The three-body problem is famously intractable in closed form — no general analytic solution exists. But the Sundog pattern isn't about solving the full dynamics. It's about asking: can an agent act usefully from an indirect, compressed signal instead of requiring the full state? In three-body systems, the full state is 18-dimensional (three position vectors, three velocity vectors). But many practical questions about three-body dynamics don't require tracking the full state at every instant. They require detecting signatures of dynamically important events: near-collisions, ejections, resonance capture, exchange events, stability transitions. These events cast shadows in lower-dimensional observables long before they resolve in full phase space. Where H(x) maps onto three-body dynamics The Sundog move would be: instead of integrating the full 18D system and asking "what happens?", instrument the system with indirect observables and ask "what do those observables tell you about what's about to happen?" Candidate indirect signals that are cheaper than full state: The moment of inertia tensor (or its scalar trace) — a 1D or 6D compression of the 9D positional configuration. When a three-body system is about to undergo a close encounter or ejection, the eigenvalue spectrum of the inertia tensor changes character before the event resolves. The smallest eigenvalue dropping signals approach to a collinear configuration; the ratio of eigenvalues signals how "hierarchical" the system is at any moment. Pairwise energy partition — for each of the three pairs, compute the two-body energy (kinetic + potential of that pair in their center-of-mass frame). In a stable hierarchical triple, one pair's energy is deeply negative (the inner binary) and one is weakly negative (the outer orbit). When the system transitions from hierarchical to democratic (the chaotic scattering regime), these energies approach each other. That crossing is an indirect signal for instability. Angular momentum exchange rates — the time derivative of the angular momentum of the inner binary. In a stable system this oscillates secularly (Kozai-Lidov). When the oscillation amplitude grows past a threshold, it signals the system is approaching a regime where the hierarchy will break. You don't need the full state to detect this; you need ∂L_inner/∂t, which is a scalar. The virial ratio (2T/|W|) — kinetic energy over potential energy. For a bound system this oscillates around 1. When it persistently exceeds 1, the system is unbound or about to eject. This is a single scalar that compresses all 18 dimensions into a stability diagnostic. What a Sundog controller for three-body would look like Imagine a scenario where you're controlling one body (say, a spacecraft in a three-body gravitational environment) and you can observe some but not all of the system state. The Sundog architecture maps directly: SCAN: Probe the dynamical landscape by executing small maneuvers and observing how the indirect signals (virial ratio, pairwise energies, angular momentum exchange) respond. This is the three-body analogue of the Lissajous scan in the photometric experiment. SEEK: Identify the region of control space where the indirect signals indicate favorable dynamics — approaching a stable resonance, avoiding ejection, capturing into a desired orbit family. TRACK: Use extremum-seeking on the indirect signal to maintain the desired dynamical regime without requiring full state estimation of all three bodies. The analogue of "tight joint limits" as a failure boundary would be: when the system enters the strongly chaotic scattering regime (all three bodies at comparable mutual distances, no hierarchy), the indirect signals become noisy and the coupling between your control action and the observable weakens. The Sundog controller would lose its gradient signal in the same way the photometric controller loses its signal when the optimum is outside the reachable workspace. What's honest vs. what's reach Honest: the indirect-signal-to-action pattern has a structural analogue in three-body dynamics. The observables listed above are well-known in celestial mechanics but typically used for post-hoc analysis, not for online control. Reframing them as control-relevant compressed signals is a genuine contribution if it leads to a controller that works. Reach: claiming Sundog "solves" three-body dynamics. It doesn't. The chaotic sensitivity is real and no compression scheme eliminates it. What the pattern could do is identify a practical operating envelope where indirect signals are sufficient for useful action — the same trade the photometric experiment documents: comparable terminal performance, slower convergence, known failure boundary. A concrete first experiment The minimum viable test: a restricted three-body problem (circular restricted, so you have two primaries and one test particle). The test particle has no access to the primaries' exact masses or positions — only to tidal acceleration (the gradient of the gravitational field at its location) and its own velocity. Can a scan-seek-track controller find and maintain a halo orbit around a Lagrange point using only the tidal signal? The oracle baseline would be a controller with full knowledge of the primaries. That's small enough to defend and connects directly to the photometric experiment's structure. Where do you want to push on this? ## Ratification - 2026-05-07 Status: ratified first as a promising research direction, and now supported by a bounded operating-envelope result in the planar restricted prototype. The hook is strong because the three-body problem is a familiar symbol of dynamical complexity, and Sundog has a clean way to enter the conversation: not by claiming to solve the dynamics, but by showing that lower-dimensional event signatures can support useful action inside a bounded operating envelope, while naming where that envelope fails. The actionable core is not "Sundog solves three-body." The actionable core is: > In selected three-body regimes, can an agent use compressed, indirect > dynamical signatures to detect instability, choose interventions, or maintain > a desired orbit family with less privileged state than an oracle controller? That is worth pursuing. It maps cleanly onto the existing Sundog pattern: 1. deny full-state access; 2. expose only indirect observables; 3. transform those observables into control-relevant signatures; 4. compare against a privileged baseline; 5. report the failure boundary. ## Actionability Audit The current proposal has one important caveat: several candidate observables are compressed, but not automatically indirect. Pairwise energies, inertia tensors, virial ratio, and angular-momentum exchange can require positions, velocities, masses, or pair identity. If those are computed from full simulator state, they are diagnostics, not Sundog-style partial observations. So the first work item is a sensor-model audit: - Mark each observable as privileged, partially observable, or locally measurable. - Separate "compressed full-state diagnostic" from "available indirect signal." - Define exactly what the controlled body can sense: local acceleration, tidal tensor estimate, range/range-rate, Doppler shift, bearing-only observations, own thrust history, own velocity, or noisy local field samples. - Preserve an oracle baseline with full CR3BP state for comparison. This keeps the claim defensible and may make the experiment more interesting. If a compressed diagnostic works only with privileged state, it can still guide feature discovery, but it should not be presented as the Sundog controller's input. ## Ratified Hook Language Safe hook: > Sundog does not solve the three-body problem. It asks a smaller, sharper > question: can an agent act usefully in three-body dynamics by reading the > signatures around instability instead of reconstructing the full state? Short version: > Three-body dynamics are chaotic. Sundog asks whether the shadow of that chaos > is enough to steer by. Avoid: - "Sundog solves the three-body problem." - "The controller predicts chaotic dynamics." - "Compressed observables replace integration." - "This proves general control under chaos." ## Roadmap ### Phase 0 - Scope and Literature Pass Goal: define the exact claim before writing the simulation. Deliverables: - A one-page claim boundary for the three-body study. - A table of candidate observables with required state access. - A chosen toy regime: start with planar circular restricted three-body problem before attempting full spatial CR3BP or general three-body scattering. - Baselines: privileged oracle, naive local controller, and passive/no-control reference. Exit criterion: we know which signals are legitimately partial and which are only full-state diagnostics. ### Phase 1 - Diagnostic Benchmark Goal: test whether compressed signals forecast useful events at all. Use full simulator state offline to compute candidate signatures, then measure whether they predict: - ejection or escape; - close approach; - hierarchy break; - transition into chaotic scattering; - departure from a desired orbit family. Metrics: - lead time before event; - precision/recall or AUROC for event detection; - robustness under noise; - failure regimes where signals become ambiguous. Exit criterion: at least one compressed signature gives nontrivial warning or classification performance against a simple baseline. ### Phase 2 - Sensor-Available Surrogate Goal: replace privileged diagnostics with locally measurable proxies. Candidate local signals: - tidal tensor estimate from small probe maneuvers or local acceleration differences; - own acceleration residual after subtracting commanded thrust; - short-horizon response curve to small thrust pulses; - bearing-only or range-rate-only measurements if the scenario allows a weak exteroceptive sensor. Exit criterion: a proxy signal preserves enough of the Phase 1 diagnostic value to justify control. ### Phase 3 - Scan/Seek/Track Controller Goal: build the Sundog controller only after the signal earns it. SCAN: perturb with small maneuvers and estimate local response of the indirect signal. SEEK: move toward regions where the signature indicates stability, capture, or lower escape risk. TRACK: maintain the desired signature using extremum seeking or receding-horizon updates without privileged full-state feedback. Metrics: - time maintained near target orbit family; - fuel or delta-v cost; - recovery after perturbation; - comparison to oracle and naive baselines; - known failure boundary. Exit criterion: the controller succeeds in a bounded operating envelope and loses cleanly where the signal-action coupling collapses. ### Phase 3.5 - Real-time Web Projection Goal: port the restricted three-body diagnostic to an interactive browser visualization before finalizing the public artifact. Deliverables: - A standalone HTML page (`threebody.html`) implementing the restricted three-body problem in JavaScript with real-time integration. - Canvas rendering showing: - the simulated three-body system (two primaries and one test particle); - shadow trails visualizing orbital history; - indirect signature overlays (virial ratio, inertia tensor eigenvalues, pairwise energies) as time-series graphs; - control overlays showing scan/seek/track phases if Phase 3 work is ready. - Visual design consistent with the site hero graph (parhelion-canvas style): gold accents, dark gradient backgrounds, clean typography, responsive layout. - UI controls for interactive exploration: - adjust initial conditions (positions, velocities); - modify system parameters (mass ratios, circular restricted vs full 3-body); - toggle signature visibility; - apply scan/seek/track control moves manually if Phase 3 controller exists, or simulate passive observation if not. - The page serves as the application landing page for the three-body experiment and directly supports Phase 4's public demonstration goal. Exit criterion: a working browser experiment demonstrates that indirect signatures respond coherently to changes in system state, providing a visual benchmark for the coupling claim before the full controller is built. Rationale: A public, visual artifact aligns with the Phase 4 objective and establishes a concrete demonstration environment early. The browser experiment can evolve from diagnostic (Phase 1/2 work) to interactive controller (Phase 3 integration) to final public artifact (Phase 4 refinement) without requiring a separate build for each stage. ### Phase 4 - Public Artifact Goal: turn the result into a strong, honest Sundog hook. Deliverables: - a small interactive visualization showing full orbit, hidden state, exposed indirect signal, and controller action; - a short writeup titled around "steering by the shadow of chaos"; - explicit claim boundary using the same discipline as the photometric experiment; - a failure-boundary panel, not a footnote. ## Validation Roadmap The browser prototype is useful, but it is not yet a completed experiment. The next phases turn the artifact into evidence by adding baseline comparisons, event labels, quantitative metrics, and a mapped operating envelope. ### Phase 5 - Reproducible Experiment Harness Goal: move from manual browser exploration to repeatable trial batches. Deliverables: - A deterministic simulation harness that can run the same dynamics outside the animation loop. - Seeded initial-condition generation for stable, near-escape, near-collision, and chaotic-scattering regimes. - A trial manifest format recording mass ratio, initial state, controller mode, thrust authority, target tidal magnitude, timestep, duration, and seed. - Per-trial logs for state, exposed signals, controller actions, event labels, and terminal outcome. Exit criterion: a trial can be replayed exactly, and the browser demo can point to the same equations as the batch harness. ### Phase 6 - Baseline Set Goal: establish what the Sundog controller is being compared against. Required baselines: - **Passive baseline:** no thrust; measures natural survival, escape, and close approach rates. - **Naive local baseline:** uses only simple local acceleration magnitude or radial heuristics, without tidal-gradient structure. - **Privileged oracle:** uses full simulator state, primary positions, and a chosen target objective. - **Ablation controllers:** SCAN-only, SEEK-only, TRACK-only, and tidal-signal shuffled/noised variants. Metrics: - time before escape or close approach; - time inside desired tidal band; - fuel or delta-v cost; - number of controller saturations; - terminal distance/energy class; - recovery after injected perturbations. Exit criterion: every claimed improvement is stated relative to at least one baseline, and every baseline uses the same initial-condition slate. ### Phase 7 - Event Labels and Diagnostic Metrics Goal: test whether compressed signatures actually forecast useful events. Event labels: - escape/ejection; - close approach; - tidal spike; - controller saturation; - loss of target tidal band; - numerical instability or invalid trajectory. Diagnostic metrics: - lead time between warning threshold and event; - precision, recall, F1, and AUROC for event detection; - false alarm rate per unit simulated time; - robustness under sensor noise; - calibration curves for warning thresholds. Exit criterion: at least one privileged diagnostic and one sensor-motivated proxy show quantified predictive value over a naive threshold baseline. ### Phase 8 - Sensor-Model Validation Goal: decide how much of the current "sensor-limited" mode is genuinely available to the controlled body. Sensor variants: - **Simulated local probe:** current browser implementation; nearby acceleration samples are computed directly from the field model. - **Micro-maneuver estimate:** the controller applies small thrust probes and estimates the response from its own acceleration history. - **Accelerometer-array estimate:** multiple local samples are modeled as if measured by a short-baseline physical sensor. - **Noisy learned surrogate:** a learned or fitted local field model receives only prior local observations. Audit questions: - Does the controller need hidden primary positions to compute the signal? - How much delay and noise can the proxy tolerate? - Does probe thrust contaminate the signal enough to change the control result? - Which sensor assumptions are plausible for spacecraft, simulation tooling, or pure software agents? Exit criterion: the writeup labels each result by sensor tier, and no result computed from full state is described as sensor-only. ### Phase 9 - Operating Envelope and Failure Map Goal: replace anecdotal failure boundaries with a map. Current command: ```bash npm run threebody:phase9 npm run threebody:phase9:refine npm run threebody:phase9:lock npm run threebody:phase9:axes npm run threebody:phase9:hazard npm run threebody:phase9:hazard:refine ``` Sweep axes: - initial position and velocity of the test particle; - mass ratio of the primaries; - thrust authority; - target tidal magnitude; - sensor noise and delay; - integration timestep; - perturbation impulse size. Outputs: - `results/threebody/phase9-operating-envelope/manifest.json`; - `trial-outcomes.csv`: primary cold-open data artifact. One row per non-passive controller trial with seed, regime, initial-condition scales, thrust, sensor noise, guard thresholds, terminal outcome, simulated time, delta-v, minimum primary distance, matched passive outcome/time, outcome effect, and failure mechanism. - `paired.csv`: legacy/internal per-trial controller rows paired against matched passive baselines. Downstream aggregates should be derivable from `trial-outcomes.csv`; - `envelope-map.csv`: one row per regime, initial-condition scale, thrust, sensor-noise, and guard setting, including survival deltas, passive survival time, controller effort, sensor-error summaries, and dominant failure mechanism; - `aggregate-envelope.csv`: same map aggregated across regimes; - `best-by-cell.csv`: best setting per regime/radius/velocity cell; - `cell-class-map.csv`: map-shaped view of best region class by radius and velocity; - `cell-delta-map.csv`: map-shaped view of best survival delta by radius and velocity; - `candidate-envelope.csv`: positive survival-delta candidates that pass the current worsened-rate filter; - success/failure heatmaps; - escape and close-approach regions; - controller saturation regions; - cases where tidal proxies are misleading; - representative trajectories for wins, losses, and ambiguous regimes. First Phase 9 smoke map: - Default run emitted 1,944 trials over stable, near-escape, and chaotic regimes, three radius scales, three velocity scales, two thrust limits, three accelerometer-noise levels, and two guard-acceleration thresholds. - Outcome counts: 972 bounded, 879 close approach, 93 escape. - Positive candidate rows: 29 of 324 `envelope-map.csv` rows after tightening the candidate definition to require positive survival delta. - Best-cell summary: 4 promising cells, 21 neutral cells, and 2 risky cells. - Interpretation: the current positive pocket is near-escape, especially around nominal/slightly-larger radius scales and low-to-moderate velocity scaling. Stable cells are mostly neutral because passive already survives; chaotic cells are mostly neutral or risky. This is the right shape for an operating envelope, but it is not yet a public performance claim. Refined near-escape map: - `npm run threebody:phase9:refine` emits a denser 7x7 near-escape radius/velocity map over radius scales `0.925..1.075`, velocity scales `0.85..1.15`, thrust limits `0.4/0.5`, sensor noise `0/0.01`, guard acceleration `2.5`, and 6 seeds. - Default refined run emitted 2,352 trials. - Positive candidate rows: 116 of 196 `envelope-map.csv` rows. - Best-cell summary: 32 promising cells, 6 neutral, 4 mixed, 3 risky, and 4 negative. - The map is no longer just anecdotal: the strongest connected positive region is the upper-right of the near-escape slice, roughly radius scale `>= 0.975` and velocity scale `>= 0.95`, with especially strong deltas above velocity scale `1.05`. Larger-radius but low-velocity cells are negative/risky, which gives the first crisp local failure boundary. - Mechanistic labels now separate `control_effort_or_saturation` from `controller_destabilized_or_shortened_passive`. In the refined map, the negative/risky rows are mostly the second mechanism: passive trajectories often last longer or survive, while guarded TRACK shortens or destabilizes them. A smaller subset shows high controller effort/saturation. No refined failure rows currently point primarily to the accelerometer sensor-noise floor. Locked near-escape map: - `npm run threebody:phase9:lock` repeats the refined 7x7 near-escape map with 24 seeds, producing 9,408 trials under `results/threebody/phase9-locked-near-escape/`. - Positive candidate rows: 132 of 196 `envelope-map.csv` rows. - Best-cell summary: 41 promising cells, 7 mixed cells, and 1 negative cell. - The connected positive pocket survives the larger seed slate. The strongest rows are still at larger radius and higher velocity: radius scale `1.05..1.075` and velocity scale `1.1..1.15` often reach survival deltas of `0.79..0.92` versus matched passive baselines. - Failure mechanisms remain honest: harmful trial rows split into 585 `controller_destabilized_or_shortened_passive` and 359 `control_effort_or_saturation`. This says the remaining boundary is mostly controller interaction with trajectories passive already handles, not a detected sensor-noise-floor failure. Mass-ratio and timestep probe: - `npm run threebody:phase9:axes` adds the first scoped mass-ratio/timestep check inside the locked high-velocity near-escape pocket. It sweeps mass ratios `0.01`, `0.3`, and `1`, timesteps `0.008`, `0.01`, and `0.012`, radius scales `1.025..1.075`, velocity scales `1.05..1.15`, thrust `0.4`, sensor noise `0`, and 8 seeds. - Default axes run emitted 1,296 trials under `results/threebody/phase9-mass-timestep/`. - Positive candidate rows: 72 of 81 `envelope-map.csv` rows. - Best-cell summary by mass/timestep: every equal-mass cell is promising; mass ratio `0.3` has 8 promising and 1 mixed cell per timestep; mass ratio `0.01` has 7 promising and 2 mixed cells per timestep. - Interpretation: within this already-favorable high-velocity near-escape pocket, the effect is not obviously an equal-mass resonance and is not sensitive to the small timestep band tested. The lower mass-ratio maps often show larger mean survival deltas because passive escapes more often there. This does not yet prove robustness outside the locked pocket. Hazard-derived guard gates: - `npm run threebody:phase9:hazard` replaces the hand-tuned guard constants in the same mass-ratio/timestep slice with thresholds derived from passive hazard-score samples. For each case/regime, the passive trials provide non-hazard samples; the controller then uses the configured quantile (`--track-guard-quantile 0.75`) for local acceleration and tidal magnitude guard thresholds, plus the complementary radius quantile for the radius gate. - Default hazard-gate run emitted 1,296 trials under `results/threebody/phase9-hazard-gates/`. - Positive candidate rows: 81 of 81 `envelope-map.csv` rows. - Outcome effects: 522 helped, 36 time-helped, 90 tied, and no hurt/time-hurt rows in the scoped run. - Failure mechanisms: no harmful rows were labeled in the hazard-gate run. - Delta-v check: average controller delta-v fell from about `2.90` in the constant-gate mass/timestep probe to about `1.74` in the hazard-gate probe. - Interpretation: inside the locked high-velocity near-escape pocket, passive hazard-score gates improve the controller story: they preserve the pocket, remove the harmful rows seen with hand-tuned constants, and reduce control effort. This is still scoped to the favorable pocket and should next be tested against the full refined near-escape grid. Full refined hazard-gate check: - `npm run threebody:phase9:hazard:refine` applies the same passive quantile-derived guard gates to the full 7x7 refined near-escape grid. - Default run emitted 2,352 trials under `results/threebody/phase9-hazard-refined-near-escape/`. - Positive candidate rows: 100 of 196 `envelope-map.csv` rows, versus 116 of 196 in the constant-gate refined run. - Best-cell summary: 27 promising, 8 neutral, 7 mixed, 5 risky, and 2 negative. - Delta-v check: average controller delta-v fell from about `2.07` in the constant-gate refined run to about `1.12` in the hazard-gate refined run. - Interpretation: hazard-derived gates are not simply better everywhere. They preserve the high-velocity positive pocket and sharply reduce control effort, but they shrink the wider candidate region and still leave low-velocity harms. This is a stronger, more honest controller: less tuned and less aggressive, but still bounded by the same broad operating-envelope geometry. Exit criterion: the public page can show where the method works, where it fails, and where the result is inconclusive. ### Phase 10 - Writeup and Claim Ratchet Goal: update the public-facing story so it reflects the Phase 9 result without turning a bounded operating-envelope pocket into a general control claim. Immediate work: - Update `docs/threebody-writeup.md` so the headline status no longer says only "promising enough to study." The earned claim is now that a guarded accelerometer-proxy TRACK controller shows a reproducible positive pocket in mapped near-escape regimes, with explicit low-velocity failure boundaries. - Update `threebody.html` copy and the applications gallery entry to make the current result inspectable from the public site. - Add a short "What changed after Phase 9" panel: initial smoke pocket, refined near-escape pocket, locked 24-seed result, mass-ratio/timestep slice, and hazard-gate refinement. - Promote the failure map to first-class narrative material. The low-velocity harm boundary and controller-shortened-passive rows are part of the result, not a footnote. - Keep the sensor-tier wording strict: the current best controller is accelerometer-proxy guarded TRACK in simulation, not a validated physical spacecraft sensor stack. Current safe claim: > The three-body page is an interactive prototype showing indirect signatures, > sensor-model separation, and a scan/seek/track controller scaffold. Claim after Phase 7: > In the tested planar restricted setup, selected compressed signatures forecast > escape or close-approach events with quantified lead time. Claim after Phase 8: > In the tested setup, a specified sensor-tier proxy preserves enough diagnostic > signal to support local control experiments. Claim after Phase 9: > In a mapped operating envelope, the Sundog-style controller outperforms > passive and naive local baselines on specified metrics, while losing to or > approaching a privileged oracle depending on regime. Current Phase 9 earned wording: > In the tested planar restricted setup, a guarded accelerometer-proxy TRACK > controller improves survival over passive behavior in a connected > near-escape operating pocket. The effect is not global: low-velocity > near-escape cells still show harms, and hazard-derived gates trade a smaller > candidate region for lower control effort. Current Phase 11 earned wording: > In the tested planar restricted setup, the guarded accelerometer-proxy TRACK > controller improves survival over passive and naive local baselines in a > robust high-velocity near-escape pocket. The result is not global: lower > velocity and equal-mass boundary cells still expose controller harms, mostly > through controller-shortened passive survival and control effort/saturation. Do not claim: - "Sundog solves three-body dynamics." - "The controller predicts chaos." - "Sensor-only control is validated" before Phase 8 passes. - "Comparable utility from partial information" before baseline metrics show the trade. Exit criterion: the writeup, page copy, and promo language all use the strongest claim actually earned by the completed validation phase, and no stronger one. ### Phase 11 - Robustness and Outside-Pocket Expansion Goal: find out whether the Phase 9 pocket is a real operating-envelope feature or a narrow artifact of one guard quantile, near-escape slice, and controller parameterization. Primary checks: - Guard-quantile sweep: compare passive hazard gates at quantiles `0.5`, `0.75`, and `0.9` on the refined near-escape grid. - Outside-pocket expansion: rerun mass-ratio/timestep probes beyond the locked high-velocity pocket, especially lower velocity scales and larger-radius low-velocity cells where the current controller harms passive outcomes. - Diagnostic-to-control chain: summarize Phase 7/8 metrics alongside Phase 9 outcomes so the public claim reads as a sequence: hazard signatures forecast events, sensor proxies preserve usable signal, controller improves outcomes only where signal-action coupling remains intact. - Failure mechanism audit: separate controller-destabilized cases from pure control-effort/saturation cases and check whether hazard gates reduce one class while worsening another. - Naive/oracle comparison pass: rerun the best Phase 11 pocket against passive, naive local acceleration, guarded accelerometer TRACK, and the privileged heuristic oracle on the same slate. Suggested commands/scripts: ```bash npm run threebody:phase11:guard-quantiles npm run threebody:phase11:outside-pocket npm run threebody:phase11:compare ``` These commands wrap `scripts/threebody-operating-envelope.mjs` with explicit output directories and manifest labels, rather than replacing the Phase 9 artifacts. The runner supports `--track-guard-quantiles` as a sweep axis for Phase 11 guard comparisons. Outputs: - `results/threebody/phase11-guard-quantiles/`; - `results/threebody/phase11-outside-pocket/`; - `results/threebody/phase11-comparison/`; - `docs/THREEBODY_PHASE11_SUMMARY.md`; - updated failure-boundary heatmaps for velocity/radius, mass ratio, timestep, and guard quantile. Exit criterion: the project can say whether the Phase 9 positive pocket is robust to guard quantile and nearby parameter expansion, and can identify the first outside-pocket region where the Sundog controller should not be used. Initial Phase 11 results: - Guard-quantile sweep emitted 7,056 trials under `results/threebody/phase11-guard-quantiles/`. The pocket survives quantiles `0.5`, `0.75`, and `0.9`; `0.5` is cheapest, `0.9` has the strongest average survival delta, and `0.75` remains a viable middle setting. - Outside-pocket expansion emitted 6,912 trials under `results/threebody/phase11-outside-pocket/`. The high-velocity pocket remains strong, while lower velocity scales and equal-mass boundary cells carry most of the mixed and negative rows. - Comparison run emitted 2,592 trials under `results/threebody/phase11-comparison/`. In the favorable high-velocity pocket, guarded accelerometer TRACK produced 81 candidate rows out of 81, while the naive local baseline produced none and the privileged heuristic oracle produced 34 of 81. This oracle is a heuristic, not an optimal controller. - Summary: `docs/THREEBODY_PHASE11_SUMMARY.md`. Phase 11 conclusion: the Phase 9 pocket is no longer just a tuned positive slice. It survives guard-quantile variation and matched baseline comparison in the favorable high-velocity near-escape region. The same runs also make the failure boundary sharper: lower velocities and equal-mass boundary cells are the first places where the controller should not be used. ## Recommendation Proceed, but make the first milestone diagnostic rather than controller-first. The idea is promising because it gives Sundog a high-recognition physics hook without needing a grandiose claim. The research discipline is to earn three things in order: first predictive signatures, then sensor-available proxies, then control. ## Implementation Status (2026-05-07) **Phase 1-3**: Prototype implemented, validation pending. The interactive browser visualization at `threebody.html` now scaffolds the diagnostic signals, sensor-limited proxy mode, and SCAN/SEEK/TRACK controller modes. It does not yet complete the benchmark requirements for event prediction, oracle comparison, or failure-boundary mapping. **Phase 4**: Draft public artifact live, still evidence-bounded. It includes: - Interactive visualization with real-time RK4 integration - Canvas rendering of system state with orbital trails - Indirect signature overlays (virial, inertia, energy, tidal tensor) - Phase 2 sensor-limited mode toggle - Phase 3 controller implementation (SCAN/SEEK/TRACK) - UI controls for exploration and experimentation - Visual design consistent with site branding **Public writeup**: See [threebody-writeup.md](threebody-writeup.md) for the complete "Steering by the Shadow of Chaos" documentation, including: - Claim boundaries and hypothesized failure modes - Sensor model audit (privileged vs. sensor-available signals) - Phase-by-phase implementation details - Connection to photometric alignment experiment - Future directions **Phase 5**: Complete for the prototype tier. A deterministic harness now lives at: ```bash npm run threebody:phase5 ``` Current Phase 5 implementation: - `public/js/threebody-core.mjs`: shared planar restricted three-body dynamics, signature computation, tidal proxy computation, prototype controller logic, event labels, and trial runner. - `public/js/threebody-browser.mjs`: browser visualization wrapper that imports the shared core module instead of carrying mirrored dynamics inline. - `scripts/threebody-harness.mjs`: CLI batch harness that writes a manifest and per-trial JSONL logs. - Default smoke output: `results/threebody/phase5-smoke/` (ignored by git). Current status against Phase 5 exit criterion: - Deterministic replay: smoke trial logs reproduce byte-for-byte across reruns. - Manifest format: present, including seed, regime, controller mode, initial state, mass ratio, timestep, duration, thrust authority, target tidal magnitude, log path, and terminal summary. - Per-trial logs: present, including state, exposed signatures, tidal tensor, thrust vector, event labels, and terminal outcome. - Browser equation sharing: complete. Both the browser and harness now use `public/js/threebody-core.mjs` for initialization, integration, signatures, tidal proxy computation, and prototype controller thrust. This does not make baseline or metric claims. It only completes the Phase 5 reproducibility substrate needed for Phase 6 and Phase 7. **Phase 6**: Started. Baseline modes and a matched-slate runner now exist: ```bash npm run threebody:phase6 ``` Current Phase 6 implementation: - Passive baseline: `off`, no thrust. - Naive local baseline: `naive`, pushes against local acceleration plus simple velocity damping, without tidal-gradient structure. - Sundog ablations: `scan`, `seek`, and `track`. - Tidal-signal ablations: `seek_noisy`, `track_noisy`, `seek_shuffled`, and `track_shuffled`. The noisy variants perturb observed tidal magnitude and gradient; the shuffled variants preserve controller wiring but replace the tidal-gradient direction with a deterministic pseudo-random direction. - Privileged baseline: `oracle`, a full-state lookahead guard that scores candidate thrust directions using simulator state. This is a privileged heuristic, not an optimal controller. - Browser controls expose `naive`, `oracle`, and the tidal-signal ablation modes for qualitative inspection. - Harness output includes: - `summary.csv`: grouped by regime and controller mode. - `paired.csv`: per-regime/per-seed deltas against matched passive and oracle rows. - `comparison.csv`: aggregate improvement/tie/worsening counts versus passive, plus mean time and fuel deltas. - The manifest declares baseline mode definitions and primary metrics: `terminalOutcome`, `simulatedTime`, `totalDeltaV`, `minPrimaryDistance`, `saturationCount`, `targetBandLossCount`, and initial Phase 7 tidal-warning diagnostics. - Default smoke output: `results/threebody/phase6-baselines/` (ignored by git). Current status against Phase 6 exit criterion: - Matched slate: present for all current modes across stable, near-escape, near-collision, and chaotic seed regimes. - Comparison artifacts: present via manifest, JSONL logs, summary CSV, paired CSV, and aggregate comparison CSV. - Interpretation: not yet complete. The initial 3-seed smoke slate exposes hard regimes and several weak controllers; it should be treated as calibration for baseline design, not as evidence for or against the Sundog claim. - Next Phase 6 work: tune the privileged baseline objective, define a larger seed slate, and decide whether the primary metric ordering should privilege survival time, bounded outcome, or fuel cost. **Phase 7**: Started. The harness now computes event-diagnostic metrics from the per-step event history: ```bash npm run threebody:phase7 ``` Current Phase 7 implementation: - Hazard label: first escape or close-approach event. - Warning scores: - `tidal_magnitude`: Frobenius norm of the local tidal tensor estimate. - `local_acceleration`: magnitude of local gravitational acceleration at the test particle; this is the naive local warning baseline. - Default warning labels: first tidal-spike event under the configured `tidalSpikeThreshold`, and first local-acceleration warning under `localAccelerationWarningThreshold`. - Warning window: samples within `eventWarningHorizon` seconds before a hazard. - Threshold sweeps: configurable `thresholdSweep` values for tidal magnitude and `localAccelerationThresholdSweep` values for local acceleration, with per-trial and aggregate precision/recall/F1/false-alarm rows. - Calibration curves: quantile-binned warning scores, with observed event-window rate per regime, controller mode, and score type. - Per-trial diagnostics: - first hazard time; - first tidal-warning time; - warning lead time; - true/false positive and true/false negative sample counts; - precision, recall, F1, false-alarm rate, and AUROC using tidal magnitude as the warning score; - the same threshold metrics and AUROC for local acceleration. - Harness output now includes `event-metrics.csv` alongside `summary.csv`, `paired.csv`, and `comparison.csv`. - Additional Phase 7 outputs: - `threshold-sweep.csv`: one row per trial, score type, and threshold. - `threshold-summary.csv`: aggregate threshold metrics by regime, controller mode, score type, and threshold. - `calibration.csv`: binned warning-score calibration rows. - Default smoke output: `results/threebody/phase7-event-metrics/` (ignored by git). Current status against Phase 7 exit criterion: - Event plumbing: present for both tidal magnitude and local acceleration. - Threshold sweeps and calibration rows: present, but only for the current small smoke slate. - Interpretation: not yet complete. Some regimes are label-degenerate under the one-second warning horizon, so AUROC or false-alarm rate may be blank when a trial has no positive/negative comparison class. Calibration bins may also be non-monotone, which means neither raw score is yet a calibrated risk probability. - Next Phase 7 work: add reliability/error summaries for the calibration bins, tune threshold grids per regime, and run a larger seed slate before ratcheting any public claim. **Phase 8**: Started. The harness now emits a sensor-model audit for the tidal tensor proxy: ```bash npm run threebody:phase8 ``` Current Phase 8 implementation: - Sensor tiers: - `simulated_local_probe`: exact virtual local probe samples from the simulator field model; this is the current reference tier. - `accelerometer_array_noisy`: short-baseline acceleration samples with deterministic Gaussian noise. - `delayed_local_probe`: exact virtual local probe estimate delayed by the configured number of integration samples. - `micro_maneuver_noisy`: larger-baseline probe with delay and extra maneuver-contamination noise. - Sensor-tier controller modes: - `seek_sensor_accel`, `track_sensor_accel` - `track_sensor_accel_guarded` - `seek_sensor_delayed`, `track_sensor_delayed` - `seek_sensor_micro`, `track_sensor_micro` - Harness options: - `--sensor-variants` - `--sensor-audit-every` - `--sensor-noise-std` - `--sensor-delay-steps` - `--micro-maneuver-contamination-std` - Calibration sweep: - `npm run threebody:phase8:calibrate` - `npm run threebody:phase8:analyze` - `npm run threebody:phase8:focus` - Dedicated script: `scripts/threebody-sensor-calibration-sweep.mjs`. - Analyzer script: `scripts/threebody-analyze-sensor-calibration.mjs`. - Sweeps accelerometer-array noise, delayed-probe latency, and micro-maneuver noise/latency/contamination with matched passive baselines for each regime and seed. - Default output: `results/threebody/phase8-calibration-sweep/` (ignored by git). - Focused accelerometer output: `results/threebody/phase8-focused-accelerometer/` (ignored by git). - Additional Phase 8 outputs: - `sensor-model-samples.csv`: per-trial/per-time sensor estimate rows. - `sensor-model-summary.csv`: aggregate tensor magnitude and component errors by regime, controller mode, and sensor variant. - Additional calibration outputs: - `paired.csv`: per-trial sensor-controller result paired against the passive baseline for the same degradation setting, regime, and seed. - `summary.csv`: outcome survival, passive deltas, delta-v, minimum-distance, and warning-score aggregates by sensor tier and degradation setting. - `sensor-error-summary.csv`: sensor-estimate magnitude/component error by tier, degradation setting, regime, and controller mode. - `envelope.csv`: analyzer output ranking controller/degradation settings by survival delta, worsened rate, and sensor error. - `regime-envelope.csv`: analyzer output preserving the per-regime failure boundary. - `candidate-envelope.csv`: analyzer output for settings that meet the current strict candidate filter. - Default smoke output: `results/threebody/phase8-sensor-model/` (ignored by git). Current status against Phase 8 exit criterion: - Sensor-tier labels: present in the manifest and summary outputs. - Reference separation: present. The exact virtual probe is explicitly labeled as a simulator-field reference, not as a hardware-valid sensor claim. - Default calibration sweep: present. The first run emitted 1,044 matched calibration trials; after adding guarded accelerometer TRACK, the default run emits 1,104 matched calibration trials. It writes paired passive-baseline comparisons, aggregate survival rows, and sensor-error rows for accelerometer noise, delayed-probe latency, and micro-maneuver degradation. - Focused accelerometer check: present. The follow-up 12-seed accelerometer run emitted 960 trials with SEEK, TRACK, and guarded TRACK. The initial 3-seed candidate for ungated `track_sensor_accel` at noise `0.03` did not survive the larger slate. The guarded accelerometer TRACK variant is now the only focused candidate: noise `0` and `0.01` improve aggregate survival versus the matched passive baseline while keeping worsened rate under the strict filter. This is a Phase 9 hypothesis, not a validated claim. - Interpretation: calibration only. The initial smoke run shows noisy accelerometer-array estimates preserve the reference tensor far better than delayed or contaminated micro-maneuver proxies in fast-changing regimes. When those degraded estimates drive SEEK/TRACK, accelerometer-array control mostly tracks the ideal behavior, while delayed and micro-maneuver variants break more stable cases. The first calibration sweep keeps the same broad shape: delayed and micro-maneuver tiers degrade quickly with latency, while accelerometer-array error grows with noise. Some noisier controller settings also produce better terminal survival on the tiny slate by changing thrust behavior, so these rows should be treated as operating-envelope clues, not as evidence that noise helps. **Phase 9**: Started. The first operating-envelope runner maps guarded accelerometer TRACK across initial-condition, thrust, sensor-noise, and guard settings: ```bash npm run threebody:phase9 ``` Current Phase 9 implementation: - Dedicated script: `scripts/threebody-operating-envelope.mjs`. - Default output: `results/threebody/phase9-operating-envelope/` (ignored by git). - Default smoke map: 1,944 trials across stable, near-escape, and chaotic regimes; three radius scales; three velocity scales; two thrust limits; three accelerometer-noise levels; and two guard-acceleration thresholds. - Outputs: - `trial-outcomes.csv` - `paired.csv` - `envelope-map.csv` - `aggregate-envelope.csv` - `best-by-cell.csv` - `cell-class-map.csv` - `cell-delta-map.csv` - `candidate-envelope.csv` - First result: 29 positive candidate rows out of 324 envelope rows. The best-cell table marks 4 cells promising, 21 neutral, and 2 risky. The positive pocket is near-escape; stable cells are mostly neutral and chaotic cells are mostly neutral or risky. - Refined near-escape result: `npm run threebody:phase9:refine` emits a denser 7x7 radius/velocity map with 2,352 trials. The best-cell matrix shows 32 promising cells, 6 neutral cells, 4 mixed cells, 3 risky cells, and 4 negative cells. The positive pocket is connected at moderate-to-high velocity scales; larger-radius low-velocity cells are the local failure boundary. - Mechanism read: `trial-outcomes.csv` is now the primary data artifact, and `envelope-map.csv` carries dominant failure mechanisms. In the refined run, harmful trial rows split into 119 `controller_destabilized_or_shortened_passive` and 65 `control_effort_or_saturation`; negative/risky map rows are mostly controller-shortened passive cases. - Locked near-escape result: `npm run threebody:phase9:lock` emits 9,408 trials over the same refined 7x7 pocket with 24 seeds. The pocket survives: 41 best-cell rows are promising, 7 are mixed, and 1 is negative. The strongest survival deltas remain at larger radius and higher velocity. - Mass-ratio/timestep result: `npm run threebody:phase9:axes` emits 1,296 trials inside the locked high-velocity pocket. The pocket survives mass ratios `0.01`, `0.3`, and `1` across timesteps `0.008`, `0.01`, and `0.012`, with 72 candidate rows out of 81. This supports the pocket as broader than the equal-mass/default-timestep geometry, within the scoped slice. - Hazard-gate result: `npm run threebody:phase9:hazard` replaces the hand-tuned guard constants in the mass-ratio/timestep slice with passive quantile-derived hazard-score gates. The scoped run keeps 81 candidate rows out of 81, records no harmful trial rows, and lowers average delta-v relative to the constant gate probe. - Full refined hazard-gate result: `npm run threebody:phase9:hazard:refine` keeps 100 candidate rows out of 196, compared with 116 under constant gates, while reducing average delta-v from about `2.07` to `1.12`. It preserves the high-velocity pocket but does not remove the low-velocity failure boundary. - Phase 11 robustness result: `npm run threebody:phase11:guard-quantiles` confirms the pocket survives guard quantiles `0.5`, `0.75`, and `0.9`, with a clear survival/effort trade. `0.5` is cheapest, `0.9` has the largest average survival delta, and `0.75` remains a viable middle setting. - Phase 11 outside-pocket result: `npm run threebody:phase11:outside-pocket` expands beyond the favorable high-velocity slice. It finds 96 promising, 36 mixed, 9 negative, and 3 neutral best cells, with most harms concentrated at lower velocity scales and equal mass ratio. - Phase 11 comparison result: `npm run threebody:phase11:compare` reruns the favorable pocket against passive, naive local acceleration, guarded accelerometer TRACK, and a privileged heuristic oracle. Guarded TRACK produces 81 candidate rows out of 81, the naive local baseline produces none, and the heuristic oracle produces 34 of 81. - Next work: make the Phase 7/8 diagnostic-to-control chain visible in the public figures, then decide whether to run a larger seed lock on the outside-pocket boundary or start the spatial/3D extension. **Interactive demonstration**: [threebody.html](../threebody.html)