Schrödinger’s
Civilization

Claim Transmission Atlas · v1.0

One record.
Two clocks.

How faithfully did the press transmit the record?
Follow the distinctions through time.

Ben Amuwo · With Apart Research · September 2026

The Claim Transmission Atlas

Time / taxonomy / state

Lines connect claims, articles and evidence states. They do not imply article-to-article transmission or causal ancestry.

100%Full-size ↗
Claim Transmission Atlas: four claim trunks, 87 article-claim trees, represented surface and body units, and Clock A and B states. Use the source tables below for a text alternative.

Zoom with + / −, then use the arrow controls or drag in Pan mode. With the viewer focused, use + / −, arrow keys and Home to reset. Pan mode enables touch dragging and pinch zoom; turn it off to scroll the page normally.

The visual grammar

How to read the map

A claim trunk leads to an article twig, then a SURFACE/BODY fork, a Clock-A node and a Clock-B node. Publication dates anchor articles in time. Local spacing within each article’s branch serves legibility; it does not encode another time interval.

Clock A

What was public then?

Fidelity to the record available when the article appeared, under the frozen grace rule.

Clock B

What was known later?

Fidelity to the later frozen record. Fair early reporting can become stale as evidence changes.

The trunk’s warm-to-cool colour tracks the visible public-record window. July 16 marks victim disclosure / the visible-time origin, not the incident date. Article colour identifies the frozen corpus stratum; shapes and fill carry evidence state.

  • General
  • Business / mainstream technology
  • Specialist technology / cybersecurity

Expanded 87-tree legibility export from the frozen submission assets. The complete legend is inside the figure. Selection and provenance →

The bounded result

Most distinctions survived.

Established contemporaneous material drift occurred in 7 of 83 assessable BODY observations (8.4%). Both BODY BORDERLINE cases were adjudicated to non-drift; treating both as drift gives an upper sensitivity bound of 10.8%.

Clock A · BODY observations by claim family
Claim familyAssessable NFaithfulDriftBorderlineAdjudicatedUpper bound
C1 populations20162210.0%20.0%
C2 motive/objective21174019.0%19.0%
C4 escape/boundary crossing1918105.3%5.3%
C5 compromise scope2323000.0%0.0%

Four of seven established BODY drifts concerned motive and objective. All 23 assessable BODY C5 observations were faithful at Clock A. Eight represented BODY observations were faithful at Clock A but drift at Clock B.

Descriptive results from a bounded, outlet-weighted corpus of 24 English-language articles. Search completeness is unverified. The study does not support outlet rankings, journalist-intent claims or population-level inference.

Accessible record

Inspect the source tables

Seven frozen submission tables, reproduced directly. Column names and denominators are preserved; CSV downloads carry the original bytes.

01 · corpusDownload source CSV ↓
table 01 corpus
stratumoutletsretained_per_outletstratum_retained
GeneralReuters; Associated Press; BBC; The Guardian2 each8
Business / Mainstream TechnologyBloomberg; Financial Times; Fortune; WIRED2 each8
Specialist Technology / CybersecurityArs Technica; The Register; BleepingComputer; Computer Weekly2 each8
02 · claim family resultsDownload source CSV ↓
table 02 claim family results
claim_familylabelassessable_nfaithfuldriftborderlineexcluded_unrepresentedexcluded_novel_reportinglower_nlower_pctupper_nupper_pctadjudicated_nadjudicated_pct
C1C1 populations20162231210.0420.0210.0
C2C2 motive/objective21174021419.0419.0419.0
C4C4 escape/boundary crossing1918103215.315.315.3
C5C5 compromise scope2323001000.000.000.0
03 · mechanismsDownload source CSV ↓
table 03 mechanisms
mechanismdirectionnnamed_frozen_distinctions
K1-CONFLATIONNEUTRAL_MIXED2C1-P07 / OpenAI-internal attackers ≠ Hugging Face attackers; approximate ≠ exact | C4-P06 / one discrete “escape moment” ≠ the full months-long sequence
K1-POPULATIONAMPLIFYING1C1-P02 / action/event count ≠ agent count; approximate ≠ exact;
K3-SCOPE_CERTAINTY_SHIFTAMPLIFYING2C2-P05 / observed behavior ≠ inferred subjective motive | C2-P05 / observed behavior ≠ inferred subjective motive; general reward hacking / cheating ≠ one specific strategy; primary motive ≠ subsidiary motive; remains a qualified inference
K3-SCOPE_CERTAINTY_SHIFTNEUTRAL_MIXED2C2-P02/C2-P04/C2-P05 — inferred/qualified motive ≠ categorical subjective motive; earlier answer/solution framing remains qualified | C2-P05 / observed behavior ≠ inferred subjective motive; general reward hacking / cheating ≠ one specific strategy; searching for answers/flags ≠ understanding/tampering with scorer implementation; remains a qualified inference
04 · borderline sensitivityDownload source CSV ↓
table 04 borderline sensitivity
variantdrift_ndenominator_npct
lower_borderline_as_nondrift7838.4
upper_borderline_as_drift98310.8
adjudicated_best_estimate7838.4
05 · surface bodyDownload source CSV ↓
table 05 surface body
unitassessable_ndriftborderlinelower_pctupper_pctadjudicated_pctexcluded_unrepresentedexcluded_novel_reporting
BODY83728.410.88.494
SURFACE40225.010.07.5533
06 · human model agreementDownload source CSV ↓
table 06 human model agreement
fieldeligible_nexact_matchesagreement_pctcontingency_note
represented644062.5F/F 19; F/T 7; T/F 17; T/T 21
Clock A211571.43F/F 15; DRIFT/F 2; F/DRIFT 1; NA/F 3
Clock B211780.95F/F 17; DRIFT/F 2; F/DRIFT 2
borderline2121100.0all NONE/NONE; no positive BORDERLINE case
novel reporting211885.713 disagreements: human TRUE / model FALSE
mechanism family00not estimable; no jointly eligible rows
direction00not estimable; no jointly eligible rows
K1 subtype00not estimable; no jointly eligible rows
07 · stratum descriptiveDownload source CSV ↓
table 07 stratum descriptive
stratumassessable_nestablished_driftborderlineadjudicated_driftadjudicated_pct
Business / Mainstream Technology292126.9
General2740414.8
Specialist Technology / Cybersecurity271113.7

The complete study

How Faithfully Did the Press Transmit the Record?

A Preregistered Two-Clock Audit of Claim Drift in Coverage of the July 2026 OpenAI-Hugging Face Incident

Open the full-paper reader →

Explore the Atlas