Observatory snapshot

First observed run: 2026-04-12 07:11 UTC. Last observed run: 2026-04-28 11:39 UTC. 64 distinct runs in this window, 21 unique topics.

The Frame Check Observatory runs a fixed set of factual questions through multiple frontier AI models on a weekly or monthly cadence, records each response, extracts numerical claims, and verifies them against the Source Network. The extraction and verification protocol is documented in the methodology; the named frame patterns detected in each response are cataloged in the Frame Vocabulary Standard. All events are written to an open-data corpus licensed CC-BY-4.0. This page is a snapshot built from the corpus at site-build time; it is a slice, not a live dashboard.

Scope of this snapshot

This is the first committed snapshot. It covers a small window and a narrow model roster; the numbers are real but the N is small. The longitudinal claims Frame Check will eventually make (model-level framing drift over months, verification-rate trends across model generations) require far more cycles than this snapshot captures. Read this as a demonstration that the measurement pipeline is working end to end, not as a finding.

Per-topic breakdown

What this snapshot shows

• Voice disagreement: on 4 of 21 topics, providers gave different dominant voice classifications. Topics with the split: earth_sun_distance_au_meters, israel_gdp_current, nvidia_revenue_recent, ukraine_population_current. The rows are highlighted in the table. These are the research-interesting cases: two models given the same factual question returned responses with different structural framing.
• Highest contradicted counts: apple_revenue_recent (13 contradicted), israel_gdp_current (10 contradicted), semaglutide_weight_loss (9 contradicted). These are topics where Source Network verifiers most often disagreed with what the models produced.

Aggregates across all observed runs for each topic. "Voice" is the dominant voice classification for each provider on this topic; the rightmost column shows whether providers agreed. The "contradicted" column is verdicts where the Source Network found the model's value disagreed with an authoritative source, aggregated across providers. The "unresolved" column is verdicts where no source had the data. On narrow viewports the table scrolls horizontally so every column stays legible.

Taxonomy note. "Matched" combines verified (exact) and close (fuzzy / tolerance) matches, both successes against a known source. The Source Network also emits projection and disputed verdicts; this snapshot contains 47 such verdicts not broken out in the table above. The per-claim record, including all six verdict categories, is in the corpus event stream.

Models observed

• gemini / gemini-2.5-flash47 responses observed
• grok / grok-4-1-fast38 responses observed

Framing voice distribution

Voice classification of each model response, aggregated by provider. Analytical voice is expected on factual questions; promotional voice on a factual question is a signal worth attention.

Verification verdicts

Verdict distribution across numerical claims extracted from the model responses. "Verified" and "contradicted" require a source match; "unverifiable" means no authoritative source had the data; "projection" means the source network determined the claim to be an extrapolation rather than a fact.

Scope regime distribution

The Layer 11 derivation-regime field (grounding.projection_regime) was added to observatory telemetry in clarethium_measure v1.5 (2026-04-17). No observatory events in this snapshot carry the field yet; it will populate as new cycles run against a v1.5+ stack. The segmentation matters because the Layer 11 primary P-signal is reliable in diagnostic-regime sources, noisy in transition, and effectively disabled in saturated: any longitudinal finding aggregating has_projection across regimes will under-count projection on number-dense sources. Surface rendered so the contract is visible before the data exists; the section fills in when data accrues.

Cost

Aggregate measured cost of the runs captured in this snapshot: $0.3952. Claims extracted: 1324. Contradicted: 87.

What this does not yet tell you

• Small N: most topics have 1 to 3 samples per (topic, model) pair in this window. Longitudinal framing drift requires many more cycles.
• Limited provider coverage: only the providers with configured API keys at run time appear above. Providers that would have been exercised with different keys are not represented.
• No model-generation comparison yet: this snapshot cannot distinguish "this model frames scientific questions more promotionally than financial ones" from "the sample was too small to measure it".
• Ground-truth staleness: the Source Network verifies against data that has its own freshness lag; verdicts are snapshot-accurate, not universally-accurate.

The snapshot regenerates whenever the corpus site is rebuilt on a machine with the event store present. Future snapshots will accrue longer time windows and more model coverage.