What can a
machine trust?

Question

How can humans, agents and machines reach verifiable understanding of the same uncertain state?

Hypothesis

Task-specific representations with an inspectable evidence chain may reduce misunderstanding and over-trust.

Method / prototype

A planned 30-part research agenda, with the eight original public protocols retained below. Compare representations against matched baselines and preserve question, hypothesis, method, observation and Systems implication.

Observation

The public register contains eight initial protocols and the Fog of Probability synthetic prototype. A protocol or prototype alone is not a completed participant study. The full 30-part itemization was not recovered from the inspected canonical records; no missing study titles or outcomes are invented here.

Systems implication

Promote an interaction only after an inspectable artifact and evaluation establish what improved, for whom, and under which limits.

Next gate

Recover the complete canonical itemization and attach an artifact and evaluation record to each executed exploration.

Question

How quickly can someone understand a representation and verify the claim it makes?

Hypothesis

Time to verifiable understanding (TTVU) may improve when visual and machine representations share one evidence contract.

Method / prototype

Trace representation → claim → state → observation → source → provenance → time/frame → uncertainty. Compile versioned events into synchronized human, agent and machine storyboards plus a machine-readable panel manifest; compare with a baseline.

Observation

A six-panel synthetic visual storyboard prototype exists. The general event-to-storyboard compiler and participant validation remain open. The evidence chain is a design requirement, not a measured TTVU improvement or a deployed general compiler.

Systems implication

Preserve semantic identity across viewpoints, unknown states and deterministic replay. Changing a view must not manufacture evidence.

Next gate

Implement one bounded event-to-panel example and evaluate verification time, correctness and missing-evidence handling.

Question

Can the same immutable physical-state snapshot drive both a compelling view and inspectable numerical values?

Hypothesis

Shared model and source identity can preserve meaning across visual inspection and machine queries.

Method / prototype

Compare WMM2025 magnetic values with published reference cases; test normal gravity, a bounded dry atmospheric column, interpolation and source admission. Keep synthetic transport and uncertainty fields explicitly separate from observations.

Observation

The separate eight-view workbench reported 30 numerical tests and 18 CPU Canvas2D browser checks. The audit reproduced model agreement and found four interface defects; a local candidate fixes those defects. This does not establish GPU, Safari, device or physical-accuracy validation. Today’s public Earth instrument separately exposes retained imagery, checksum-verified ETOPO samples and a fixed-date magnetic reference.

Systems implication

Model agreement, visual quality and operational navigation accuracy are different claims. Sampling resolution, time, height, units and display transformations travel with the data.

Next gate

Qualify the hardened binding on the actual renderer and obtain GPU/device evidence; admit real vector products before presenting physical flow.

Question

What position, navigation and time remain supportable when connectivity, GNSS or clock references degrade?

Hypothesis

A local evidence path with explicit validity, uncertainty and qualified fallback can remain interpretable through disconnection and recovery.

Method / prototype

Prepare mission references before disconnection; process available observations locally; test communication loss, GNSS unavailability, suspect evidence and clock loss separately. Qualify UAV, UUV, UTV, XR and AV profiles individually.

Observation

The Ojai capture candidate has Linux Swift and synthetic-stress evidence. Apple and physical capture acceptance remain pending in the cited workstream. A common state interface is not evidence of multi-platform PNT qualification.

Systems implication

Keep deterministic estimation and control authority separate from agent orchestration. Last-known position, camera pose, relative registration and a current global fix must remain distinguishable.

Next gate

Record one end-to-end local reference session, replay its failures and quantify performance against an independent reference before expanding platform coverage.

Question

Which claims can a consumer make from incomplete, conflicting, stale or authenticated evidence?

Hypothesis

A shared typed boundary may reduce ambiguous state claims across human, agent and machine consumers.

Method / prototype

Preserve observed versus inferred state; source authentication versus ranging authentication; uncertainty versus integrity; supporting versus conflicting evidence. Carry source identity, lineage, units, time, frame and missingness through consumer policy.

Observation

The shared-state reference implementation reported 42 local tests and nine synthetic browser scenarios. Its receipt identifies an unmerged dependency and a draft, undeployed implementation. Classifier scores and declared lineage independence do not become certainty.

Systems implication

Unresolved conflict or insufficient evidence must not produce an unsupported representative state. Checksums identify retained bytes; they do not establish physical accuracy or independent timestamp attestation.

Next gate

Resolve dependencies, qualify an observation adapter and recorded-trace replay, and compare identical inputs and consumer policies before promotion.

Question

Can direct GNSS RF be retained as raw evidence and converted into a trusted-state input?

Hypothesis

A low-cost receiver experiment may expose the full path from raw RF to replayable observations and a position/time solution.

Method / prototype

Proposed chain: active GPS L1 antenna → RTL-SDR-class receiver → Mac → GNSS-SDR → raw I/Q and decoded observations → Spacetime evidence/state.

Observation

FR-001 is pending review. No purchase, successful acquisition or receiver qualification is claimed. The proposed hardware chain remains a feasibility experiment.

Systems implication

Success would require live acquisition/tracking, navigation-message and ephemeris decoding, pseudorange/Doppler, position/time and retained raw RF for replay.

Next gate

Review feasibility and the $150 maximum request before purchasing parts and executing the proof.

Question

Can external signals become bounded experiments whose measured outcomes improve the next decision?

Hypothesis

Source qualification, deduplication and explicit promotion gates may make research more useful than a stream of untested summaries.

Method / prototype

Observe/discover → hypothesize → select → build → test → publish → measure → learn → update trusted state. Record attempted, built, evaluated, published, learned and adopted separately, alongside CI success and recovery time.

Observation

Signal-intake receipts establish retrieval, deduplication and ledger updates, with NISAR AOI/QC and AI4ATLAS/weather benchmark investigations active. Those intake records do not establish downstream product outcomes. Autonomous execution, evaluation, publication and recovery remain the target operating model.

Systems implication

A completed HCI exploration requires a viewable artifact plus an evaluation result. Zero completed explorations in a week is an escalation condition; a schedule or dispatch is not delivery.

Next gate

Attach each selected signal to a bounded experiment and return measured results; make failed execution create visible recovery work.

The comparison used Southampton Shoal Channel LB 6, station s08010, bin 4, and an assumed 12-minute observation delay. The ten-day holdout covered 22–31 August 2026. Only 1,654 of 2,390 scheduled comparisons had the required observations; missing values were not interpolated. Historical availability, coefficient vintage, upstream training overlap, row-level quality and physical depth remain unresolved.

The retained numerical review independently recomputed this negative result. Retain the harmonic baseline; do not promote the correction. This is one retrospective station/bin comparison, not an operational navigation result.

At the 18 September, 17:01 Pacific audit, the research schedule remained active at 08:00 and 17:00 Pacific, but collection was blocked before execution by a host runtime dependency. Six scheduled captures had completed; eight misses were recorded and two further expired slots had not yet been recorded by the collector. The current attempt created no experiment artifact.

A read-only reconciliation found 14 observed targets, one missing after seven days, six awaiting exact observations and zero pending targets. This did not change the persisted collector state. The earlier snapshot’s three immature targets must not be read as current.

Small lead-time groups (6, 3 and 5 cases), missingness and uncertain source issue times prevent performance ranking. Acquisition time is first seen by this instrument, not first publication by NOAA. Missed slots are not backfilled.

Identity

FE-002 / RSH-007. Retained prospective study, not a released Assure capability.

Snapshot

Reconciliation: 2026-09-19T00:01:03.349+00:00. This dated record is not current collection status.

Result and limits

Predictive performance remains inconclusive. No method ranking is supported. Missing observations remain missing; acquisition time is first seen by this instrument, not first publication by NOAA.

Attribution and access

Named contributor, license and maintainer are not supplied here. Raw-data access is unknown. Public summaries do not supply raw receipts or a reproducibility package.

Separate evidence

The 17 September 2026 field record and 2026-09-18 program register retain their own dates and counts. They are not this reconciliation.