# Three-Body Phase 15C - Multi-Step Counterfactual Horizon Audit Spec Phase 15 ended as **Fail-Magnitude**: the guarded TRACK survival envelope was stable across the precision ladder, and ablations collapsed or inverted, but the privileged one-step counterfactual missed the pre-registered magnitude bar. Phase 15B then showed that the `1e-9` denominator floor is real and near-universal, but it is **not** the full explanation: floor-hit steps carry no TRACK-specific hidden positive one-step signal, and shuffled arms can score higher on floor-hit positive rate while failing to produce candidate cells. Phase 15C is the pre-registered follow-on named by Phase 15B: a diagnostic multi-step counterfactual horizon audit. It asks whether TRACK's favorable pocket is explained by cumulative trajectory steering over short horizons rather than by one-step energy reduction. Phase 15C does **not** revise the Phase 15 verdict, retune the controller, change guard thresholds, or upgrade the earned claim. It is a mechanism diagnostic only. ## 1. Decision Lock - **Frozen controller.** Reuse the Phase 13/15 guarded TRACK controller, Phase 13 passive-derived guard thresholds, thrust limit, sensor settings, and action coupling. No guard retuning, thrust retuning, sensor retuning, or new controller mode is admitted. - **Same cell slate.** Reuse the Phase 15B lock slate: `near_escape`, `dt=0.004`, mass ratios `0.01,0.3,1`, radius scales `1.025,1.05,1.075`, velocity scales `0.95,1.05,1.1,1.15`, thrust limit `0.4`, sensor noise `0`, hazard quantile guard `0.75`, eight seeds, duration `16`. - **Same mode slate.** Reuse the Phase 15B audit modes: `off`, `track_sensor_accel_guarded`, `track_sensor_accel_signal_delay`, `track_sensor_accel_signal_shuffle`, `track_sensor_accel_action_shuffle`, and `track_sensor_accel_sign_flip`. - **Audit only.** The audit is additive. Default Phase 13/14/15 paths remain unchanged. New receipts must be written only behind a new explicit multi-step audit flag. - **Horizon set.** The frozen horizons are simulation-step counts `N in {4, 8, 16, 32}`. With `dt=0.004`, these correspond to `0.016, 0.032, 0.064, 0.128` seconds of lookahead. - **Primary stratification.** Read results by `candidateEnvelope=true/false`, with `signal_delay` candidate vs non-candidate rows elevated as a primary diagnostic because Phase 15B found their floor-step sign behavior diverged. - **No claim upgrade.** A positive Phase 15C read can support a mechanism explanation for Phase 15's survival pocket; it cannot convert Phase 15 to Pass or broaden the gravity claim. ## 2. Scope Phase 15C owns: - additive multi-step counterfactual audit receipts in `public/js/threebody-core.mjs`; - additive CSV columns in `paired.csv`, `trial-outcomes.csv`, and `aggregate-envelope.csv`; - future npm scripts: - `npm run threebody:phase15c:multistep-smoke` - `npm run threebody:phase15c:multistep` - outputs under: - `results/threebody/phase15c-multistep-counterfactual-smoke/` - `results/threebody/phase15c-multistep-counterfactual-lock/` - this spec and a future `PHASE15C_RESULTS.md`. Phase 15C does not own alternate hazard scores, warning-quality reruns, spatial/3D extension, spacecraft-domain extension, controller redesign, or any public-copy upgrade. ## 3. Command Shape The following command names are reserved but not runnable until an implementation commit adds the flag, CSV columns, and npm scripts. Smoke: ```bash npm run threebody:phase15c:multistep-smoke ``` The smoke must be capped under the repository's ten-minute rule. Its intended shape is the Phase 15B smoke cell with horizons `4,8,16,32`: `massRatio=1`, `dt=0.004`, radius `1.025`, velocity `1.1`, the six Phase 15B modes, one seed, duration `4`. It is a column-flow and sanity check only. Lock: ```bash npm run threebody:phase15c:multistep ``` The lock uses the Phase 15B lock slate and all four horizons. It is expected to exceed the inline ten-minute rule, so the implementation commit must record a capped rate probe, wall-clock estimate, resume/readback path, and operator-run or long-runner plan before the lock is started. ## 4. Multi-Step Counterfactual Definition For each eligible simulation step in a trial, compute a matched set of counterfactual rollouts from the same current state: - `actual`: the controller-selected action for the current mode at the current step, followed by the mode's frozen policy for the remaining `N - 1` steps; - `noop`: zero thrust at the current step, followed by the same frozen policy for the remaining `N - 1` steps; - `oracleStrict`: the Phase 15 forward-oracle strict action at the current step, followed by the same frozen policy for the remaining `N - 1` steps. For every horizon `N`, report: - raw terminal score difference `score(actual_N) - score(noop_N)`; - raw oracle gap `score(oracleStrict_N) - score(noop_N)`; - normalized score `clamp((score(actual_N) - score(noop_N)) / max(abs(score(oracleStrict_N) - score(noop_N)), 1e-9), -1, 1)`; - positive-rate indicator for the raw terminal difference; - floor-hit indicator for the same `1e-9` normalizer floor used in Phase 15B. The scoring function must be the same terminal-state score family used by the Phase 15 counterfactual receipt unless the implementation amendment discovers that the Phase 15 function is not separable from one-step-only fields; any such change must be filed before running the smoke. ## 5. Metrics Trial-level fields, per horizon `N`: - `counterfactualH{N}EligibleSteps` - `counterfactualH{N}MeanEffectVsNoop` - `counterfactualH{N}MeanAbsEffectVsNoop` - `counterfactualH{N}PositiveRate` - `counterfactualH{N}MeanRawNormalizer` - `counterfactualH{N}NormalizerFloorRate` - `counterfactualH{N}MeanScore` - `counterfactualH{N}FloorPositiveRate` - `counterfactualH{N}NonFloorMeanScore` Aggregate rows report corresponding means/totals by `mode x candidateEnvelope x horizon`. Primary readouts: - **Horizon lift:** TRACK candidate rows show increasing or materially higher normalized score / positive rate at `N in {8,16,32}` than at `N=4`. - **Ablation separation:** TRACK horizon-lift exceeds delayed/shuffled/sign-flip arms on candidate rows. - **Non-candidate control:** non-candidate TRACK rows do not show the same horizon-lift pattern, or are explicitly interpreted as a broad mode effect rather than a candidate-envelope mechanism. - **Signal-delay asymmetry:** `signal_delay` candidate and non-candidate splits are reported separately at every horizon, with no pooling before the primary read. ## 6. Pre-Registered Branches **Multi-step steering supported:** candidate TRACK rows show a monotone or material horizon lift from `N=4` to at least one of `N=16` or `N=32`, and at least two of `signal_delay`, `signal_shuffle`, `action_shuffle`, and `sign_flip` trail TRACK by `>= 0.10` on positive rate or normalized score in the same candidate split. Interpretation: Phase 15's survival pocket is consistent with cumulative trajectory steering not visible to the one-step yardstick. **Multi-step steering rejected:** TRACK candidate rows remain near chance or do not improve with horizon, or shuffled/mistimed arms match or exceed TRACK at all horizons. Interpretation: the Phase 15 survival pocket is not explained by this multi-step counterfactual; look next at the hazard score, event-warning quality, or a controller-design limitation. **Mixed / partial diagnostic:** horizon lift appears only in non-candidate rows, only in `signal_delay`, or only at one horizon without separation from ablation arms. Interpretation: Phase 15C localizes a new diagnostic but does not explain the mechanism. All branches preserve the Phase 15 formal verdict. ## 7. Readback After the smoke, record: - command and wall-clock; - trial count; - presence of all new horizon columns in `paired.csv`, `trial-outcomes.csv`, and `aggregate-envelope.csv`; - per-horizon TRACK candidate and non-candidate read for the smoke cell; - whether the lock command is worth staging unchanged; - capped per-trial or per-cell rate if the smoke is representative enough to estimate lock runtime. After the lock, record: - branch; - candidate/non-candidate split by mode and horizon; - the signal-delay candidate/non-candidate asymmetry; - whether the next registered move should be hazard-score audit, event-warning-quality rerun, controller redesign, or Phase 16/3D deferral.