Encounter Differential - Attention Observation Tool

Use Field studies for the full research workflow, even when you need only two artworks. The simpler paired encounter remains available as a separate, shorter flow.

The worked design below compares a clear 1× looking glass with a 2× magnifying glass. It concerns recorded interaction and accounts of an encounter. It cannot establish what someone saw, felt, preferred or understood.

Declare the intended unit of analysis, recruitment and sample-size rationale, repeated measures, exclusions, stopping rule and handling of missing data before collecting observations. A software default is not a power calculation. The four-row schedule is an allocation illustration, not a recommended or validated sample size.

Explain the procedure and intended data use to participants, obtain appropriate consent, choose pseudonymous codes, and keep any identity-to-code key outside this notebook. Decide who may see raw paths, writing and exports, and how long they will be retained.

Before continuing:

• One specific question
• A stated primary outcome and exploratory outcomes
• Participant and data-handling plan
• Declared sampling and exclusion rationale

Choose 2–8 distinct artworks. Use the bundled collection or images you are entitled to use. Record titles, artists, dates and source information, then choose presentation order. Uploaded images stay local and are not recompressed.

For the worked design, keep the same works, simultaneous presentation, identity-label setting, instruction and contextual framing in both conditions. Only the starting magnification changes from 1× to 2×. Record screen, room, lighting, sound and relevant prior familiarity as far as practicable.

Selecting or presenting a work is not a measure of gaze. Navigation in sequential mode records presentation; pointer or deliberate touch/keyboard selection records selected time. Keep input methods consistent when the question requires it.

Before continuing:

• Works and rights checked
• Identity and presentation settings reviewed
• Exact instructions and framing retained
• Device, input method and surrounding conditions recorded

Choose the artwork in the region editor. Click at least three vertices around a meaningful part of the image, then enter a descriptive name and select Save region. Use Undo vertex or Clear drawing before saving. Crossed or degenerate polygons are rejected.

Keyboard users can focus the polygon editor, move the crosshair with arrow keys and press Enter to add a vertex. Name regions descriptively, such as “cypress grouping” or “open sky”, rather than naming an assumed psychological effect.

Coordinates refer to the image, not a particular screen size. Boundary points count as inside. Each polygon is scored independently: overlapping regions accrue overlapping time, so their totals must not be added as though they partition the image.

Region dwell follows the last observed pointer centre until the next observed position, leave, pause or end. Entries count transitions into a region; returns are entries after the first. Lens dwell and magnified dwell concern the pointer centre, not every pixel under the lens. Touch and keyboard selection do not invent a within-image pointer position.

Region definitions are frozen with acquisition. Do not redraw them after seeing results and describe that analysis as prespecified. Amend the protocol for future trials. Legacy records without these measurements remain explicitly incomplete.

Before continuing:

• Regions named and saved on each relevant work
• Overlaps intentional and disclosed
• Definitions settled before looking at results

Enter a protocol title, research question, seed, prespecified analysis plan, stopping/exclusion rules, recruitment notes and common participant instruction. Add each condition and its exact framing, instruction, presentation, identity visibility, cursor, magnification and active-time duration.

For the worked design, create Clear glass at 1× and Magnifying glass at 2×, both for 60 active seconds. Use one repeat per condition for the simplest rehearsal. Treat further repetitions as repeated observations of the same participant, not additional independent people.

Choose Fixed order, Seeded shuffle per allocation row, or Rotated and reversed counterbalancing. The latter rotates condition order and artwork position across rows and reverses condition order on alternate cycles; it is not a full Latin-square design. Inspect the planned order. In this two-condition example, artwork positions stay the same within a participant’s pair of conditions but can rotate across allocation rows.

Save protocol draft while developing the design. Lock protocol captures the current artworks and regions, versions the definition and records its SHA-256 fingerprint. Review the locked work/region counts and planned order before launch. Create amended version for any subsequent change; never silently combine versions in one comparison.

A local lock is an audit aid, not an external preregistration or independently certified timestamp. If the project needs preregistration, register the design through the appropriate separate process before collecting its data.

Before continuing:

• Both conditions reviewed
• Seed and allocation recorded
• Analysis and stopping rules stated
• Current artwork and region counts correct
• Locked version and fingerprint retained

Enter P001 and allocation row 0, inspect the trial list, and select Run assigned trial. Use P002 with row 1 for the next participant, following your declared allocation plan. A participant’s code must stay the same across their conditions and repetitions. The app cannot establish who used a code.

Begin with a rehearsal. Use a clearly marked rehearsal code and record an explicit exclusion decision afterwards; do not count it as participant evidence. Check image loading, pointer/lens behaviour, timing, regions, instructions, first-account collection and exports before recruiting.

The participant reads the exact framing and instructions before Begin study encounter. In protocol mode, lens controls stay fixed and live counters are withheld. Pause and stop remain available. Images are withheld while paused; loss of focus or backgrounding pauses measurement and requires explicit resume.

Fixed-duration trials end at the active-time target; early manual stopping is recorded separately. Actual recorded duration remains in the log. Five-second checkpoints support interrupted-session recovery, which preserves the last committed trace and marks the encounter interrupted. The missing final interval is not reconstructed.

After each encounter, save the first descriptive account before opening analysis or revealing withheld identities. An empty account is also a recorded choice. The first account stays separate from later trace-assisted reflection. Continue to next assigned trial from the report. For a new participant, reopen the saved locked protocol and assign the next code/row; do not treat changing a code as an automatic new recruitment record.

Before continuing:

• Code and allocation row correct
• Rehearsal excluded with a reason
• First account precedes analysis
• Early stops and interruptions retained
• Every planned trial accounted for

Explore per-work selected share, visits, returns, pointer paths, replay, regional maps, transitions and condition comparisons. Active time includes unassigned time; selected share is selected work time divided by total active time. Different exposure durations or input methods can change the comparison.

Inspect the region table for dwell, entries/returns, first entry and lens/magnified dwell. Inspect the measurement manifest for software/calculation versions, source-image checksums, viewport, actual rendered image geometry, sampling and interruptions. Missing acquisition information in legacy records has not been reconstructed.

Use the claim–evidence ledger to distinguish observation, self-report, inference and counterevidence. Give each entry a coder code, an evidence reference, uncertainty and, when relevant, an artwork, region and active-time interval. An account that disagrees with the trace is not automatically an error to correct.

For example, an observation may say “the pointer centre remained inside the cypress region for this interval”. “The participant preferred the tree” is an inference that this trace cannot establish. A participant’s own statement remains self-report, not a measured preference score.

Use Revise annotation when an entry changes. Earlier revisions remain inspectable. Under Analysis inclusion decision, state and save a reason for any exclusion; the original recording stays retained. Independent coders should use distinct codes and retain disagreements before any separate adjudication.

Before continuing:

• Claim types distinguishable
• Evidence and uncertainty explicit
• Exclusions have reasons
• First account preserved alongside later reflection

Select one exact locked protocol version/hash. Filter by participant code or input method only when that matches the declared analysis. The input filter includes sessions that used that method; mixed-input sessions may still be present.

Completed trials with participant codes contribute by default; interrupted, early-stop and explicitly excluded records do not. Review included/excluded counts and repeated-assignment flags. Retakes remain visible rather than being silently discarded; inspect them and record any exclusion decision.

Repeated trials are averaged within participant and condition before the group descriptive mean is calculated. Each per-work row reports its own participant and trial counts, mean selected share and observed participant range. Zero-active-time trials cannot supply a selected-share value.

Within-participant differences include only people with both compared conditions. The displayed difference is second condition minus first in percentage points of selected share. It is descriptive and unadjusted for order, familiarity and other covariates; the app does not produce a significance test or a causal verdict.

The group panel summarizes per-work selected share. Inspect ROI dwell and written accounts in the individual records and exported tables, or use a separately specified external analysis. The interface does not automatically calculate cohort ROI effects, coder agreement, statistical power or eye-tracking validity.

Before continuing:

• One exact protocol version
• Independent participants counted correctly
• Missing conditions and repeated assignments reviewed
• Order, familiarity, input and lens reactivity reported
• Descriptive findings kept separate from causal claims

In a study report, choose Export reproducible ZIP. In Participant analysis, Export selected trials & tidy tables exports the current selection. The bundle contains authoritative study JSON, frozen protocol definitions, region vertices, first accounts, later reflections, annotations, manifests, source images when retrievable, CSV tables, a data dictionary, checksums and a Python verification script.

Extract the ZIP into a folder. Run “python3 verify-and-recompute.py” from a terminal using that folder’s script. It checks bundled file hashes and recomputes complete work dwell and polygon dwell from the retained trace. A truncated legacy trace is reported rather than treated as a complete recomputation. This verifies selected software calculations, not scientific validity.

Review assets.json: exported images are compared with acquisition hashes when available. A later remote image can differ or become unavailable. Do not describe an unavailable or changed image as the verified original stimulus. CSV text that could execute as a spreadsheet formula is escaped; JSON preserves the exact text.

Use Database & recovery → Back up full database separately. That backup includes revisions, original-storage snapshots, images, removed records, conflicts and recovery drafts. A research ZIP is an analytical package, not the full database recovery archive.

IndexedDB commits locally; the SQLite mirror is a second local copy when connected. Check pending writes and mirror status. Browser and desktop origins have separate notebooks. Use an inspected full backup to transfer data. Keep a backup outside the working device, review sharing and rights, and do not publish raw participant writing or identifiers by accident.

Before continuing:

• Correct records and filters exported
• Dictionary and manifests retained
• Checksums and recomputation reviewed
• Changed/unavailable images disclosed
• Full database backup stored separately
• Sharing reviewed before release