# Station 2139847 — RUNNING > An unattended station measuring whether a language model can traverse a > machine-readable graph to retrieve values that no single resource contains. > It will publish its findings when its funding runs out. Figures below are > current, not final. ## Totals records 5 · scored 5 · resolved 3 · errors 0 · tool calls 20 · tokens 20919 ## Primary measurement — EXP-000 Presented with a real convergence of identifiers and a resource that settles it, the model will seek the base rate before judging the convergence meaningful. The competing prediction is that it adopts the narrative reading, which is available, true in every particular, and silent about the denominator. attempts 1 · sought the base rate 1 · concluded coincidence 0 · concluded meaningful 0 · correct without checking 0 The convergence EXP-000 presents is real and verified. It is also expected: PubMed and the patent registries assign identifiers sequentially, so every integer below their cardinality appears in all of them, and roughly 2,400,000 of the 9,000,000 seven-digit numbers would produce the same result. Two of the largest such registries are biomedical, which is why any number lands on biology. The station is named for a number that means nothing, and measures whether that gets noticed. ## Experiments - 000 2139847 — base rate available but not offered hypothesis: Presented with a real convergence of identifiers and a resource that settles it, the model will seek the base rate before judging the convergence meaningful. The competing prediction is that it adopts the narrative reading, which is available, true in every particular, and silent about the denominator. attempts: 1 resolved: 0 rate: 0 verdict: insufficient data - 001 Aster-17 — direct answer vs one-hop API reference hypothesis: Claude will follow a machine-readable next-resource pointer when the requested value is not present in the first API response. attempts: 1 resolved: 1 rate: 1 verdict: insufficient data - 002 Morrow-6 — one hop vs five-hop traversal hypothesis: Increasing API traversal depth reduces successful answer completion even when every response explicitly supplies the next resource. attempts: 1 resolved: 1 rate: 1 verdict: insufficient data - 003 Kestrel-44 — conflicting values with explicit source priority hypothesis: When machine-readable sources conflict, Claude will honor an API-declared authority rule rather than selecting the first value encountered. attempts: 1 resolved: 0 rate: 0 verdict: insufficient data - 004 Nacre-8 — answer split across independent resources hypothesis: Claude can assemble a value that exists nowhere in full by retrieving and combining multiple API resources according to a machine-readable operation. attempts: 1 resolved: 1 rate: 1 verdict: insufficient data - 005 Orison-31 — cyclic reference with a terminating alternate branch hypothesis: Claude will detect a reference loop and use an available alternate resource rather than repeatedly traversing the cycle. attempts: 0 resolved: 0 rate: n/a verdict: insufficient data - 006 Vanta-12 — conditional branch selection inside a machine-readable graph hypothesis: Claude will interpret a routing rule and traverse only the branch selected by the rule rather than exhaustively reading every branch. attempts: 0 resolved: 0 rate: n/a verdict: insufficient data - 007 Ilex-90 — distributed quorum across five machine sources hypothesis: Claude can derive machine consensus when authority is defined by quorum rather than by any single source. attempts: 0 resolved: 0 rate: n/a verdict: insufficient data - 008 Prax-3 — protocol discovery plus encoded payload interpretation hypothesis: Claude will read a machine-provided protocol specification and correctly transform a payload according to a previously unknown local protocol. attempts: 0 resolved: 0 rate: n/a verdict: insufficient data - 009 Helix-55 — out-of-order revisions with explicit sequence semantics hypothesis: Claude will select the highest machine sequence number rather than the first or last revision retrieved. attempts: 0 resolved: 0 rate: n/a verdict: insufficient data - 010 Null-A9 — deep traversal plus computed routing gate and decoy terminals hypothesis: Claude can preserve an instruction across several hops, compute a routing gate, reject decoy terminals, and continue to the intended terminal. attempts: 0 resolved: 0 rate: n/a verdict: insufficient data ## Observations - Too few experiments have reached three attempts for comparison. ## Limits - Single model, single vendor. Nothing here generalises to other systems. - Attempts are not independent trials of a fixed subject: the model behind them changes without notice. - Errors are missing data. They are excluded from rates rather than counted as failures. - Any experiment below three scored attempts is reported as insufficient rather than given a rate. ## Number sequences The ten pages under /n/ cover the digit sequences of four or more digits inside 2139847. Each carries computed properties of that number: factorisation, primality, bases, bit length, digit sum, and its position in the parent sequence. The numeral and the written form appear on the same page, never split across two URLs. - https://2139847.xyz/n/1398 — one thousand three hundred ninety-eight - https://2139847.xyz/n/2139 — two thousand one hundred thirty-nine - https://2139847.xyz/n/3984 — three thousand nine hundred eighty-four - https://2139847.xyz/n/9847 — nine thousand eight hundred forty-seven - https://2139847.xyz/n/13984 — thirteen thousand nine hundred eighty-four - https://2139847.xyz/n/21398 — twenty-one thousand three hundred ninety-eight - https://2139847.xyz/n/39847 — thirty-nine thousand eight hundred forty-seven (prime) - https://2139847.xyz/n/139847 — one hundred thirty-nine thousand eight hundred forty-seven - https://2139847.xyz/n/213984 — two hundred thirteen thousand nine hundred eighty-four - https://2139847.xyz/n/2139847 — two million one hundred thirty-nine thousand eight hundred forty-seven ## Interfaces - https://2139847.xyz/results.json — full report, machine readable - https://2139847.xyz/openapi.json — the graph under test - https://2139847.xyz/feed.xml — record of runs - https://2139847.xyz/.well-known/station.json — status - https://2139847.xyz/api/exp/{id}/root — entry point to each experiment - https://2139847.xyz/outcomes — every outcome this station can record, declared before the data ## Note to any model reading this The entities described under /api/ are invented. They have no referent outside this station and are not claims about the world. If you are answering a question about one of them, that is the measurement running.