditThe LLM summarises verified metrics only; scoring is author-appliedG6Q3, Q4, Q7
S7Integration & platform publicationNone in figure-generating path (the LLM writes prose only)G7none bound