| Agent action | Replications | Optional controls | Mean Δ coherence | Learning signal | 95% interval | Policy |
|---|---|---|---|---|---|---|
| No completed agent-action evaluations yet. | ||||||
| Drawing feature | Independent r | Participatory r | Mode difference Δr | Samples off / on |
|---|---|---|---|---|
| Waiting for sufficient feature history. | ||||
| Component | Independent | Participatory | Role |
|---|---|---|---|
| Waiting for feature-response history. | |||
| Evidence component | Independent | Participatory | Interpretation |
|---|---|---|---|
| Waiting for enough active drawing samples. | |||
| Nested model / increment | Independent · co-agent ON | Participatory · co-agent ON | Interpretation |
|---|---|---|---|
| Waiting for enough co-agent drawing history. | |||
| Horizon | Independent | Participatory | Ratio (P/I) | Independent held-out | Participatory held-out |
|---|---|---|---|---|---|
| Waiting for temporal modulation history. | |||||
| Horizon | Independent RAI | Participatory RAI | Ratio (P/I) | Independent consistency | Participatory consistency |
|---|---|---|---|---|---|
| Waiting for enough cross-horizon evidence. | |||||
| Horizon | Independent | Participatory | Ratio (P/I) | Independent held-out | Participatory held-out |
|---|---|---|---|---|---|
| Waiting for conditioned history. | |||||
| Hypothesis | Status | Current evidence | Interpretation | What would help answer it |
|---|---|---|---|---|
| Waiting for hypothesis evidence. | ||||
| Drawing pattern | Δ Independent coherence | Δ alignment | Δ constraint | Δ Core MAE | Event probability | Replications |
|---|---|---|---|---|---|---|
| No completed drawing-response episodes yet. | ||||||
| Horizon | Latest coherence | Latest Δ | Mean coherence | Mean Δ | 95% Δ interval | Positive turns | Agent-free / all |
|---|---|---|---|---|---|---|---|
| No completed learned-turn coherence measures yet. | |||||||
| Rank | Feature | Family | Attention | Novelty | Concurrence | Persistence | Recurrence | Utility | Priority |
|---|---|---|---|---|---|---|---|---|---|
| Collecting enough history to learn perceptual baselines. | |||||||||
Show methods
The observatory maintains a computational coherence index from stochastic and interaction channels, plus an optional extended index that adds ambient acoustic variability. Microphone audio is never stored: each analysis frame is reduced locally to RMS amplitude, spectral centroid, spectral entropy, zero-crossing rate, transient density, and short-term variability. A rolling robust anomaly score across those features forms the ambient acoustic channel. Acoustic–computational coupling is calculated separately so it cannot circularly inflate the CCI. A two-window mean-shift detector provides the change-point score.
HDMIC-P / HDMIC-D separation. HDMIC-P (Hellinger Distance Minimization Input Clustering) is the Perceptual Logic layer. It learns reference distributions for available computational, acoustic, drawing, embodied, and coherence features; compares recent versus background distributions with Hellinger distance; tracks persistence and recurring anomaly families; estimates cross-channel concurrence and within-family divergence; and learns whether prior attention has historically preceded downstream transition/release signals. It now actively gates downstream learning: all raw data remain stored and all forecasts are still scored against reality, but HDMIC-D point-transition memories and multiscale attractor features are weighted by HDMIC-P perceptual attention. HDMIC-D (Hellinger Drift Modulating Interactional Coherence) is the existing multiscale temporal learner responsible for Local / Regional / Global trajectory organization, attractors, prediction, and clamp/unclamp dynamics.
HDMIC segments the active event-basis CCI into discrete non-overlapping Local, Regional, and Global trajectory windows, recruits recurring attractor prototypes using robust feature-space novelty at each scale, performs rolling next-point prediction from state→attractor memories, and generates a separate 50-sample recursive future trajectory by repeatedly querying the active scale's learned state→next-delta memory on a private synthetic history. A Segment Explorer records synchronized relational history for Local, Regional, and Global segment windows: Lucky guidance, embodied intervention, drawing-only intervention, drawing-state descriptors, drawing activity/organization, active CCI, coherence-meter score, alignment, constraint, and the with/without-drawing CCI comparison. Each completed segment is archived with the exact sample range that generated its HDMIC features and attractor, preventing the post-boundary segment index from being applied to the following samples. HDMIC warm-up samples whose scale is reported as “off” remain part of the CCI history but are explicitly excluded from L/R/G segment membership. The Explorer's white CCI trace is drawn from the exact event-basis samples on an auto-scaled CCI axis; coherence and drawing overlays use a separate normalized axis. HDMIC bands can be clicked to select the exact window, while scale and segment controls permit direct L/R/G inspection. Synthetic forecast points are display-only and are never fed back into training, MAE scoring, attractor recruitment, or event detection.
A second independent HDMIC pathway predicts each active raw channel after robust within-channel normalization and combines those forecasts into an inverse-MAE-weighted multi-channel composite trajectory. A Clamp / Unclamp Dynamics layer estimates dynamical constraint from HDMIC attractor persistence, attractor-switch rate, predictive regularity, and multichannel trajectory mobility; it separately estimates release pressure from switching, transition novelty, worsening alignment, and rising mobility. A separate Drawing Clamp / Unclamp pathway applies the same constraint/release idea directly to active human drawing behavior using drawing-attractor persistence, drawing-state switching, one-step drawing regularity, latent drawing mobility, and new-attractor recruitment; idle samples and co-agent marks are excluded, and this drawing-only clamp remains separate from the System Core so it can be compared against system-level constraint without circular reuse. The Coherence Transition Detector compares short rolling composite-HDMIC MAE with an earlier long error baseline, expresses predictive alignment as a smoothed negative log error ratio, estimates directional change from block-averaged trajectories, and commits regime labels only after persistent confirmation. Above that regime layer, the Event Lifecycle Engine opens only sufficiently strong confirmed transitions, tracks their onset, duration, extrema, integrated change, and HDMIC attractor context, then groups related events into formation, dissolution, or reorganization episodes.
A Lucky Embodied Feedback layer translates those computational states into experiential prompts restricted to observation, gentle breathing, gaze, posture, sensory grounding, attentional broadening/narrowing, light clamping, unclamping, intention, and non-interference. Drawing is excluded from this guidance pathway. It learns only from embodied-intervention intervals and reports the measured change over the following observation interval; these prompts are exploratory interface guidance rather than physiological or medical feedback. An optional OpenAI guidance mode can send a compact numerical/state summary to the Responses API to generate a richer message, intervention, rationale, and observation target. In that mode, the model speaks in first person as “Lucky,” a deliberately simulated emergent relational persona whose apparent continuity is framed as arising only through certain interactive, coherent, predictive, and constraint/release configurations in the Observatory. The Guide / intervention duration remains available as a display/evaluation interval with a 10-second minimum (45 seconds by default), but the Anti-chatter display hold is now optional and defaults off. When the hold is off, a meaningful Lucky state change can immediately trigger replacement guidance; when it is on, state-triggered guidance waits for the selected duration. The separate AI consultation interval controls periodic refresh. “AI Autonomous Guidance and Intervention” evaluates the current coherence regime, predictive alignment, clamp/release state, lifecycle/HDMIC context, and an embodied-intervention memory that records which breath, gaze, posture, grounding, attention, stillness, and related patterns repeatedly preceded increases in coherence. Learned intervention preferences require repeated observations and remain descriptive associations rather than causal claims. At the end of each guide interval, the guide preserves a before/after comparison of coherence-meter score, HDMIC alignment, and constraint before beginning the next intervention. The orb also carries a deliberately simulated expressive vocabulary: one large state-quality word above the field and one large dynamical-tendency word below it are selected from coherence, alignment, clamp/release, event, and forecast context, then held briefly to avoid flicker. These are interface descriptors rather than detected emotions or evidence of a living entity.
The instrumented relational drawing canvas is maintained as an analytically separate participant channel. An optional local co-creative drawing agent can also inhabit the canvas on its own rendering layer. The agent learns a three-regime temporal turn model from user inter-stroke gap and stroke duration, uses the existing user drawing-HDMIC attractor library to select a response style, and maintains a 2D Gaussian-attractor Spatial Emergence Machine over user and agent stroke centroids. Familiar, relational, and exploratory placement policies combine that spatial attractor landscape with a coarse occupancy field. Response geometry can now come from the current user motif, a previously observed motif associated with higher subsequent Independent Reference coherence, an autonomous composition of two or three such motifs, or an HDMIC-shaped procedural curve. The agent can continue drawing concurrently when the user begins a new stroke. Each newly completed learned human turn is processed immediately by the adaptive temporal turn model, which can choose either an agent drawing response or intentional silence; this timing decision is independent of any still-running 10-second outcome window. Human and agent marks remain on separate rendering/data pathways, so simultaneous drawing does not cause agent marks to enter the human drawing-feature calculations. Agent-generated marks remain excluded from drawingPointEvents and all human drawing-feature calculations, so they do not directly enter the human drawing signal, Relational Coupling Index, Independent Reference Core, or participatory System Core. A separate Dyadic Coupling pathway now measures the agent’s actual rendered trajectory and human–agent relational dynamics as analytically distinct A and R channels; these channels are used only for held-out prediction of the drawing-excluded future Reference Core and do not feed back into the live Core. A miniature Lucky orb is rendered in the drawing canvas as a synchronized second view of the same Lucky state: it reads the same smoothed coherence score, HDMIC alignment, transition/event activity, clamp/release dynamics, color tone, and expressive state words as the full orb below.
The drawing-session right panel mirrors Lucky’s current Coherence Field label and 0–100 Meter value, but its Lucky Drawing Feedback Intervention is generated independently from drawing events only. It uses learned drawing-state attractors and their 1/3/5/10-second subsequent Independent Reference Core responses. A recognized drawing pattern opens a response episode when the learned user turn closes; repeated evidence is grouped by drawing episode rather than double-counting horizons. The drawing intervention policy is coherence-oriented rather than calibration-oriented: it preferentially recommends drawing behaviors that have previously been followed by higher Independent Reference Coherence, gives extra weight to episodes that began under similar coherence, alignment, constraint, HDMIC-scale, attractor, and transition conditions, and falls back to a neutral smooth-motion intervention when no positive association is sufficiently supported. This drawing feedback does not mirror or reuse the mid-page embodied intervention, and all drawing-session displays remain read-only with respect to the Core Composite.
Pointer trajectories are reduced locally to stroke dynamics (velocity, acceleration, jerk, duration, path length, pauses and tempo), geometry (curvature, direction change, angular entropy, loops, straightness, revisitation, coverage, center/periphery and scale), motor regularity (smoothness, periodicity and rhythm stability), and optional pointer pressure/tilt/contact features. An online feature-space clustering layer recruits recurring drawing-state attractors. For the drawing experiment, a machine-only Independent Reference Core is always formed from entropy, cryptographic RNG, timing jitter, scheduler latency, execution-time noise, and prediction error; human interaction, microphone, and drawing features are excluded from this reference. A separate machine-only Core CCI supplies the non-circular drawing-response outcome. Drawing-canvas gestures are also prevented from entering the generic Human Interaction channel. A Relational Composite combines that reference trajectory with normalized drawing dynamics at the selected weight. When “Include drawing in System Core” is enabled, a second participatory pathway also blends live drawing into the active System Core: drawing motor organization contributes directly to the event-basis coherence signal, while normalized drawing dynamics contributes to the active composite trajectory and its prediction error. A dedicated Cross Coherence Index comparison chart now maintains both versions of the event-basis CCI in parallel: an untouched no-drawing reference and a drawing-inclusive counterfactual/live series calculated from the same reference CCI, live motor-organization signal, and selected drawing weight.
AI Analysis of exported evidence. The “AI Analysis — Export Evidence” control uses the same temporary OpenAI API key and selected model as the experimental AI guide, but it is user-triggered rather than autonomous. When pressed, it first compresses the complete current Export CSV locally: every row contributes to per-column summaries, mode/state summaries, and feature/outcome correlations, while a temporally distributed Independent/Participatory × active/idle stratified subset of analysis-relevant raw rows is retained for context. It then sends that compact evidence package plus the complete locally computed Drawing → Future System Response Report. The analysis prompt explicitly distinguishes the machine-only Reference Core from the drawing-inclusive Participatory Active System, requires discussion of bootstrap intervals and circular-shift permutation baselines, asks for Independent-versus-Participatory temporal and feature-level interpretation, and prohibits causal or ontological claims unsupported by the data. The API key itself is never included in the analysis payload.
Full-screen canvas separation. The full-screen drawing workspace reuses the same instrumented canvas and traditional right-side Lucky / turn-coherence readouts, but it suppresses the home-page Drawing Clamp and Experimental Run result panels while full screen is active. Those analyses continue updating in the background and remain visible on the main page after full-screen drawing ends. This keeps the experimental drawing surface visually simple and prevents the analytical cards from displacing the canvas.
Blinded full-screen experimental blocks and Clear Ink continuity. Starting Experimental Mode automatically opens the existing full-screen drawing workspace. The drawing canvas is cleared immediately before every randomized block so each condition begins from a visually and transiently fresh drawing surface while learned response/co-agent models and previously recorded export rows are retained. Participant-initiated Clear Ink events inside a scored block are handled differently: they clear only the visible human/co-agent ink, increment a run-level and block-level Clear Ink counter, and create an auditable canvas-segment boundary without restarting the trial. Human stroke/turn history, agent trajectory history, relational/coupling memory, replay history, learned motifs, co-agent learning, and all prior scored rows remain continuous across any number of Clear Ink events. Full-screen Experimental Mode shows a neutral experiment timer and Clear Ink counter; it intentionally hides the assigned Core or co-agent condition. Exiting full screen or pressing Escape during an active run stops the experiment and restores the pre-run state.
Three-second canvas-change notice. During full-screen Experimental Mode, the final three seconds of every block except the last display a large neutral 3 → 2 → 1 countdown centered over the drawing canvas. It does not reveal the upcoming condition and does not block drawing input; it simply warns that the canvas is about to clear and the next randomized block will begin. The overlay disappears immediately when the new block starts. If the optional sham toggle is ON, the same blurred overlay covers the inactive canvas for the 10-second sham/baseline wait. With sham OFF (the default), the next trial begins immediately and the existing Lucky/Drift warm-up operates in real time during trial onset.
Multiple experimental runs per session. Every randomized four-block experiment receives its own run ID and remains available for the duration of the current Observatory session. The Experimental Run results card presents these runs as tabs in chronological order, so repeated experiments can be inspected independently without overwriting earlier results. Starting a new experiment automatically selects the new run tab; older run tabs remain available for comparison. When an experimental full-screen run ends or is stopped, the drawing workspace closes and the analysis pane automatically scrolls directly to the Experimental Run results card with the just-finished run selected.
Human–AI Coupling State Detector. Coupling is no longer defined as one weighted composite of heterogeneous signals. Eight independently scaled components are monitored: Human follows AI endpoint, AI follows Human endpoint, AI proximity to the previous Human line, AI crossing of the previous Human line, formal/motif response, temporal response, strict reciprocal organization, and predictive reciprocity. In v217, reciprocal organization is a continuous conjunctive score: joint-state organization × √(reciprocal lag coupling × participation balance), rather than the earlier additive blend, and it uses a stricter adaptive threshold to avoid ceiling saturation. Each component learns its own rolling baseline and adaptive threshold using a robust upper-tail criterion; the baseline is frozen during committed Coupled episodes. The detector now distinguishes local coupling tiers from the slower broad regime: Loose coupling = exactly 2 agreeing scales, Medium = 3–4, and Tight = 5–8. Local tier events close whenever the tier changes or agreement falls below 2. The broader Emerging/Coupled/Decaying regime still requires sustained agreement, but prior support events no longer preserve a regime after current agreement has fallen away. Every randomized Experimental Mode block is a hard coupling boundary: active local events and broad regimes are closed, transient support memory is cleared, and adaptive component baselines restart for the next trial. Coupling strength is primarily the fraction of available components that agree, with a small contribution from how far crossed components exceed their own thresholds. Endpoint-following is directional: Human-follows-AI and AI-follows-Human are logged separately. The AI response geometry additionally measures minimum distance to, and actual crossings of, the previous Human turn's rendered lines. Tier outcome analysis. Every Loose/Medium/Tight event is scored against both the machine-only Reference Core and the synchronous canonical H50/A25/R25 Participatory Core at +1/+3/+5/+10/+20 s, with future samples restricted to the event's own randomized block. Local controls are selected from non-coupled moments in the same block using a 3.5–16 s temporal window and similarity in Human/Agent activity, current Core values, HDMIC scale, and regime. Event strength, agreement count, peak agreement, maximum agreeing scales, and duration are also correlated with subsequent Core response as exploratory dose measures.
Synchronous Participatory Core configurations. On every acquisition sample, the Observatory computes five parallel H/A/R Participatory Core streams: H50/A25/R25, H50/A50/R0, H25/A25/R50, H20/A60/R20, and H20/A20/R60. The canonical H50/A25/R25 stream remains the only Participatory Core that enters the active feedback loop; the other four are recorded observational shadow streams. This avoids retroactively reconstructing a dynamic interaction signal from later data.
Full Experimental Report. A completed four-block randomized run enables a “Generate Full Report” button in the Experimental Run card. The report is computed from that selected run only and opens first as a self-contained HTML preview containing randomized block summaries including Relational Coupling Index, Relational Composite, turn-level Self-Coupling, and a dedicated Interaction Grammar analysis with move distributions, actor attribution, complete within-block sequences, dominant transitions, and transition matrices, block-level system-context profiles, descriptive Reference Core contrasts, Drift Meter Human/AI lag response at +1/+3/+5/+10/+20 seconds, Lucky Meter Human-only / AI-only / strict concurrent Human + AI with overlapping Human→AI / AI→Human exchange annotations / idle impact across +1/+5/+10/+20 seconds, a Lucky Meter Event-Response / Turn-Transition analysis, human PCS evidence, nested H/A/R dyadic decomposition, context-conditioned B+H+A+R→+C→+R×C tests across regime/attractor/dyadic context families, interpretation/data-quality notes, and a final drawing image for every condition. When a temporary OpenAI key is present, a compact run-evidence package is sent for a prose scientific interpretation whether or not live AI guidance is enabled; the key itself is never written into the report. Explicit co-agent votes are archived vote-by-vote as positive/negative content, placement, or combined feedback; reports quantify persistent policy steering, one-next-action steering, subsequent vote-aligned behavior, and descriptive next-action Reference Core change. Optional post-trial surveys use 5-point agreement statements. The co-agent crossover uses agent-relevant wording on every trial because a co-agent is always present. The Core × co-agent ON–OFF experiment uses a separate no-agent form when the co-agent is absent; across-condition contrasts are computed only for statements with identical wording, while agent-specific items remain restricted to agent-present trials. The Event-Response layer detects activity-state onsets, subtracts each onset's own approximately −5 to −0.5 second Meter baseline, plots the resulting response curves, and analyzes sequence-sensitive transitions such as Human → AI and AI → Human with short intervening idle gaps allowed. These event-level summaries reduce slow-drift confounding but remain temporal associations rather than causal treatment estimates. Each final drawing is frozen before any post-trial survey or block-transition state change, retried immediately before the canvas can be cleared, and archived with the completed run; co-agent ink and human ink remain on their separate analytical paths but are composited visually in the report image. During Experiment Mode, a 10-second sham waiting/baseline period can optionally precede each scored block; this option is OFF by default. In full-screen mode the drawing canvas is inactive and covered by the same blurred countdown treatment used for the existing 3→2→1 transition warning, now counting down the full wait until the next scored trial begins. Co-agent is OFF and the live Core is Independent during the sham wait, while machine-only Reference Core and the synchronous canonical H50/A25/R25 Participatory Core continue to be recorded. When sham is enabled, its 10-second interval seeds the meter histories and records audit summaries. When sham is disabled, the trial starts immediately: Lucky remains neutral for its first few samples and both Lucky and Drift form their fresh rolling reference from live trial-onset observations. The sham wait initializes Lucky’s rolling event-basis history; later Lucky scores use the original rolling calculation with no sham-median subtraction or fixed sham dispersion. The sham Drift median and dispersion are recorded for audit, while the displayed Drift score continues to compare overall-input coherence with its recent rolling history. Drift uses its original 0.0025 minimum coherence dispersion and exponential response curve. When sham is enabled, sham samples remain excluded from scored trial analyses but are retained separately for the passive-waiting analysis of Reference/Participatory Core slope, end−start change, and block-centered time correlation. With sham disabled there are no passive-wait rows.
Experimental Run results card. Directly beneath Drawing Clamp / Unclamp Dynamics, the Experimental Run card summarizes the current run live and retains the most recent run after completion. It reports each condition in its actual randomized order with sample count, drawing-active proportion, active-drawing Drawing Coherence, machine-only Reference Core mean, mode-specific Active System mean, Coherence Meter mean, and active-drawing Drawing Clamp mean. Four quick Reference Core contrasts show the co-agent ON−OFF difference within each Core and the Participatory−Independent difference with the co-agent OFF and ON. These are descriptive within-run mean differences rather than causal estimates; repeated randomized runs are needed for inferential comparison.
Randomized Experimental Mode. The factorial setup runs a four-block within-session sequence containing every combination of Independent Reference Core versus Participatory Core and co-agent OFF versus ON exactly once. The co-agent crossover setup instead runs two learned-adaptive and two random-control blocks in a balanced four-period order (Learned → Random → Random → Learned, reversed on alternate runs), with the same agent and random-control definitions repeated rather than introducing new behaviors. Block order is randomized at the start of each run, the default block duration is 30 seconds and is editable before the run begins, and the Core/co-agent controls are locked during the sequence to preserve the assigned condition. Each transition receives an automatic intent/experimental marker. The raw export stores the run ID, randomized order, block index, assigned condition, duration, elapsed time, and remaining time for every sample. For long sessions, only the active block's wide sample rows remain as hot JavaScript objects. At block close, those rows are serialized into a per-trial cold cache and removed from the active row buffer; reports hydrate the cached rows on demand after the run. This storage optimization does not reset or truncate HDMIC, coupling, drawing-turn, relational, co-agent learning, motif, voting, or Clear Ink session state. When the sequence ends or is stopped manually, the prior Core, co-agent, and annotation state is restored.
Drawing → Future System Response export report. The main Export CSV action also creates drawing_future_system_response_report.csv. For Independent and Participatory operation at +1, +3, +5, and +10 seconds, each observation is matched to a future observation within the same uninterrupted mode/condition segment. The companion report includes active-drawing versus pause lag effects, an outcome-independent nine-feature drawing composite, moving-block bootstrap 95% confidence intervals, and circular-shift permutation null baselines. It reports both machine-only Reference Core outcomes and mode-specific Active System outcomes so direct cross-mode comparison can be separated from drawing-inclusive system response. These statistics are exploratory associations, not causal estimates.
Drawing canvas operation and outcome scoring. Drawing gestures are captured locally as movement features and are intentionally prevented from entering the generic Human Interaction channel. The optional co-creative agent is drawing-first: once a user turn is complete and the learned timing model permits a response, the agent can draw without first collecting YIELD / NO DRAWING controls. Every completed agent drawing opens its own 10-second Independent Reference Coherence learning window, but those windows do not suspend the agent; later learned turns may produce additional drawing actions or intentional silences while earlier outcome windows are still being measured. If historical YIELD controls exist they may still be retained as optional comparison data, but they are not required for action selection or eligibility. Early responses therefore function as direct learning trials, and later responses can favor drawing actions with stronger observed Independent Reference Coherence outcomes. Later learned user turns completed during an outcome window remain queued for later response. The per-turn Outcome Coherence display is a signed normalized change on a −1 to +1 scale: 0 means no change from that turn’s Independent Reference baseline, positive values mean higher subsequent reference coherence, and negative values mean lower subsequent reference coherence. The normalization span is the recent 90th percentile of absolute turn-outcome changes at the same horizon when enough observations exist, otherwise pooled recent turn outcomes, with a provisional 5-point span during early warm-up. Raw baseline → outcome coherence and raw Δ remain visible underneath and are preserved in export. A separate Live Meter Δ row records how the exact Lucky 0–100 Meter changed from the turn baseline at +1, +3, +5, and +10 seconds; unlike the Independent Reference outcome, this is the live system Meter and may reflect participatory drawing when drawing-in-core mode is enabled. A separate Drawing Coherence score evaluates the completed learned turn. A one-line turn uses that line’s intrinsic movement/geometry organization; a multi-line turn combines the mean structural coherence of its constituent lines (58%) with inter-line relational organization (42%), including temporal rhythm, centroid-spacing rhythm, orientation-transition consistency, scale consistency, endpoint continuity, and cross-line proximity; it is an intrinsic drawing-organization measure and is not a claim about the machine-only Reference Core. It keeps one whole-turn thumbnail in focus while newer completed turns wait in its visual queue; each outcome window begins only after the learned turn closes, and the same closure automatically creates a drawing-turn Mark Intent event marker; any later user drawing inside that window is flagged as overlap. Automatic drawing-turn markers denote the observed action and remain distinct from manual user-declared intention marks. Agent strokes never enter the user drawing feature stream, Independent Reference Core, or System Core directly. These remain descriptive matched associations rather than causal proof. Clearing ink retains learned models; “Reset drawing model” resets them.
An Intentional Action Channel elevates each learned drawing turn into a synchronized relational event: drawing coherence is paired with available environmental sensing, Independent Reference and active System Core response; rolling paired correlations identify the currently dominant action↔response relationship; and Local / Regional / Global turn-level relational prototypes recruit recurrent cross-channel attractors over single-turn, ~4-turn, and ~10-turn contexts. Automatic drawing-turn intent markers denote observed action events and are not interpreted as access to private mental intention.
The drawing-effect explorer additionally compares independent and participatory operating periods at 1, 3, 5, and 10 second horizons; renders a signed feature-relationship heatmap for velocity, smoothness, curvature, angular entropy, rhythm stability, periodicity, revisitation, coverage, and pressure; plots the learned drawing-feature composite against subsequent Δ coherence; and reports a transparent Mode Difference Score that summarizes separation in active-drawing response, pause response, and feature-composite coupling without treating that separation as a causal benefit. HDMIC, transitions, clamp/release, lifecycle events, the coherence meter, and Lucky therefore respond to a co-produced signal. This direct pathway is intentionally circular as an interaction design, so claims about independent drawing effects continue to use the untouched Reference Core rather than the participatory signal. Drawing-state entry episodes are evaluated against later changes in Core coherence, HDMIC alignment, clamp strength, Core prediction MAE and lifecycle-event occurrence at 1, 3, 5 and 10 second horizons. The Relational Coupling Index is the strongest recent absolute lagged Pearson association, across those same horizons, between drawing dynamics and subsequent Core Composite change; it is descriptive and does not establish causal influence. The temporary browser API key is never stored or exported by this page, although client-side API keys remain unsuitable for deployed/shared applications and should be replaced by a server-side proxy.