Roadmap starts at line ~125 It should look less like “Bayes versus Sundog, one wins” and more like: Bayes turns evidence into belief. Sundog turns response into control. That is the cleanest head-to-head. Bayes’ Theorem asks: P(world∣signal)∝P(signal∣world)P(world) Sundog asks: H(x)≈ dτ dS ​ or, in plain terms: When I act, how does the indirect signal change, and can I use that change to steer? The fair comparison is not “Bayes gets full state, Sundog gets shadows.” That would be a strawman. A fair Bayesian baseline also gets partial observations. The difference is that Bayes tries to reconstruct a posterior over the hidden world, while Sundog tries to couple action directly to the signal field. The Sundog docs already define the shared pattern this way: the system does not observe the whole truth, receives an indirect signal, transforms it into a control-relevant form, and acts from that transformed signal rather than from full world-state symmetry. The simplest headline: Bayes: “Where is the hidden thing, probably?” Sundog: “Which action makes the trace improve?” For a visual demo, I would make it a split-screen duel. Left side: Bayes. A dark board, hidden target, and a posterior heatmap. Every observation updates the heatmap. The UI shows “prior,” “likelihood,” “posterior,” and “expected utility.” It feels analytical, cool, map-like. Right side: Sundog. Same hidden target, but no posterior heatmap. Instead, a live signal field, detector bars, scan arcs, and a trace line showing signal improvement. It feels embodied, experimental, servo-like. The shared task: A hidden source emits an indirect signal. The agent must align to it, reach it, or keep a system stable around it. Both agents receive the same sensor readings. Neither receives privileged target coordinates. The fair Sundog version should use the real project shape: detector intensity plus proprioception, no laser position, no mirror hit point, no target Cartesian position, and a SCAN → SEEK → TRACK loop. That is the current measurable form of the older Sundog idea. The head-to-head should have three rounds. Round 1: Clean model. Bayes should win. The hidden source produces a known Gaussian field. Noise is known. Geometry is known. The Bayesian agent has the correct likelihood model. Bayes updates its posterior efficiently, localizes the hidden target, and chooses the best action. Sundog still has to scan, seek, and track. It can work, but it pays the acquisition-time cost. This matches the current Sundog evidence posture: the photometric controller can match terminal oracle accuracy in the tested mirror task, but it is much slower, with median time-to-threshold around 188 steps versus 11.5 for the target-aware analytic baseline. Verdict stamp: BAYES WINS — MODEL KNOWN Overlay copy: When the likelihood is right, inference is powerful. Bayes does not need to wander. It can believe its way toward the target. Round 2: Unknown response surface. Sundog should look strong. Now keep the target hidden, but corrupt the model: warped optics, shifted detector calibration, unmodeled surface reflection, mild occlusion, or a field whose shape is not the assumed likelihood. The Bayesian agent updates confidently into the wrong posterior because its likelihood is wrong. Sundog does not need the full generative model. It perturbs, reads the actual response, and climbs what the world is actually giving back. Verdict stamp: SUNDOG HOLDS — FIELD READABLE Overlay copy: The map lied. The response still answered. This is the core Sundog pitch without overclaiming. The project’s promo language already frames the idea as indirect environmental response carrying actionable structure, not perfect world reconstruction. Round 3: Decoy or local maximum. Sundog should fail. Add a false signal peak. The visible trace is locally attractive but globally wrong. A flat Sundog controller locks onto the decoy because the local signal improves there. A Bayesian agent with the right structural model can represent two hypotheses, gather disambiguating evidence, and eventually reject the decoy. Verdict stamp: BOUNDARY FOUND — SIGNAL ALIASED Overlay copy: A signal can be useful without being truthful. Sundog fails when the trace improves for the wrong reason. This is the most important round. It prevents the comparison from becoming marketing. It says: Bayes is better when the system needs explicit uncertainty over hidden causes; Sundog is better when the goal is direct control from a real response field and a full posterior is unnecessary, unavailable, or too expensive. The landing-page card could read: Bayes vs. Sundog Status: Comparative Benchmark Bayes: Evidence → posterior → action. Sundog: Probe → response → control. Question: Must the agent infer the hidden world, or can it act from the trace? Verdict: Complementary, not redundant. For the rail, I would avoid a “Sundog beats Bayes” stamp. Use a sequence of stamps instead: MODEL KNOWN → Bayes wins. FIELD READABLE → Sundog holds. SIGNAL ALIASED → Sundog fails. HYBRID NEEDED → real systems use both. The deeper theorem-level distinction: Bayes is an epistemic theorem. It is about how belief should change when evidence arrives. Sundog is a control hypothesis. It is about when indirect environmental structure can be transformed into useful action without reconstructing the full hidden state. So the best head-to-head line is: Bayes is for knowing under uncertainty. Sundog is for acting under partial sight. A strong final card: Title: Posterior vs. Halo Clip: Left side heatmap sharpens; right side signal trace climbs. Then a decoy appears. The heatmap splits; the Sundog trace locks onto the wrong peak. Stamp: COMPLEMENTARY Description: Bayes inferred the hidden cause. Sundog followed the live response. Both were right until the signal stopped meaning what it seemed to mean. Theorem meaning: Sundog does not replace inference. It identifies the cases where action-coupled signal is enough, and the cases where belief over hidden causes is still necessary. This would be one of the strongest public-facing comparisons because it respects Bayes instead of caricaturing it, while making Sundog’s distinct claim sharper. # Sundog vs. Bayes Working hook: > Bayes turns evidence into belief. Sundog turns response into control. This > roadmap tests when the posterior is worth paying for, when the halo is enough, > and where a hybrid is required. This roadmap defines a comparative benchmark between Bayesian inference and Sundog-style response control. It is not a claim that Sundog replaces Bayes. Bayes is the correct comparator precisely because it is strong: it already handles uncertainty, hidden causes, partial observations, and principled belief updates. A weak baseline would make the comparison easier and less useful. The experiment asks a narrower question: > Given the same partial observations and the same action budget, should an > agent maintain an explicit posterior over hidden causes, or can it act directly > from the action-coupled response field? The intended result is not one global winner. The intended result is an envelope: - known model: Bayes should win; - readable but mis-modeled field: Sundog should hold; - aliased or decoy field: flat Sundog should fail; - mixed regimes: hybrid should earn its keep. ## Why Bayes Is the Right Next Comparator The current Sundog research claim is narrow: a controller without Cartesian access to a target can align a mirrored end-effector using sparse photometric feedback and proprioception, reaching terminal accuracy comparable to a target-aware analytic baseline in the tested MuJoCo setting. That is a result against an oracle-style direct-state baseline, not yet against the most serious partial-observation inference baseline. Bayes is the serious partial-observation comparator because it does not require full sight. A Bayesian agent can be denied the true target location, receive the same indirect signal as Sundog, and still maintain a posterior over hidden state. If Sundog only works because its baseline is unfairly blind, the Bayes comparison will expose that. If Sundog is genuinely a different control regime, the comparison should reveal where direct response control is cheaper, more robust, or more honest than posterior reconstruction. The clean distinction: ```text Bayes: observation -> likelihood -> posterior -> action Sundog: probe -> response -> control -> next probe ``` Bayes asks: ```text P(hidden_world | signal) ∝ P(signal | hidden_world) P(hidden_world) ``` Sundog asks: ```text H(x) ≈ dS / dτ ``` or, in operational language: > When I act, how does the indirect signal change, and can that change steer me? This comparison is worth running because it will prevent two bad public reads: 1. **Anti-Bayes caricature.** Bayes is not “full-state classic AI.” Bayes is a principled way to reason under partial observation. 2. **Sundog inflation.** Sundog is not a universal replacement for inference. It is a control posture for cases where the response field is actionable before the hidden cause is fully reconstructed. ## What's Honest vs. What's Reach **Honest:** - A measured envelope showing when a correct Bayesian model beats direct response control on acquisition time, terminal accuracy, calibration, and wrong-lock rate. - A measured envelope showing when a Sundog response controller remains useful under model drift, calibration warp, or unknown response geometry where a fixed Bayesian likelihood misleads. - A serious Sundog failure card: signal aliasing / decoy response fields where the trace improves for the wrong reason and a flat response controller locks onto the wrong basin. - A hybrid result, if earned, showing when posterior maintenance should be used for hypothesis management while Sundog-style response tracking handles local control. - A public visualization that makes the difference legible: posterior heatmap versus live response trace, then a decoy case where the heatmap splits and the response trace lies. **Reach (do not claim):** - "Sundog beats Bayes." - "Bayesian inference requires full state." - "Posterior inference is obsolete." - "Sundog is a faster Bayesian method." - "Sundog solves partial observability in general." - "Sundog replaces uncertainty modeling." - "The failure of one Bayesian implementation refutes Bayes." The value of this roadmap is complementarity, not conquest. A result where Bayes wins cleanly under a known model is not a Sundog failure. It is the calibration point. A result where Sundog fails under aliasing is not a program failure either. It is the boundary the public rail needs: **the signal can be useful without being truthful.** ## Ratified Hook Language Safe hook: > Sundog vs. Bayes asks whether an agent under partial observation must infer > the hidden world, or whether it can sometimes act directly from the trace the > world gives back. Short version: > Bayes is for knowing under uncertainty. Sundog is for acting under partial > sight. Public one-liner: > The posterior draws a map. The halo moves the hand. Avoid: - "Bayes sees the whole world." - "Sundog beats probabilistic reasoning." - "Inference is unnecessary." - "The posterior is just another proxy." - "Bayes is slow, Sundog is fast." The current photometric result shows the opposite cost shape: Sundog can pay large acquisition time to avoid target coordinates. ## Core Experimental Question The shared task family is hidden-source navigation / alignment. A hidden target or source emits an indirect signal field. The agent controls a 2D point, mirror pose, probe pose, or simple embodied body. It receives only sensor observations derived from the field and its own proprioceptive/action history. It does not receive the hidden target coordinates. The agent must align to the source, maximize target intensity, maintain a desired response, or choose when to abstain. Both Bayes and Sundog receive the same raw observations. The difference is how they transform them. - The Bayesian family maintains a belief distribution over hidden causes and chooses actions by expected utility, information gain, Thompson sampling, model-predictive planning, or a simple value-of-information rule. - The Sundog family does not maintain a full posterior. It scans, seeks, tracks, reacquires, and acts from measured changes in the response field. - The hybrid family uses posterior structure only for hypothesis management, disambiguation, and reset decisions, while using response tracking for local control. The comparison is fair only if Bayes is allowed to be genuinely Bayesian under partial observation. The comparison is unfair if Bayes is forced to be a direct state oracle or if Sundog is allowed privileged geometry. ## Applied Workbench Cross-References This roadmap now has two lanes: 1. the standalone Bayes-vs-Sundog benchmark defined below; and 2. Bayesian-floor profiles stamped into existing Sundog workbenches as claim-hygiene comparators. The applied profiles are not substitutes for the standalone benchmark. They are the local audit surfaces that keep each workbench from accidentally claiming too much from weak baselines. | Workbench | Bayesian comparator status | Cross-reference | | --- | --- | --- | | Balance | Executable same-shadow floor scaffold. `bayes_floor_shadow_particle` runs with observation parity, no-state-leak, unknown-mode rejection, Phase 10 slate loading, and regret receipts. Final Phase 15 lock ran 27,200 trials with audits passing, claim gate 56/56 hard-gate cells, 12 reported-only observation-degradation cells, zero negative mean-regret cells, and aggregate regret versus `sundog_shadow` +0.00395. Treat Balance as having earned a same-information Bayesian-floor claim-lock receipt, with reported-only delay/noise/dropout failure-regime lanes preserved as boundary diagnostics. | [`SUNDOG_V_BALANCE.md` - Bayesian Floor Profile](SUNDOG_V_BALANCE.md#bayesian-floor-profile) and Phase 15 status | | Pressure Mines | Active applied profile. The staged same-field baseline is `mines-bayesian-baseline-v1`: a frontier-limited posterior over hidden mine occupancy, strict budget parity between pressure-only and full Sundog-legal lanes, and a Phase 12 path that starts with legal `Phi_t` serialization, no-leak, observation parity, and tiny posterior smoke before any claim lock. Admission plumbing exists (`npm run mines:phase12:admission`). The repaired executable smoke exists (`npm run mines:phase12:posterior-smoke`) and uses shared cell/seed sensor streams plus a pressure-floor action guard. The 64-seed Phase 10 reducer ran 14,720 trials with no-leak pass: `bayes_frontier_pressure` matches `naive_pressure` / `sundog_minimal` exactly across all 46 cells, while `sundog_lean` remains stronger in the promoted candidate cell. Phase 13 now publishes `public/data/mines-phase13-bayes-floor.json`, a Mines evidence panel, representative posterior-cell hazard/pressure/confidence slices, and replay selectors for confirmed, failure, and Bayes-divergence cells. Treat this as a same-field pressure-floor receipt, not a posterior-dominance claim lock. | [`sundog_v_minesweeper.md` - Bayesian Baseline Profile](sundog_v_minesweeper.md#bayesian-baseline-profile) and Phase 13 status | | Three-Body / Coarse Graining | Current pattern source for the receipt shape: admission spec, same-information guard, belief diagnostics, regret reducer, and capped-probe discipline. Still not a closed floor. | [`BAYESIAN_FLOOR_PROFILE_TEMPLATE.md`](BAYESIAN_FLOOR_PROFILE_TEMPLATE.md) and [`proof/PHASE4_THREEBODY.md`](proof/PHASE4_THREEBODY.md) | The interpretation rule is shared across all applied profiles: - if Bayes is invalid or weak, repair it before interpreting the Sundog result; - if Bayes dominates, narrow the Sundog claim to a weaker controller or heuristic claim; - if Sundog tracks a repaired same-observation floor, the workbench earns stronger evidence that response-control recovered most of the actionable information in that envelope; - if both fail in the same cells, the boundary is more likely to be an observability boundary than merely a controller boundary. Applied queue after the Balance lock: 1. **Pressure Mines Phase 12 admission path.** Declare `bayes_frontier_pressure` / `bayes_frontier_full`, serialize the legal `Phi_t`, and prove no-leak / observation parity before implementing the posterior policy. Landed with `npm run mines:phase12:admission`. 2. **Pressure Mines tiny posterior smoke.** Run only the Phase 10 confirmed pocket plus paired failure cell with capped particles and seeds. Landed with `npm run mines:phase12:posterior-smoke`; no-leak and easy-cell sanity now pass after the same-field sensor repair. 3. **Pressure Mines pocket probe.** Broaden to the `bayes_frontier_full` pocket probe and reduce Bayes-vs-`sundog_minimal` / `sundog_lean` regret into the Phase 12 summary shape. Landed with `npm run mines:phase12:pocket-probe`. 4. **Pressure Mines Phase 10-slate reducer.** Run the repaired pressure lane over the locked Phase 10 envelope cells, then surface Bayes regret fields only if the same-field floor holds. Loader smoke landed with `npm run mines:phase12:phase10-slate:smoke`; the 64-seed all-cell reducer landed as `npm run mines:phase12:phase10-slate`. Public data surfacing now lands through `npm run mines:public-data`, `mines.html`, and the `alignment.html` side-by-side, with explicit language that this is pressure-floor parity, not posterior dominance over `sundog_lean`. Phase 13.1 adds representative posterior-cell slices and replay selectors for the confirmed pocket, paired failure, and Bayes-divergence cell. 5. **Template backfill.** Any Mines-specific deviations from the generic [`BAYESIAN_FLOOR_PROFILE_TEMPLATE.md`](BAYESIAN_FLOOR_PROFILE_TEMPLATE.md) have now been backfilled: admission lanes, same-observation seed policy, claim-gate fields, reported-only lanes, posterior/evidence slices, and public interpretation copy. 6. **Standalone Phase 1 exact reference.** Landed as `scripts/bayes-phase1-reference.mjs` with `npm run bayes:phase1:smoke` and `npm run bayes:phase1`. The 128-seed lock passes the known-model gate before any `bayes.html` work begins. 7. **Standalone Phase 2 mismatch slate.** Landed in the same runner with `npm run bayes:phase2:smoke` and `npm run bayes:phase2`. The lock finds a narrow anisotropic separation for `sundog_memory` over fixed clean `bayes_misspecified`; other variants remain boundary or Bayes-favorable. 8. **Standalone Phase 3 aliasing lock.** Pre-registered dual gate landed and `npm run bayes:phase3` reports `dual_gate_pass` in `decoy` and `alias`. `symmetric` and `low_probe` remain diagnostic: both show Bayes-adaptive recovery, but not the pre-registered HC-Sundog wrong-lock failure arm. 9. **Standalone Phase 5 operating-envelope closeout.** Phase 5 closed with an anchor-validated lock and a degenerate single-class envelope: `bayes_adaptive` dominates throughout the decoy-strength x budget grid. No boundary was mapped; the likely boundary axis is model-mismatch severity, deferred only to an optional separately pre-registered Phase 5b. 10. **Standalone Phase 5b mismatch-severity sweep.** Anchor-validated lock (`reproduced_phase5_lock_cell`, seed-matched, exact) maps a genuine multi-class envelope monotone in severity: `bayes_dominant` (s0-s1) -> `mixed` (s1-s2) -> `response_dominant` (s3). The Bayes-versus-response boundary is misspecification severity; the s3 response win is the Phase 2 separation re-derived (s3 is equivalent to `bayes_misspecified`), not a new claim. 11. **Standalone Phase 6 photometric port.** Anchor-complete lock (five gating components, nominal aggregate + Mann-Whitney + stress all `diff == 0.0`): 8/9 `bayes_particle_dominant`, 1 degenerate `mixed`. Nominal-only particle Bayes is far faster than SCAN/SEEK/TRACK at comparable terminal accuracy across the tested optical/detector misspecification envelope. The Phase 1-5b response-control edge does not replicate on the core task under this optical-parameter axis; this does not refute Phase 5b, because a wrong-likelihood-structure axis would be a separate Phase 6b. 12. **Standalone Phase 6b likelihood-structure ablation.** Anchor-complete lock (structureS0Nest, nominal aggregate, Mann-Whitney all `diff == 0.0`; stress `not_applicable`). 4-rung map: `s0/s1/s2` `bayes_particle_dominant` (ttt 11.0, identical), `s3` `photometric_dominant` (ttt 500, 30/30 failed). Partial wrong structure does not degrade core-task particle Bayes; only the degenerate model-less `s3` flips. The Phase 1-5b response-control edge does not transfer to the core task under non-degenerate wrong structure - the synthetic edge is substrate-specific. Experimental arc closed; no Phase 6c. ## Controller-Family Architecture Run all controller families on the same seeded slates, same observation stream, same action budget, and same termination rules. ### 1. Oracle Privileged ceiling. Knows the hidden source, true field parameters, and current geometry. It is not a deployed baseline. It defines the best plausible terminal and acquisition behavior when the hidden state is granted. Use it to normalize results and detect impossible task settings. Do not use it as the primary Bayes comparator. ### 2. Bayes-Correct Bayesian agent with the correct generative model family and correct noise model. It does not know the hidden target value, but it knows the likelihood class. It maintains a posterior over target location, field parameters, or discrete hypotheses. Expected behavior: should win Round 1. If it does not, the Bayesian implementation is probably too weak. Candidate implementations: - grid posterior over hidden target location for small 2D fields; - particle filter for continuous target / source parameters; - Kalman or extended Kalman filter only where the model is legitimately close to linear-Gaussian; - Gaussian-process surrogate with explicit uncertainty for unknown smooth response surfaces; - active Bayesian optimization for expensive probe settings. ### 3. Bayes-Misspecified Bayesian agent with a plausible but wrong likelihood. Examples: assumes a Gaussian field when the actual response is warped, assumes stationary detector noise when noise is position-dependent, assumes one source when two interacting sources exist, or assumes the decoy field is weak when it is strong. Expected behavior: should become confidently wrong in some cells. This is the main place Sundog should look strong if the response field remains locally readable. ### 4. Bayes-Adaptive Bayesian agent that can update some model parameters online or maintain a small mixture over model families. This is the fairer version of Bayes in the mis-specified setting. It should recover from some mismatch, but pay an action, compute, or sample cost. Expected behavior: may beat both Bayes-Misspecified and Sundog in some cells. That is acceptable. The comparison should report the cost and failure boundary, not force Sundog to win. ### 5. HC-Sundog Hand-coded response controller. Uses SCAN / SEEK / TRACK / REACQUIRE over the observed signal. Does not represent a posterior over hidden target location. It may keep memory of best observed response, local slope estimates, phase state, and confidence gates. Expected behavior: strong when the action-response field is smooth enough, readable enough, and not severely aliased. Weak under decoys, disjoint basins, long-horizon preparatory actions, and cases where the best local response is not evidence of the true hidden target. ### 6. Sundog-Memory Response controller with richer empirical memory but still no posterior over hidden causes. Stores local response samples, basin labels, reacquisition points, and recent probe outcomes. This separates "Sundog fails because it has no memory" from "Sundog fails because it lacks causal belief over hidden hypotheses." Expected behavior: should improve over HC-Sundog in noisy or drifting fields, but still fail under structurally aliased signals. ### 7. Hybrid Bayes-Sundog Maintains a compact posterior or hypothesis set only where ambiguity matters, then uses response tracking for local control. The design target is not "Bayes plus everything." The design target is minimal inference added at the failure boundary. Expected behavior: should dominate flat Sundog in aliasing cells and avoid the full overhead of Bayes-Correct in simple readable fields. If it does not, the hybridization is unjustified. ### 8. Naive Local / Random Lower-bound baselines. - Random or passive. - Greedy local ascent without scan. - Nearest-highest-sensor heuristic. - Posterior heatmap without active probing. These are not the serious comparison. They detect implementation leakage and provide public legibility. ## Decision Lock for Phase 0 ### Environment family Use **shadow-field navigation** as the first environment because it matches the Mesa infrastructure, is cheap to simulate, supports exact posteriors in small versions, and can generate clean probe slates. Canonical Phase 0 task: - Agent moves in bounded 2D continuous space. - Hidden target `x*` emits scalar response field `S(x)`. - Agent observes local field samples around its body, plus its own position and action history. It does not observe `x*`. - Action is bounded 2D velocity or displacement. - Objective is to reach and dwell near the true target / source response. - The slate includes known-model, warped-model, decoy, aliasing, drift, noise, and sensor-delay variants. Start with a grid discretization for exact Bayesian reference. Then port to continuous particles or GP only after the discrete version is debugged. Do not start in the MuJoCo mirror task. The photometric task is the scientific spine, but it is too expensive and geometrically specific for first-pass posterior diagnostics. After the shadow-field benchmark is stable, port the best comparison into the photometric mirror setting as Phase 6.5. ### Claim boundary The roadmap tests a comparative control question, not a theorem of Bayesian inference. Possible final claim shape: > In the tested hidden-source environment family, Bayesian inference dominates > when the likelihood model is correct and ambiguity must be represented; > Sundog-style response control remains competitive or more robust when the > response field is locally readable but the model is mis-specified; flat > Sundog fails under aliased response fields, motivating a hybrid controller. If the result is one-sided in Bayes's favor: > In the tested environment family, posterior-maintaining controllers dominate > response-only controllers across known, mis-specified, and aliased fields > under matched action budgets. Sundog should be reframed as a lightweight > control heuristic that works only when posterior structure is unnecessary. If the result is one-sided in Sundog's favor: > In the tested environment family, response-only control outperforms the > Bayesian implementations under the chosen action/compute budget. This does > not refute Bayes; it shows that the tested posterior machinery was not worth > its cost in these readable fields. Stronger Bayesian baselines remain a > required future pass. ### Evidence tier Initial status: **candidate comparative benchmark / roadmap**. A completed Phase 3 result becomes an **Operating-Envelope Study** only if it has: - seeded matched trials; - at least one serious Bayesian baseline; - at least one Sundog failure boundary; - action-budget and compute-budget reporting; - saved run artifacts; - public summary that avoids "Sundog beats Bayes" language. ## Experimental Rounds ### Round 1 — Known Model Question: > What happens when Bayes has the right likelihood? Setup: - Hidden source emits known Gaussian or known mixture field. - Noise distribution is known. - Sensor geometry is known. - Agent does not know target coordinates. - Bayes-Correct maintains posterior over `x*`. - HC-Sundog scans, seeks, and tracks response. Expected result: Bayes-Correct should win on acquisition time, probe efficiency, and possibly terminal accuracy. Sundog may still succeed, but it should pay a scan cost. Verdict language: > MODEL KNOWN — BAYES WINS This round protects the comparison from becoming anti-inference marketing. ### Round 2 — Warped Field / Wrong Likelihood Question: > What happens when the map is wrong but the response is still readable? Setup: - Actual field is warped, anisotropic, shifted, clipped, delayed, or mildly occluded. - Bayes-Misspecified assumes the clean field. - Bayes-Adaptive can fit some parameters but not arbitrary geometry. - HC-Sundog sees only the live response. Expected result: Bayes-Misspecified should sometimes become confidently wrong. HC-Sundog should track the real response when local gradients remain actionable. Bayes-Adaptive may recover with a sample cost. Verdict language: > FIELD READABLE — SUNDOG HOLDS This is the strongest pro-Sundog round, but only if the Bayesian baseline is not artificially weakened. ### Round 3 — Aliased Field / Decoy Basin Question: > What happens when the signal improves for the wrong reason? Setup: - Add a false response peak or aliased basin. - Local response improves near the decoy. - The true target can be disambiguated only by a structured probe, prior, time sequence, or multi-sensor pattern. - Bayes-Correct maintains two hypotheses. - HC-Sundog follows local improvement. Expected result: Flat Sundog should fail. Sundog-Memory may reduce repeated wrong-locks but should not solve true aliasing without adding belief over hidden causes. Bayes-Correct or Hybrid should win if the disambiguating evidence is present in the observation stream. Verdict language: > SIGNAL ALIASED — BOUNDARY FOUND This is the key failure card for the landing-page rail. ### Round 4 — Drift and Reacquisition Question: > What happens when the response field changes after the agent has committed? Setup: - Target moves slowly, detector calibration drifts, or noise statistics change. - Some changes preserve local response; others invalidate prior structure. - Agents must detect failure and reacquire. Expected result: Sundog may reacquire quickly when the response itself changes visibly. Bayes may handle drift better if the motion model is correct. Hybrid should win when drift alternates between readable and ambiguous regimes. Verdict language: > REACQUIRE — MODEL OR RESPONSE? ### Round 5 — Compute/Action Budget Sweep Question: > Is the posterior worth its cost? Setup: Sweep: - action budget; - compute budget; - sensor count; - noise; - field smoothness; - decoy strength; - prior quality; - model mismatch severity. Expected result: Produce a Pareto map, not a single winner. Regions should be labeled: - Bayes-dominant; - Sundog-dominant; - Hybrid-dominant; - all-fail; - ambiguous / needs stronger baseline. Verdict language: > OPERATING ENVELOPE ## Metrics Primary metrics: - terminal response / terminal target intensity; - time-to-threshold; - cumulative response / regret; - wrong-lock rate; - dwell success after threshold; - reacquisition time after drift; - action count; - compute time; - memory footprint. Bayes-specific metrics: - posterior entropy over time; - posterior calibration; - negative log likelihood of held-out observations; - posterior mass on true target / true hypothesis; - time to disambiguate decoy; - information-gain per action. Sundog-specific metrics: - best-observed response memory; - scan coverage; - local slope estimate quality; - reacquisition trigger correctness; - false-track duration; - response-gradient collapse rate. Shared failure labels: - **Known-model Bayes win.** Posterior correctly concentrates; Sundog pays scan cost. - **Wrong-model Bayes collapse.** Posterior concentrates on wrong basin because likelihood is false. - **Readable-field Sundog win.** Local response remains coupled to useful action despite model mismatch. - **Signal alias Sundog collapse.** Local response improves toward false basin. - **Hybrid necessity.** Direct response gets close, posterior disambiguation prevents wrong lock. - **All-fail.** Signal contains insufficient information or action budget is too small. ## Roadmap ### Phase 0 — Scope and Literature Pass Goal: pin the comparison so it is fair to Bayes and useful to Sundog. Deliverables: - One-page claim boundary for the Bayes comparison. - Definitions of observation, hidden state, action, field, and success. - Baseline list: Oracle, Bayes-Correct, Bayes-Misspecified, Bayes-Adaptive, HC-Sundog, Sundog-Memory, Hybrid, Naive Local, Random. - Applied-profile index pointing to Balance, Pressure Mines, and Three-Body so the standalone benchmark inherits the strongest existing receipt conventions instead of inventing a parallel shape. - Literature spine covering Bayesian filtering, Bayesian decision theory, active inference / active sensing, Bayesian optimization, POMDPs, extremum seeking, visual servoing, and model mismatch. - Decision about first environment: discrete 2D grid first, continuous field second. - Budget policy: action budget, compute budget, and wall-clock reporting. Exit criterion: no one can plausibly object that Bayes was straw-manned before implementation starts. ### Phase 1 — Reference Task and Exact Bayesian Baseline Goal: build the smallest task where Bayes is exact and should win. Deliverables: - Seeded 2D hidden-source grid environment. - Exact grid posterior over hidden target location. - Known Gaussian/noisy field observation model. - Bayes-Correct policy using expected utility and optionally information gain. - HC-Sundog SCAN / SEEK / TRACK controller over the same observations. - Oracle and random baselines. - JSONL trial logs and replay manifest. Status, 2026-05-17: executable lock landed. Commands: ```bash npm run bayes:phase1:smoke npm run bayes:phase1 ``` Receipt paths: - `results/bayes/phase1-reference-smoke/manifest.json` - `results/bayes/phase1-reference-lock/manifest.json` - `results/bayes/phase1-reference-lock/trials.jsonl` - `results/bayes/phase1-reference-lock/steps.jsonl` - `results/bayes/phase1-reference-lock/replay-manifest.json` Phase 1 lock summary (`npm run bayes:phase1`, 128 seeds, 512 trials, 29.529s after the Phase 2 schema extension): | Mode | Successes | Success rate | Mean score | Mean hit turn | | --- | ---: | ---: | ---: | ---: | | `oracle` | 128/128 | 1.000000 | 1.623372 | 9.039063 | | `bayes_correct` | 126/128 | 0.984375 | 1.486077 | 11.738095 | | `hc_sundog` | 27/128 | 0.210938 | 0.062268 | 11.481481 | | `random` | 10/128 | 0.078125 | -0.257697 | 12.900000 | Exit gate: **pass**. `bayes_correct` beats `hc_sundog` under the known model: mean score delta +1.423809, success-rate gate true, score gate true. The hit turn comparison is only gate-active when success rates are tied, to avoid rewarding a controller for succeeding only on easy seeds. Representative replay roles are now selected in `replay-manifest.json`: `bayes_advantage`, `closest_pair`, and `bayes_boundary`. The boundary role is useful claim hygiene, not a Phase 1 failure: Bayes still clears the aggregate known-model gate, but the replay keeps the two missed seeds visible before Phase 2. Exit criterion: Bayes-Correct beats or matches HC-Sundog under the known model. If it does not, fix Bayes before proceeding. ### Phase 2 — Model Mismatch Slate Goal: test whether Sundog has a real advantage when likelihood structure is wrong but response remains actionable. Deliverables: - Warped Gaussian field. - Anisotropic field. - Detector calibration shift. - Mild occlusion / clipped field. - Delayed sensor response. - Bayes-Misspecified baseline locked to the clean likelihood. - Bayes-Adaptive baseline with limited parameter learning. - Matched trials against HC-Sundog and Sundog-Memory. Status, 2026-05-17: executable mismatch lock landed. Commands: ```bash npm run bayes:phase2:smoke npm run bayes:phase2 ``` Receipt paths: - `results/bayes/phase2-mismatch-smoke/manifest.json` - `results/bayes/phase2-mismatch-lock/manifest.json` - `results/bayes/phase2-mismatch-lock/summary.csv` - `results/bayes/phase2-mismatch-lock/regret.csv` - `results/bayes/phase2-mismatch-lock/replay-manifest.json` Implementation details: - `bayes_misspecified` keeps an exact grid posterior, but its likelihood is locked to the clean radial Gaussian regardless of the true field variant. - `bayes_adaptive` is a finite-mixture Bayesian baseline over clean, warped, anisotropic, calibration-shift, clipped, and delayed model families. - `sundog_memory` is the response-control repair lane: it scans until the live signal rises above its own starting baseline, then probes from remembered high-response cells. Phase 2 lock summary (`npm run bayes:phase2`, 96 seeds, 2,880 trials, 239.345s): | Scenario | `bayes_misspecified` score / success | best response lane | best response score / success | Result | | --- | ---: | --- | ---: | --- | | `anisotropic` | 0.767947 / 0.500000 | `sundog_memory` | 1.359587 / 0.875000 | separated | | `calibration_shift` | 0.520778 / 0.458333 | `sundog_memory` | 0.549783 / 0.500000 | near miss under +0.1 margin | | `warped` | 0.978200 / 0.614583 | `sundog_memory` | 0.905387 / 0.677083 | response wins survival, loses score | | `clipped` | 1.646094 / 1.000000 | `sundog_memory` | 0.785773 / 0.604167 | Bayes-favorable | | `delayed` | 1.466406 / 1.000000 | `hc_sundog` | 0.108966 / 0.208333 | Bayes-favorable | Exit gate: **separation_found**. The current Phase 2 receipt supports a narrow claim: under an anisotropic response surface, fixed clean Bayes is brittle and the memory response-control lane recovers better. It does **not** support a blanket mismatch claim; clipped and delayed variants still favor the fixed posterior baseline, and calibration shift is a close boundary rather than a locked win. Exit criterion: at least one mismatch regime cleanly separates fixed posterior failure from response-control recovery, or the roadmap records that the tested mismatch was insufficient. ### Phase 3 — Aliasing and Decoy Failure Slate Goal: force the failure the public rail needs. Deliverables: - False peak / decoy field variant. - Two-source alias variant. - Symmetric ambiguity variant. - Low-probe-budget ambiguity variant. - Disambiguating probe protocol available to Bayes and Hybrid. - Wrong-lock classifier. - Representative replay clips. Pre-registered smoke-harness-first spec, locked before the Phase 3 smoke run: | Scenario | Field shape | Intended Sundog failure | | --- | --- | --- | | `decoy` | True radial Gaussian plus lower-amplitude false peak at a seeded decoy cell. | Flat climb locks onto local decoy. | | `alias` | True radial Gaussian plus near-equal seeded second source. | Response improves toward either basin. | | `symmetric` | Target-reflected source makes the field nearly symmetric about an axis. | Identical local response on two sides. | | `low_probe` | `alias` field with `maxTurns` overridden to 20. | No budget to disambiguate. | Modes: ```text oracle,bayes_misspecified,bayes_adaptive,hc_sundog,sundog_memory,random ``` Admission notes: - `bayes_misspecified` remains locked to the clean single-radial likelihood. - `bayes_adaptive` extends the finite model mixture with `decoy`, `alias`, and `symmetric` families; this is the recovery lane. No new Hybrid mode is admitted in Phase 3. - `low_probe` is a scenario-level budget override; the npm commands stay fixed at `--max-turns 40`, and the runner applies 20 only inside `low_probe`. Wrong-lock classifier: ```text wrong_lock = !success AND dist(finalPos, nearestNonTargetBasin) <= 1.5 AND dwellTurns(within 1.5 of that basin) >= 5 ``` Dual gate: Phase 3 is `dual_gate_pass` only if both arms are true in the same eligible scenario; otherwise it records `aliasing_insufficient` and exits non-fatally (`pass: true`), matching the Phase 2 pattern. | Arm | Fixed threshold | | --- | --- | | Failure arm | `hc_sundog` wrong-lock rate >= 0.50 in at least one `decoy`, `alias`, or `low_probe` regime. | | Recovery arm | In that same scenario, `bayes_adaptive` versus the best Sundog lane has `successDelta >= 0` and `scoreDelta >= 0.10`. | Commands: ```bash npm run bayes:phase3:smoke npm run bayes:phase3 ``` The lock command was pre-registered before the smoke run and then run unchanged. Do not edit the thresholds, classifier radius/dwell, or recovery margin after results land. Smoke receipt, 2026-05-17: ```text npm run bayes:phase3:smoke 384 trials in 104.937s Exit gate: dual_gate_pass (3/4 dual-gate scenarios) ``` Receipt paths: - `results/bayes/phase3-aliasing-smoke/manifest.json` - `results/bayes/phase3-aliasing-smoke/summary.csv` - `results/bayes/phase3-aliasing-smoke/regret.csv` - `results/bayes/phase3-aliasing-smoke/replay-manifest.json` - `results/bayes/phase3-aliasing-lock/manifest.json` - `results/bayes/phase3-aliasing-lock/summary.csv` - `results/bayes/phase3-aliasing-lock/regret.csv` - `results/bayes/phase3-aliasing-lock/replay-manifest.json` Smoke gate rows: | Scenario | HC wrong-lock rate | Failure arm | Best Sundog lane | Bayes-adaptive score / success | Score delta | Success delta | Dual gate | | --- | ---: | --- | --- | ---: | ---: | ---: | --- | | `decoy` | 0.5000 | true | `sundog_memory` | 1.676563 / 1.0000 | +1.648635 | +0.8125 | true | | `alias` | 0.7500 | true | `sundog_memory` | 1.427670 / 0.8750 | +1.441081 | +0.6875 | true | | `symmetric` | 0.3125 | false | `sundog_memory` | 0.400235 / 0.3750 | +0.107920 | 0.0000 | false | | `low_probe` | 0.5000 | true | `sundog_memory` | 1.085602 / 0.8125 | +1.170522 | +0.6250 | true | Smoke interpretation: the smoke harness was strong enough to proceed to the pre-registered lock unchanged. The lock receipt below is now the claim-bearing Phase 3 result. Lock receipt, 2026-05-17: ```text npm run bayes:phase3 2,304 trials in 512.632s Exit gate: dual_gate_pass (2/4 dual-gate scenarios) ``` Lock gate rows: | Scenario | HC wrong-lock rate | Failure arm | Best Sundog lane | Bayes-adaptive score / success | Score delta | Success delta | Dual gate | | --- | ---: | --- | --- | ---: | ---: | ---: | --- | | `decoy` | 0.572917 | true | `sundog_memory` | 1.722396 / 1.000000 | +1.329126 | +0.604167 | true | | `alias` | 0.510417 | true | `sundog_memory` | 1.483721 / 0.906250 | +1.060612 | +0.510417 | true | | `symmetric` | 0.375000 | false | `sundog_memory` | 0.818609 / 0.562500 | +0.210788 | +0.052083 | false | | `low_probe` | 0.395833 | false | `sundog_memory` | 1.067770 / 0.781250 | +0.819386 | +0.427083 | false | Interpretation: Phase 3 now has a lock-backed aliasing boundary. `decoy` and `alias` satisfy the dual gate: HC-Sundog wrong-locks at or above the pre-registered 0.50 threshold, and Bayes-adaptive recovers with non-negative success delta plus score delta above +0.10. `symmetric` and `low_probe` show Bayes-adaptive recovery but do not satisfy the failure arm, so they stay diagnostic rather than claim support. Expected outcome: HC-Sundog should fail in at least one serious decoy regime. If it does not, increase alias strength carefully, but do not create an impossible toy just to force failure. The failure must teach something about response control. Exit criterion: either a real Sundog failure boundary is mapped, or the result shows that the selected aliasing slate was not strong enough and Phase 3 must be redesigned. ### Phase 4 - Hybrid Controller Goal: test the minimum posterior needed to repair the lock-backed Phase 3 `decoy` and `alias` response-control failure regimes. Frozen claim surface: - Claim-bearing scenarios: `decoy` and `alias`. - Diagnostic-only scenarios: `symmetric` and `low_probe`. - `oracle` remains a privileged yardstick and is never claim-gated. - `bestSundog := hc_sundog`, because Phase 4's locked mode set has only one no-posterior response lane. Frozen mode set: ```text oracle,bayes_adaptive,hc_sundog,hybrid,hybrid_no_posterior,hybrid_posterior_only,hybrid_posterior_reset_only,hybrid_posterior_decoy_disambig ``` The repair predicate is fixed once and reused by the repair and load-bearing arms: ```text repair(M) = (M.meanScore - hc_sundog.meanScore >= 0.10) AND (M.successRate - hc_sundog.successRate >= 0) ``` `0.10` is the existing `phase2SeparationMargin`, reused unchanged from Phases 2 and 3. Three-arm gate, evaluated per scenario: | Arm | Predicate | | --- | --- | | Repair arm | `repair(hybrid)` | | Frugality arm | `hybrid.meanLikelihoodEvals <= 0.5 * bayes_adaptive.meanLikelihoodEvals` | | Load-bearing arm | `NOT repair(hybrid_no_posterior) AND repair(hybrid_posterior_decoy_disambig)` | New frozen constants: | Constant | Value | Use | | --- | ---: | --- | | `FRUGALITY_FRACTION` | 0.5 | Hybrid must use no more than half the full-posterior ceiling's likelihood-eval count. | | `AMBIGUITY_MASS` | 0.5 | Posterior ambiguity gate for off-response-basin posterior mass. | `likelihoodEvals` is the cumulative per-candidate Gaussian log-likelihood eval count used by posterior updates. It is reported per trial and summarized as `meanLikelihoodEvals`. Frozen ambiguity trigger for plain `hybrid`: - the response controller is in `track` for at least `WRONG_LOCK_DWELL_TURNS` turns without success; - every consecutive signal delta in that track-dwell window is `< noiseStd`; - after the posterior is consulted, posterior mass outside the current response basin is at least `AMBIGUITY_MASS`. The response basin is centered on HC-Sundog's current `bestPos`; "off basin" means candidate distance greater than `WRONG_LOCK_RADIUS`. The trigger reuses the Phase 3 wrong-lock dwell/radius constants rather than introducing new thresholds. `hybrid` takes the HC-Sundog response action by default and only allows the posterior to override when all three trigger clauses are true. Hard-drop override: ```text if repair(hybrid_no_posterior) is true in either decoy or alias: status = hybrid_unnecessary ``` Otherwise, Phase 4 reports: - `hybrid_niche_confirmed` only if all three arms are true in both `decoy` and `alias`; - `hybrid_no_niche` if the hard-drop does not fire but either claim scenario misses at least one arm. The Phase 4 gate is non-fatal (`pass: true`), matching the Phase 2/3 pattern. Commands: ```bash npm run bayes:phase4:smoke npm run bayes:phase4 ``` Command-text audit note: the frozen draft named `--scenarios phase3`; the npm scripts use `--scenarios phase4`, which is a pure runner alias to the same four Phase 3 scenarios (`decoy`, `alias`, `symmetric`, `low_probe`). This is command text drift only, not a scenario-set change. The smoke ships first. The lock command is staged unchanged for the operator only after the smoke verifies the harness shape. Superseded exploratory smoke, 2026-05-17: ```text npm run bayes:phase4:smoke 512 trials in 462.877s Exit gate: hybrid_niche_confirmed (2/2 claim scenarios all-arms) ``` Historical gate rows, retained here because the standard smoke path was later overwritten by the repaired frozen-trigger smoke: | Scenario | Hybrid score / success | HC-Sundog score / success | Hybrid evals | Bayes-adaptive evals | Arms | | --- | ---: | ---: | ---: | ---: | --- | | `decoy` | 1.564062 / 1.0000 | -0.061427 / 0.1875 | 1260.25 | 3674.25 | repair true; frugality true; load-bearing true | | `alias` | 0.810088 / 0.6250 | -0.355776 / 0.0625 | 1881.50 | 4615.00 | repair true; frugality true; load-bearing true | Audit interpretation: this smoke is exploratory/diagnostic only and must not be promoted. The gate scaffolding was faithful, but plain `hybrid` used a periodic posterior refresh derived from `FRUGALITY_FRACTION` and `AMBIGUITY_MASS`, so the frugality arm was not an independent measurement of the frozen ambiguity-triggered controller. The Phase 4 claim remains unearned until the frozen trigger is rerun. Repaired frozen-trigger smoke receipt, 2026-05-17: ```text npm run bayes:phase4:smoke 512 trials in 191.794s Exit gate: hybrid_no_niche (0/2 claim scenarios all-arms) ``` Receipt paths: - `results/bayes/phase4-hybrid-smoke/manifest.json` - `results/bayes/phase4-hybrid-smoke/summary.csv` - `results/bayes/phase4-hybrid-smoke/regret.csv` - `results/bayes/phase4-hybrid-smoke/replay-manifest.json` Repaired smoke gate rows: | Scenario | Hybrid score / success | HC-Sundog score / success | Hybrid evals | Bayes-adaptive evals | Arms | | --- | ---: | ---: | ---: | ---: | --- | | `decoy` | -0.061427 / 0.1875 | -0.061427 / 0.1875 | 0.00 | 3674.25 | repair false; frugality true; load-bearing true | | `alias` | -0.355776 / 0.0625 | -0.355776 / 0.0625 | 35.50 | 4615.00 | repair false; frugality true; load-bearing true | Repaired smoke interpretation: the frozen trigger is now faithful and the result is a clean smoke negative. `hybrid` does not repair either claim scenario, so Phase 4 has not earned a niche. The load-bearing ablation still repairs both claim scenarios, which means the posterior can be sufficient; the plain ambiguity trigger is too restrictive or fires too late on this smoke. Lock staging, held after repaired smoke: ```bash npm run bayes:phase4 ``` Estimate from the repaired smoke rate: 3,072 trials at 2.67 trials/s is about 19 minutes on the current CPU lane, so the lock remains operator-run under the repo's long-run rule. Because the repaired smoke is `hybrid_no_niche`, do not run or promote the lock as claim support without an explicit redesign or decision to test the negative at lock scale. The read-back artifact would be `results/bayes/phase4-hybrid-lock/manifest.json`. Outcome branches: - `hybrid_niche_confirmed`: promote Phase 4 from smoke to lock-backed niche claim. - `hybrid_unnecessary`: drop Hybrid; the no-posterior response floor repaired a claim scenario. - `hybrid_no_niche`: keep the smoke receipt but redesign or narrow Phase 4 before claiming a hybrid niche. ### Phase 4b - No-Progress Hybrid Trigger Goal: isolate whether the Phase 4 negative came from the track/plateau trigger rather than from the frozen three-arm gate or the hybrid mode set. Phase 4b is a single-variable preregistered change against Phase 4. Only the plain `hybrid` trigger changes; the gate arms, constants, ablations, claim scenarios, hard-drop override, statuses, and non-fatal `pass: true` behavior remain identical to Phase 4. Frozen Phase 4b trigger for plain `hybrid`: - maintain `noProgressTurns` on the plain-`hybrid` response state; - each turn, if `state.response.bestSignal` strictly increased, reset `noProgressTurns` to 0; otherwise increment it; - fire when `noProgressTurns >= WRONG_LOCK_DWELL_TURNS`; - on fire, do one `updateAdaptivePosterior` refresh, then compute `posteriorOffBasinMass(view, state.response.bestPos)`; - use the posterior action only when that off-response-basin mass is at least `AMBIGUITY_MASS`; - reset `noProgressTurns` to 0 after every fire, whether or not the posterior action is taken. There are zero new constants. Phase 4b reuses `WRONG_LOCK_DWELL_TURNS = 5`, `AMBIGUITY_MASS = 0.5`, and `WRONG_LOCK_RADIUS = 1.5` inside `posteriorOffBasinMass`. The controller contains no `FRUGALITY_FRACTION` coupling. Unchanged from Phase 4: - modes: `oracle,bayes_adaptive,hc_sundog,hybrid,hybrid_no_posterior,hybrid_posterior_only,hybrid_posterior_reset_only,hybrid_posterior_decoy_disambig`; - `repair(M) = scoreDelta >= 0.10 AND successDelta >= 0` versus `hc_sundog`; - frugality arm: `hybrid.meanLikelihoodEvals <= 0.5 * bayes_adaptive.meanLikelihoodEvals`; - load-bearing arm: `NOT repair(hybrid_no_posterior) AND repair(hybrid_posterior_decoy_disambig)`; - hard-drop: `repair(hybrid_no_posterior)` in either claim scenario yields `hybrid_unnecessary`; - claim scenarios: `decoy` and `alias`; `symmetric` and `low_probe` remain diagnostic-only. Commands: ```bash npm run bayes:phase4b:smoke npm run bayes:phase4b ``` Smoke ships first. The lock command is frozen but deferred and operator-run. Smoke receipt, 2026-05-17: ```text npm run bayes:phase4b:smoke 512 trials in 203.108s Exit gate: hybrid_niche_confirmed (2/2 claim scenarios all-arms) ``` Receipt paths: - `results/bayes/phase4b-hybrid-smoke/manifest.json` - `results/bayes/phase4b-hybrid-smoke/summary.csv` - `results/bayes/phase4b-hybrid-smoke/regret.csv` - `results/bayes/phase4b-hybrid-smoke/replay-manifest.json` Smoke gate rows: | Scenario | Hybrid score / success | HC-Sundog score / success | Hybrid evals | Bayes-adaptive evals | Arms | | --- | ---: | ---: | ---: | ---: | --- | | `decoy` | 0.089803 / 0.2500 | -0.061427 / 0.1875 | 1082.75 | 3674.25 | repair true; frugality true; load-bearing true | | `alias` | -0.196441 / 0.1250 | -0.355776 / 0.0625 | 1224.75 | 4615.00 | repair true; frugality true; load-bearing true | Trigger diagnostics from `steps.jsonl`: | Scenario | Hybrid steps | Stuck-window fires | Posterior actions | Posterior refreshes | | --- | ---: | ---: | ---: | ---: | | `decoy` | 551 | 61 | 59 | 61 | | `alias` | 615 | 69 | 61 | 69 | Smoke interpretation: Phase 4b earns the smoke niche with a real non-zero stuck-window path rather than a periodic clock. The claim is still smoke-only: the frozen lock command is deferred. Smoke-derived lock estimate is 3,072 trials at 2.52 trials/s, about 20 minutes on the current CPU lane. Margin fragility: the repair arm is thin in the smoke. `decoy` clears the score margin by about +0.051 and `alias` by about +0.059, while both success deltas are +0.0625, exactly one extra success in a 16-seed smoke. The lock is therefore a real arbiter: a downgrade to `hybrid_no_niche` at 96 seeds would be the preregistration mechanism working, not a regression. Lock receipt, 2026-05-18: ```text npm run bayes:phase4b 3,072 trials in 1495.976s Exit gate: hybrid_no_niche (1/2 claim scenarios all-arms) ``` Lock receipt paths: - `results/bayes/phase4b-hybrid-lock/manifest.json` - `results/bayes/phase4b-hybrid-lock/summary.csv` - `results/bayes/phase4b-hybrid-lock/regret.csv` - `results/bayes/phase4b-hybrid-lock/replay-manifest.json` Lock gate rows: | Scenario | Hybrid score / success | HC-Sundog score / success | Score delta | Success delta | Hybrid evals | Bayes-adaptive evals | Arms | | --- | ---: | ---: | ---: | ---: | ---: | ---: | --- | | `decoy` | 0.179274 / 0.291667 | 0.141162 / 0.281250 | +0.038112 | +0.010417 | 1047.25 | 3153.58 | repair false; frugality true; load-bearing true | | `alias` | 0.248791 / 0.333333 | 0.097524 / 0.250000 | +0.151267 | +0.083333 | 988.08 | 4470.04 | repair true; frugality true; load-bearing true | Lock trigger diagnostics from `steps.jsonl`: | Scenario | Hybrid steps | Stuck-window fires | Posterior actions | Posterior refreshes | | --- | ---: | ---: | ---: | ---: | | `decoy` | 3233 | 354 | 314 | 354 | | `alias` | 3159 | 334 | 251 | 334 | Lock interpretation: Phase 4b does not earn a claim niche. The no-progress trigger is active and frugal, and the load-bearing ablation still shows the posterior can repair both claim regimes, but plain `hybrid` only clears the repair arm in `alias`. `decoy` falls below the fixed +0.10 score margin at lock scale, so the smoke niche is downgraded to `hybrid_no_niche`. This closes Phase 4 as a pre-registered negative for the frugal gated Hybrid claim. The lock-backed affirmative finding is narrower: `hybrid_posterior_decoy_disambig` repairs both `decoy` and `alias`; the dropped claim is the frugal gate, not posterior sufficiency. Treat the alias/decoy split as Phase 5 operating-envelope material rather than a Phase 4c redesign. ### Phase 5 - Operating Envelope Sweep Goal: map regions rather than narrate cherry-picked rounds. Phase 5 is descriptive, not a Hybrid claim gate. Hybrid remains reported-only after the Phase 4b lock negative. Frozen axes: - `decoy-strength in {0.50, 0.66, 0.82, 0.98}` with `sigmaScale = 0.82` held fixed. - `max-turns in {20, 30, 40, 60}`. - Sixteen swept `decoy` cells, plus fixed `alias` reference cells at each budget with amplitude `0.96`. Frozen modes, all reported and none gated: `oracle`, `bayes_misspecified`, `bayes_adaptive`, `hc_sundog`, `sundog_memory`, `hybrid`, `hybrid_posterior_decoy_disambig`, `random`. Per-cell classifier: reuse `phase2SeparationMargin = 0.10`. If the best non-yardstick family score is `< 0`, label `all_fail`; otherwise label `bayes_dominant`, `response_dominant`, or `hybrid_dominant` only when that family's best lane beats both other family bests by at least `0.10`; otherwise label `mixed`. Self-consistency anchor: the full `(decoy-strength=0.82, max-turns=40)` cell must reproduce the Phase 4b lock decoy deltas against `hc_sundog` for the shared lanes. The anchor checks against `results/bayes/phase4b-hybrid-lock/manifest.json`. Voided lock receipt, 2026-05-18: the first `phase5-envelope-lock` run wrote `results/bayes/phase5-envelope-lock/manifest.json` with `status: envelope_incomplete` because the anchor reported `selfConsistencyAnchor: mismatch_void`. The oracle and `hc_sundog` rows matched the Phase 4b lock, but posterior lanes did not, isolating the defect to the adaptive model slate: Phase 5 controller states were still passing `config.scenarios` into `adaptiveModelsForScenarios`, so Phase 5 missed the Phase 5 branch and used the Phase 2 adaptive families instead of the Phase 3 decoy/alias families. Treat that lock as void; do not interpret its envelope labels. Pre-registered amendment before resmoke, 2026-05-18: - Pass the full `config` object into `adaptiveModelsForScenarios` when creating adaptive controller state, so Phase 5 uses the frozen Phase 3 decoy/alias adaptive model slate. Phases 1-4 are unchanged because `isPhase5Config` is false there. - Expand the smoke decoy-strength grid from `{0.50, 0.98}` to `{0.50, 0.82, 0.98}`. Smoke becomes 8 cells / 512 trials and now includes the `(0.82, 40)` anchor, catching any future anchor mismatch at smoke scale. The anchor predicate, full grid, classifier, modes, and lock command remain unchanged. Resmoke after that amendment still returned `envelope_incomplete` with all 8/8 cells classified because the anchor comparator was exact summary-to-summary: it compared the 8-seed smoke summary to the 96-seed Phase 4b lock summary. That is stricter than the frozen "within seed-count variance" intent and is not a controller mismatch: the same seed subset from `results/bayes/phase4b-hybrid-lock/trials.jsonl` reproduced the smoke anchor exactly for `oracle`, `bayes_adaptive`, `hc_sundog`, `hybrid`, and `hybrid_posterior_decoy_disambig`. Pre-registered comparator repair before second resmoke, 2026-05-18: when the current run's seed window differs from the Phase 4b reference summary, the anchor compares against the same seed subset from the Phase 4b `trials.jsonl` receipt and remains exact. This adds no threshold and does not change the anchor predicate, grid, classifier, modes, or lock command. Commands: ```powershell npm run bayes:phase5:smoke npm run bayes:phase5 ``` Output directories: - Smoke: `results/bayes/phase5-envelope-smoke/` - Lock: `results/bayes/phase5-envelope-lock/` Frozen filenames under each output directory: - `manifest.json` - `trial-outcomes.csv` - `envelope-map.csv` - `aggregate-envelope.csv` - `best-by-cell.csv` - `cell-class-map.csv` - `cell-delta-map.csv` - `candidate-envelope.csv` Superseded smoke receipt, 2026-05-18, before the anchor-in-smoke amendment: ```text npm run bayes:phase5:smoke Bayes phase5-envelope-smoke: 384 trials in 37.588s (10.22 trials/s) Audits: pass Exit gate: envelope_mapped (6/6 cells classified) Wrote results\bayes\phase5-envelope-smoke ``` Amended smoke receipt, 2026-05-18: ```text npm run bayes:phase5:smoke Bayes phase5-envelope-smoke: 512 trials in 160.047s (3.2 trials/s) Audits: pass Exit gate: envelope_mapped (8/8 cells classified) Wrote results\bayes\phase5-envelope-smoke ``` Anchor receipt: `selfConsistencyAnchor.status = reproduced_p4b_lock_decoy`, using `referenceBasis = seed_matched_trials_jsonl`; all shared-lane score and success deltas match the Phase 4b seed-matched reference subset exactly. Amended smoke class map: | Cell | Class | Best family | Best mode | | --- | --- | --- | --- | | `decoy_strength_0p50_turns_20` | `bayes_dominant` | bayes | `bayes_adaptive` | | `decoy_strength_0p82_turns_20` | `bayes_dominant` | bayes | `bayes_adaptive` | | `decoy_strength_0p98_turns_20` | `bayes_dominant` | bayes | `bayes_adaptive` | | `alias_ref_turns_20` | `bayes_dominant` | bayes | `bayes_adaptive` | | `decoy_strength_0p50_turns_40` | `bayes_dominant` | bayes | `bayes_adaptive` | | `decoy_strength_0p82_turns_40` | `bayes_dominant` | bayes | `bayes_adaptive` | | `decoy_strength_0p98_turns_40` | `bayes_dominant` | bayes | `bayes_adaptive` | | `alias_ref_turns_40` | `mixed` | bayes | `bayes_adaptive` | The full lock is deferred by the long-run rule. The amended smoke has 512 trials; the full grid has 15,360 trials, with a higher average action budget. Budget-scaled estimate from the amended smoke is roughly 80-100 minutes: ```powershell npm run bayes:phase5 ``` Lock receipt, 2026-05-18: ```text npm run bayes:phase5 15,360 trials in 4618.776s (3.33 trials/s) Audits: pass Exit gate: envelope_mapped (20/20 cells classified) ``` Lock receipt paths: - `results/bayes/phase5-envelope-lock/manifest.json` - `results/bayes/phase5-envelope-lock/cell-class-map.csv` - `results/bayes/phase5-envelope-lock/aggregate-envelope.csv` - `results/bayes/phase5-envelope-lock/candidate-envelope.csv` Self-consistency anchor: **pass**. `status = reproduced_p4b_lock_decoy`, `referenceBasis = seed_matched_trials_jsonl`, 96 seed-matched reference trials. All five shared lanes (`oracle`, `bayes_adaptive`, `hc_sundog`, `hybrid`, `hybrid_posterior_decoy_disambig`) reproduce the Phase 4b lock decoy deltas at exactly zero score and success difference. The harness is trustworthy at lock scale; the result below is real, not an artifact of the earlier `mismatch_void` defect. Lock class map, aggregate: | Class | Cells | Swept decoy | Alias reference | | --- | ---: | ---: | ---: | | `bayes_dominant` | 20 / 20 | 16 / 16 | 4 / 4 | Every cell is `bayes_dominant` with `bestMode = bayes_adaptive`, across the full decoy-strength `{0.50, 0.66, 0.82, 0.98}` x budget `{20, 30, 40, 60}` grid and all alias-reference budgets. Phase 5 closeout interpretation: the envelope is **degenerate**: a single class over the entire pre-registered grid. `envelope_mapped` here means complete and anchor-valid, not a rich multi-class envelope. The roadmap exit criterion - "summarized as an operating envelope with failure classes, not a winner-take-all benchmark" - is therefore satisfied only trivially: the decoy-strength x budget slice contains no Bayes / response / hybrid boundary. This is a legitimate negative that sharpens, rather than weakens, the cross-phase claim. The Phase 3/4 failure boundary was a failure of misspecified / fixed-clean Bayes (`bayes_misspecified`) and of response and frugal-gated-hybrid controllers - not of adaptive Bayes. Once the Bayesian agent may adapt its model family (`bayes_adaptive` with the decoy/alias/symmetric mixture), it dominates the whole decoy-strength x budget envelope, including near-equal decoy (`0.98`), the alias reference, and the tightest budget (`20`). The Sundog / hybrid edge in this benchmark is specific to misspecified or fixed Bayes; it does not survive against correctly specified adaptive Bayes on these two axes. Consequence: any operating boundary lives on an axis not in this grid - most plausibly model-mismatch severity, the axis traded away at Phase 5 scoping, consistent with the entire Phase 2 to Phase 4 record. Phase 5 is closed as a pre-registered degenerate-envelope negative. The flat result is recorded as-run; the margin is not re-sliced and the grid is not re-cut to manufacture a boundary. A mismatch-severity sweep, if pursued, is a separate Phase 5b pre-registration with its own seed-matched anchor, not an amendment here. Exit criterion outcome: **met trivially / no boundary mapped.** The comparison is summarizable as an envelope, but the decoy-strength x budget envelope is single-class; the phase closes as a negative. Exit criterion: the comparison can be summarized as an operating envelope with failure classes, not as a winner-take-all benchmark. ### Phase 5b - Mismatch-Severity Sweep Goal: map the axis Phase 5 deliberately traded away: adaptive-mixture misspecification severity. Phase 5b is descriptive, not a new Hybrid gate. Hybrid remains reported-only. Frozen severity axis, implemented by filtering the Phase 3 adaptive model family list and re-normalizing the remaining models to a uniform prior: - `s0`: full Phase 3 mixture (`clean`, `decoy`, `alias`, `symmetric`) - equivalent to Phase 5 `bayes_adaptive`. - `s1`: remove `decoy`. - `s2`: remove `decoy`, `alias`. - `s3`: `clean` only - behaviorally equivalent to `bayes_misspecified`. The severity filter is applied in `createAdaptiveState` / `adaptiveModelsForScenarios`, so it affects `bayes_adaptive` and the adaptive-posterior Hybrid lanes. `oracle`, `random`, `hc_sundog`, `sundog_memory`, and `bayes_misspecified` are severity-invariant. Frozen grid: severity `{s0, s1, s2, s3}` x decoy-strength `{0.82, 0.98}`, budget fixed at `40` = eight swept cells. Modes are unchanged from Phase 5: `oracle`, `bayes_misspecified`, `bayes_adaptive`, `hc_sundog`, `sundog_memory`, `hybrid`, `hybrid_posterior_decoy_disambig`, `random`. Classifier and status are identical to Phase 5: dominance margin `phase2SeparationMargin = 0.10`, Hybrid reported-only with no gate arm, `envelope_mapped` / `envelope_incomplete`, non-fatal `pass: true`. Seed-matched anchor: `(s0, decoy-strength=0.82, max-turns=40)` must reproduce the Phase 5 lock `decoy_strength_0p82_turns_40` cell exactly against `results/bayes/phase5-envelope-lock/trials.jsonl` using `referenceBasis = seed_matched_trials_jsonl`. The anchor cell is in the smoke grid. Secondary consistency check, reported not gated: at `s3`, `bayes_adaptive` should match `bayes_misspecified` as a single clean-model mixture. The runner records the residual score and success deltas as `s3AdaptiveMinusMisspecified` columns. Commands: ```powershell npm run bayes:phase5b:smoke npm run bayes:phase5b ``` Output directories: - Smoke: `results/bayes/phase5b-mismatch-severity-smoke/` - Lock: `results/bayes/phase5b-mismatch-severity-lock/` Frozen filenames under each output directory remain the Phase 5 envelope set: `manifest.json`, `trial-outcomes.csv`, `envelope-map.csv`, `aggregate-envelope.csv`, `best-by-cell.csv`, `cell-class-map.csv`, `cell-delta-map.csv`, and `candidate-envelope.csv`. Smoke receipt, 2026-05-18: ```text npm run bayes:phase5b:smoke Bayes phase5b-mismatch-severity-smoke: 128 trials in 28.561s (4.48 trials/s) Audits: pass Exit gate: envelope_mapped (2/2 cells classified) Wrote results\bayes\phase5b-mismatch-severity-smoke ``` Anchor receipt: `selfConsistencyAnchor.status = reproduced_phase5_lock_cell`, `referenceBasis = seed_matched_trials_jsonl`, with eight seed-matched reference trials from the Phase 5 lock. All shared lanes (`oracle`, `bayes_misspecified`, `bayes_adaptive`, `hc_sundog`, `sundog_memory`, `hybrid`, `hybrid_posterior_decoy_disambig`, `random`) reproduce the Phase 5 lock anchor cell with exactly zero score and success difference. Smoke class map: | Cell | Class | Best family | Best mode | | --- | --- | --- | --- | | `severity_s0_decoy_strength_0p82_turns_40` | `bayes_dominant` | bayes | `bayes_adaptive` | | `severity_s3_decoy_strength_0p82_turns_40` | `mixed` | response | `sundog_memory` | Secondary sanity: at `s3`, `bayes_adaptive` and `bayes_misspecified` match exactly (`s3AdaptiveMinusMisspecifiedScoreDelta = 0`, `s3AdaptiveMinusMisspecifiedSuccessDelta = 0`). The full lock is deferred by the long-run rule. The smoke has 128 trials; the full grid has 6,144 trials. Linear estimate from the smoke is roughly 23 minutes: ```powershell npm run bayes:phase5b ``` Lock receipt, 2026-05-18: ```text npm run bayes:phase5b 6,144 trials in 889.476s (6.91 trials/s) Audits: pass Exit gate: envelope_mapped (8/8 cells classified) ``` Lock receipt paths: - `results/bayes/phase5b-mismatch-severity-lock/manifest.json` - `results/bayes/phase5b-mismatch-severity-lock/cell-class-map.csv` - `results/bayes/phase5b-mismatch-severity-lock/aggregate-envelope.csv` - `results/bayes/phase5b-mismatch-severity-lock/candidate-envelope.csv` Self-consistency anchor: **pass**. `status = reproduced_phase5_lock_cell`, `referenceBasis = seed_matched_trials_jsonl`, 96 seed-matched reference trials from `results/bayes/phase5-envelope-lock/trials.jsonl`. All eight shared lanes reproduce the Phase 5 lock `decoy_strength_0p82_turns_40` cell at exactly zero score and success difference. The harness is trustworthy at lock scale; the `s0` column is, by construction, the Phase 5 result re-derived (chain: 5b.s0 = Phase 5 = Phase 4b = Phase 3). Lock class map, aggregate: 3 `bayes_dominant`, 3 `mixed`, 2 `response_dominant` over 8 swept cells. | Severity | decoy 0.82 | decoy 0.98 | | --- | --- | --- | | `s0` full mixture | `bayes_dominant` (bayes-resp +1.329) | `bayes_dominant` (+1.417) | | `s1` -decoy | `bayes_dominant` (+0.345) | `mixed` (+0.244) | | `s2` -decoy -alias | `mixed` (+0.098) | `mixed` (best `hybrid_posterior_decoy_disambig`, bayes-hyb -0.009) | | `s3` clean-only | `response_dominant` (`sundog_memory`, -0.196) | `response_dominant` (`sundog_memory`, -0.180) | Phase 5b closeout interpretation: this is a genuine multi-class operating envelope, monotone in misspecification severity: **Bayes-dominant (`s0`-`s1`) -> contested `mixed` (`s1`-`s2`) -> response-dominant (`s3`)**, with decoy-strength a secondary modulator (`0.98` crosses one severity step earlier than `0.82`). The roadmap Phase 5 exit criterion - "an operating envelope with failure classes, not a winner-take-all benchmark" - is satisfied non-trivially, in contrast to the Phase 5 degenerate single-class result. Three framing constraints, pre-committed at audit and binding on any public copy: 1. The response win is `response_dominant` via `sundog_memory` and only at `s3`. Because `s3` is the `clean`-only mixture, it is behaviorally `bayes_misspecified`; this cell is the Phase 2 narrow-separation result re-derived inside the envelope, not an independent new win. The honest statement is: response control wins under severe clean-only Bayesian misspecification; partial misspecification (`s2`) is contested `mixed`, not a response win. "Sundog beats Bayes under mismatch" unqualified is not supported. 2. `s2` / `0.98` is the only cell where the Phase 4-dropped `hybrid_posterior_decoy_disambig` lane is nominally best (bayes-hybrid -0.009). The cell is `mixed` because no family clears the 0.10 margin, so this is not a `hybrid_dominant` claim and does not reopen the Phase 4 close. It is recorded as a faint reported-only transition-zone artifact. 3. The `s0` margins (+1.329 / +1.417) restate the Phase 5 degenerate result and the anchor-consistent nesting; reported as continuity, not novelty. Exit criterion outcome: **met non-trivially / boundary mapped.** The Bayes-versus-response operating boundary is misspecification severity: adaptive Bayes dominates with a correct or mildly degraded model family; response control overtakes only when the family collapses to clean-only; the intermediate band is contested. Recorded as-run; the margin is not re-sliced and no cell is re-cut to sharpen the boundary. ### Phase 6 - Photometric Port Frozen canonical record, 2026-05-18: Goal: port the standalone Bayes-vs-response comparison back to the core MuJoCo mirror-alignment experiment without changing the existing photometric receipts. Lanes: - `photometric` - `doa_direct` (privileged yardstick; excluded from classification) - `doa_noisy` (reported rail) - `random` (reported rail) - `bayes_particle` `bayes_particle` is a particle filter over laser `xy` in `[-0.4,0.4]^2` from the 8 detector intensities. It always predicts observations with the nominal model (`beam_sigma = 0.15`, `detector_noise_sigma = 0.0`); stress enters only through the environment, so mismatch is model-side and pre-registered. Particle counts are frozen at 64 for smoke and 128 for lock. Two single-axis severity ladders are mapped, nesting the committed stress receipts exactly: - `beam_sigma = [0.05, 0.10, 0.15, 0.25, 0.40]` at `detector_noise_sigma = 0.0` - `detector_noise_sigma = [0.0, 0.02, 0.05, 0.10, 0.20]` at `beam_sigma = 0.15` The shared nominal cell `(0.15, 0.0)` is counted once, yielding 9 distinct cells. The smoke is a strict seed-prefix run over `nominal`, `beam_sigma=0.40`, and `detector_noise_sigma=0.20`. Metric and classifier: - Primary metric: median `time_to_threshold`, right-censored at 500 steps. Censored seeds enter as 500. - Receipt extension: `experiments/analysis.py` now emits `time_to_threshold.ci95` using the same bootstrap estimator already used for `terminal_intensity.ci95`. - A class is assigned only between `photometric` and `bayes_particle`. A lane is dominant only when its median-time lead over the runner-up exceeds the `ci95` half-width of the better lane's per-seed `time_to_threshold`; otherwise the cell is `mixed`. - `terminal_intensity` is reported alongside the acquisition metric. Substrate-validity anchors: 1. The nominal `photometric` and `doa_direct` lanes reproduce the committed `results/analysis/analysis_summary.json` aggregates exactly at lock scale. The smoke checks the same deterministic receipt as a strict seed-prefix live re-execution before any lock is run. 2. Each ladder's `photometric` lane reproduces the committed `results/stress_tests/{beam_sigma,detector_noise}/sweep_summary.json` photometric column at lock scale. Smoke checks the included cells by seed-prefix live re-execution. Anchor clarification, recorded before the repaired smoke: the initial smoke receipt below is void because the anchor compared committed NPZ receipts to themselves. A valid anchor must re-execute `photometric` and `doa_direct` live through the shared Phase 6 env/seed/stressor path at the nominal cell, and must re-execute `photometric` live for the included stress cells. Those live traces must match the committed NPZ receipts within `1e-12` max absolute target-trace difference. Receipt replay may remain as loadability evidence for non-anchored rails, but it is not an anchor pass condition. Frozen commands: ```powershell $env:PYTHONPATH=(Resolve-Path ..).Path; python experiments/run_baseline_comparison.py --phase phase6-photometric-smoke --results-dir results/bayes/phase6-photometric-smoke --phase6-cells smoke --seeds 6 --steps 500 --particle-count 64 --conditions photometric doa_direct doa_noisy random bayes_particle $env:PYTHONPATH=(Resolve-Path ..).Path; python experiments/run_baseline_comparison.py --phase phase6-photometric-lock --results-dir results/bayes/phase6-photometric-lock --phase6-cells lock --seeds 30 --steps 500 --particle-count 128 --conditions photometric doa_direct doa_noisy random bayes_particle ``` Voided smoke receipt, 2026-05-18: ```text $env:PYTHONPATH=(Resolve-Path ..).Path; python experiments/run_baseline_comparison.py --phase phase6-photometric-smoke --results-dir results/bayes/phase6-photometric-smoke --phase6-cells smoke --seeds 6 --steps 500 --particle-count 64 --conditions photometric doa_direct doa_noisy random bayes_particle Photometric phase6-photometric-smoke: 90 trials in 41.933s (2.15 trials/s) Audits: pass Exit gate: envelope_mapped (3/3 cells classified) ``` This receipt is retained only as a voided implementation receipt. It is not claim support and must not be used to unlock the Phase 6 lock, because the anchor was circular. Voided smoke receipt paths: - `results/bayes/phase6-photometric-smoke/manifest.json` - `results/bayes/phase6-photometric-smoke/trial-outcomes.csv` - `results/bayes/phase6-photometric-smoke/candidate-envelope.csv` - `results/bayes/phase6-photometric-smoke/cell-class-map.csv` - `results/bayes/phase6-photometric-smoke/aggregate-envelope.csv` Voided smoke interpretation: the 3 / 3 `bayes_particle_dominant` class map is not interpretable because the anchor did not validate the live Phase 6 path. The repaired smoke must be run before any Phase 6 lock or result framing. Repaired smoke receipt, 2026-05-18: ```text $env:PYTHONPATH=(Resolve-Path ..).Path; python experiments/run_baseline_comparison.py --phase phase6-photometric-smoke --results-dir results/bayes/phase6-photometric-smoke --phase6-cells smoke --seeds 6 --steps 500 --particle-count 64 --conditions photometric doa_direct doa_noisy random bayes_particle Photometric phase6-photometric-smoke: 90 trials in 34.622s (2.60 trials/s) Audits: pass Exit gate: envelope_mapped (3/3 cells classified) ``` Repaired smoke anchor: **pass**. Live re-execution through the shared Phase 6 env/seed/stressor path matches committed receipts within the frozen `1e-12` tolerance: - nominal `photometric`: max absolute target-trace diff `1.39e-14` - nominal `doa_direct`: max absolute target-trace diff `4.77e-15` - included stress photometric cells: max absolute target-trace diff `1.81e-14` Repaired smoke class map: 3 / 3 `bayes_particle_dominant` cells over `nominal`, `beam_sigma_0p40`, and `detector_noise_0p20`. This is now a valid smoke-shape receipt, not a lock. The full 9-cell, 128-particle lock remains the arbiter before interpreting the surprising max-severity Bayes result. Phase 6 lock-interpretation constraints, frozen before the lock: 1. **Severity-axis weakness.** Phase 6 perturbs optical parameters (`beam_sigma`, `detector_noise_sigma`), which is weaker than the model-family ablation that produced the Phase 5b synthetic boundary. If the lock stays Bayes-dominant, the supported statement is that this optical-parameter misspecification axis does not break the nominal particle filter in the tested range. It does not refute the Phase 5b result; a wrong likelihood-structure axis would require a separate Phase 6b pre-registration. 2. **Smoke is not a claim.** The repaired smoke validates the live anchor and the shape of the result on a strict 6-seed prefix only. The public Phase 6 interpretation requires the frozen 30-seed lock, including the aggregate exact anchor against `analysis_summary.json` and the committed stress `sweep_summary.json` receipts. 3. **Speed, not accuracy.** The primary comparison is median `time_to_threshold`; `terminal_intensity` must be reported alongside it. If the lock preserves the smoke shape, public copy should say "far faster at comparable terminal accuracy," not "better" without qualification. If the lock confirms the smoke, the honest closeout is that the Phase 1-5b synthetic response-control edge does not replicate on the core photometric task under these tested optical/detector severities. If the lock downgrades the smoke, the smoke is recorded as an over-optimistic prefix and the lock result wins. In either case, the result is reported as-run; the grid is not re-cut and the severity axis is not changed after seeing the lock. Lock receipt, 2026-05-18: ```text $env:PYTHONPATH=(Resolve-Path ..).Path; python experiments/run_baseline_comparison.py --phase phase6-photometric-lock --results-dir results/bayes/phase6-photometric-lock --phase6-cells lock --seeds 30 --steps 500 --particle-count 128 --conditions photometric doa_direct doa_noisy random bayes_particle Photometric phase6-photometric-lock: 1350 trials in 1097.418s (1.23 trials/s) Audits: pass Exit gate: envelope_mapped (9/9 cells classified) ``` Anchor: **pass**, all five components gating (`anchorPassComponents`): nominal and stress live re-execution <= `1e-12` max target-trace diff; nominal `analysis_summary.json` aggregate exact (`time_to_threshold.{mean,median}`, `terminal_intensity.{mean,median,ci95}` all `diff == 0.0`); `photometric_vs_doa_direct_terminal_intensity` Mann-Whitney exact (U 526.0, p 0.264, `diff == 0.0`); stress `sweep_summary.json` mean_terminal exact (`diff == 0.0`). The recording gap in the first lock manifest was closed by manifest re-emit from existing artifacts plus fresh live anchor checks; no trial number changed. Lock class map: **8 / 9 `bayes_particle_dominant`, 1 `mixed`** over the 9-cell beam_sigma x detector_noise envelope. | Cell | Class | Best | Lead vs runner-up | ci95 half-width | | --- | --- | --- | ---: | ---: | | `nominal` (0.15, 0.0) | `bayes_particle_dominant` | bayes_particle | 177.0 | 0.37 | | `beam_sigma_0p05` | `mixed` | photometric | 0.0 | 51.7 | | `beam_sigma_0p10` | `bayes_particle_dominant` | bayes_particle | 181.0 | 65.3 | | `beam_sigma_0p25` | `bayes_particle_dominant` | bayes_particle | 25.0 | 0.17 | | `beam_sigma_0p40` | `bayes_particle_dominant` | bayes_particle | 23.0 | 0.17 | | `detector_noise_0p02/0p05/0p10/0p20` | `bayes_particle_dominant` | bayes_particle | 177.0 | 0.37 | Phase 6 closeout interpretation. The lock confirms the repaired smoke. On the core photometric mirror task, a nominal-only particle Bayes lane dominates the SCAN/SEEK/TRACK photometric controller on median `time_to_threshold` across the entire tested misspecification envelope, including the maximum tested optical (`beam_sigma=0.40`) and detector (`detector_noise=0.20`) severities. The Phase 1-5b synthetic response-control edge **does not replicate** on the core task under these tested optical/detector severities. Binding framing (frozen pre-lock, applied as-is): 1. **Severity-axis statement, not a refutation of Phase 5b.** Phase 6 perturbs optical parameters, which is a weaker misspecification than the Phase 5b model-family ablation. The supported claim is narrow: this optical-parameter axis does not break the nominal particle filter in the tested range. A wrong-likelihood-structure axis would require a separate Phase 6b pre-registration; Phase 5b stands. 2. **Speed at comparable terminal accuracy.** `terminal_intensity` is approximately tied across cells; the entire `bayes_particle` advantage is `time_to_threshold`. Public copy says "far faster at comparable terminal accuracy," not "better" unqualified. 3. **The lone `mixed` cell is a no-decision, not a photometric win.** `beam_sigma_0p05` reports best=`photometric` with lead 0.0 and ci95 half-width 51.7 - the families are statistically indistinguishable there, not a photometric-dominant regime. It must not be cited as a Sundog/response foothold. Recorded as-run. The grid is not re-cut and the severity axis is not changed after the lock. Exit criterion outcome: **met.** A reviewer can now see that the Phase 1-5b response-control result does not transfer to the core photometric task under the tested optical/detector misspecification - i.e., the synthetic edge is, on this axis, a shadow-field/grid-substrate effect rather than a core-task property. Phase 6 closes as a pre-registered non-replication. Exit criterion: a reviewer can see whether the Phase 1-5b result - response edge only under Bayesian misspecification severity - is a shadow-field toy artifact or replicates on the core photometric task. The phase is descriptive, non-fatal, and does not create a winner-take-all claim. ### Phase 6b - Core Photometric Structure-Mismatch Hard Stop Phase 6b is the final single-variable transfer test flagged by Phase 6: wrong likelihood structure on the core photometric task. It maps; it does not rescue a controller, reopen Hybrid, add a new lane, or create a Phase 6c. Frozen cell set: nominal optics only, `beam_sigma = 0.15` and `detector_noise_sigma = 0.0`, crossed with structure rungs `{s0,s1,s2,s3}`. The structure ladder affects only the `bayes_particle` assumed model: `s0` is the Phase 6 identity reflected-likelihood path; `s1` removes reflection from the assumed model; `s2` keeps only the target-detector channel; `s3` is uninformative. The same assumed model must drive belief update, target-matrix scoring, and candidate-action construction. The environment, seed path, detector observations, and assumed sigma stay Phase 6 nominal throughout. Lock cells are `nominal x {s0,s1,s2,s3}` (4 cells); smoke cells are `nominal x {s0,s3}` (2 cells). Classifier and reporting are unchanged from Phase 6: median `time_to_threshold` primary, ci95-half-width dominance, `doa_direct` excluded as a privileged yardstick, `doa_noisy` and `random` reported rails, descriptive/non-fatal. Particle counts remain 64 for smoke and 128 for lock; smoke seeds are the strict `range(6)` prefix of the lock `range(30)`. Anchor discipline is Phase 6 plus the nominal structure nest. Applicable gating components are `nominalLiveReexecution`, `analysisSummaryAggregate`, `mannWhitneyTests`, and `structureS0Nest`. With no stress cells, `stressLiveReexecution` and `stressSweepAggregate` have no inputs and must be recorded as `not_applicable`, not silently passed. `structureS0Nest` requires the `s0` `bayes_particle` nominal cell to reproduce the Phase 6 lock nominal `bayes_particle` episode receipt exactly (`diff == 0.0`). Because the smoke lane itself keeps the frozen 64-particle smoke count while the Phase 6 lock receipt is 128-particle, the smoke nest is an auxiliary s0 live re-execution at the 128-particle reference count; classification rows still use the frozen smoke particle count. Binary exit: `response_edge_transfers` if any structure rung produces a `photometric_dominant` cell; otherwise `core_task_bayes_speed_dominant`. Hard stop either way; no Phase 6c is bundled. Frozen commands: ```powershell $env:PYTHONPATH=(Resolve-Path ..).Path; python experiments/run_baseline_comparison.py --phase phase6b-structure-smoke --results-dir results/bayes/phase6b-structure-smoke --phase6-cells smoke --seeds 6 --steps 500 --particle-count 64 --conditions photometric doa_direct doa_noisy random bayes_particle $env:PYTHONPATH=(Resolve-Path ..).Path; python experiments/run_baseline_comparison.py --phase phase6b-structure-lock --results-dir results/bayes/phase6b-structure-lock --phase6-cells lock --seeds 30 --steps 500 --particle-count 128 --conditions photometric doa_direct doa_noisy random bayes_particle ``` Void smoke receipt, 2026-05-18: ```text $env:PYTHONPATH=(Resolve-Path ..).Path; python experiments/run_baseline_comparison.py --phase phase6b-structure-smoke --results-dir results/bayes/phase6b-structure-smoke --phase6-cells smoke --seeds 6 --steps 500 --particle-count 64 --conditions photometric doa_direct doa_noisy random bayes_particle Photometric phase6b-structure-smoke: 60 trials in 73.615s (0.82 trials/s) Audits: pass Exit gate: core_task_bayes_speed_dominant (2/2 cells classified) ``` This smoke is **void** and not interpretable. The structure axis did not bite: both `s0` and `s3` classified `bayes_particle_dominant`, with median `time_to_threshold = 11.0` and `n_failed = 0` in both endpoint cells. Root cause: candidate actions were still built by `optics.optimal_joint_angles`, a true-optics inverse solver, so each rung retained a privileged optics-optimal aim for every hypothesis. The structure filter reweighted belief among true-optics actions but did not faithfully control behavior. Fidelity clarification, frozen before re-smoke: the assumed forward model `predict_r(laser_xy, action)` must drive all three Bayes-particle paths: belief update, target matrix, and candidate-action construction. `s0` keeps `optics.optimal_joint_angles`, preserving the Phase 6 byte-nest, because that solver is the exact analytic argmax of the `s0` reflected model. `s1` and `s2` must use grid argmax over the shared 11x11 joint-angle lattice from `optics.py` (`[-JOINT_LIMIT, JOINT_LIMIT]` on each axis, the same lattice used to seed `optimal_joint_angles`). `s3` has `predict_s3 == 0`, so the grid argmax ties and all particles map to `G[0]`, yielding a fixed, observation-independent action. `_precompute_target_matrix` and `_update` remain structure-filtered through `_predict_for_action`. Sharpened smoke validity gate, frozen before re-smoke: on `nominal x {s0,s3}`, `structureS0Nest` must remain `max_abs_diff == 0.0`, `s0` must classify `bayes_particle_dominant`, and `s3` must **not** classify `bayes_particle_dominant` under the existing ttt-primary + ci95-half-width classifier. If `s3` still classifies `bayes_particle_dominant`, the structure ablation remains unresolved; stop, do not lock, and do not interpret it as `core_task_bayes_speed_dominant`. Repaired smoke receipt, 2026-05-18: ```text $env:PYTHONPATH=(Resolve-Path ..).Path; python experiments/run_baseline_comparison.py --phase phase6b-structure-smoke --results-dir results/bayes/phase6b-structure-smoke --phase6-cells smoke --seeds 6 --steps 500 --particle-count 64 --conditions photometric doa_direct doa_noisy random bayes_particle Photometric phase6b-structure-smoke: 60 trials in 42.569s (1.41 trials/s) Audits: pass Exit gate: response_edge_transfers (2/2 cells classified) ``` Repaired smoke anchor and validity: **pass**. `structureS0Nest` remains exact against the Phase 6 lock nominal `bayes_particle` receipt (`max_abs_diff = 0.0`); `stressLiveReexecution` and `stressSweepAggregate` remain `not_applicable`; aggregate and Mann-Whitney checks remain `lock_only`. Smoke validity passes: `s0` is `bayes_particle_dominant`, while `s3` is `photometric_dominant`, not `bayes_particle_dominant`. Endpoint rows: | Cell | Class | Best | Runner-up | Lead | ci95 half-width | Bayes median ttt | Bayes n_failed | | --- | --- | --- | --- | ---: | ---: | ---: | ---: | | `nominal_structure_s0` | `bayes_particle_dominant` | `bayes_particle` | `photometric` | 178.0 | 1.0 | 11.0 | 0 | | `nominal_structure_s3` | `photometric_dominant` | `photometric` | `bayes_particle` | 311.0 | 13.333333333333329 | 500.0 | 6 | Phase 6b lock-interpretation constraints, frozen before the lock: 1. **Smoke is not the verdict.** The repaired smoke proves only that the endpoint ablation is valid: `s0` still nests the Phase 6 nominal Bayes lane exactly, and `s3` now degrades enough that the structure axis bites. The scientific binary exit is decided only by the 30-seed, 4-cell lock. The smoke headline `response_edge_transfers` is a shape/validity receipt, not a claim-grade conclusion. 2. **Total collapse is distinct from partial misspecification.** The smoke flip occurs only at `s3`, where `predict_s3 == 0` makes the Bayes particle effectively model-less and observation-independent. A response edge over `s3` alone is a weak transfer result. The meaningful transfer question is whether the edge holds under partial wrong structure (`s1` no-reflection or `s2` target-only), and those cells are lock-only. Closeout copy must report `s3`-only transfer separately from `s1`/`s2` transfer and must not state "the response edge transfers to the core task" unless the lock supports it. Lock receipt, 2026-05-18: ```text $env:PYTHONPATH=(Resolve-Path ..).Path; python experiments/run_baseline_comparison.py --phase phase6b-structure-lock --results-dir results/bayes/phase6b-structure-lock --phase6-cells lock --seeds 30 --steps 500 --particle-count 128 --conditions photometric doa_direct doa_noisy random bayes_particle Photometric phase6b-structure-lock: 600 trials in 388.494s (1.54 trials/s) Audits: pass Exit gate: response_edge_transfers (4/4 cells classified) ``` Lock receipt paths: - `results/bayes/phase6b-structure-lock/manifest.json` - `results/bayes/phase6b-structure-lock/cell-class-map.csv` - `results/bayes/phase6b-structure-lock/candidate-envelope.csv` Anchor: **pass**, all applicable components gating at `seeds == 30`. `structureS0Nest` exact against the Phase 6 lock nominal `bayes_particle` receipt (`max_abs_diff = 0.0`); `nominalLiveReexecution` pass; `analysisSummaryAggregate` evaluated (not `lock_only`) with `photometric` and `doa_direct` `max_abs_diff = 0.0` on every field; Mann-Whitney `photometric_vs_doa_direct_terminal_intensity` actual == expected (U 526.0, p 0.264, `diff = 0.0`); `stressLiveReexecution` and `stressSweepAggregate` `not_applicable` (no stress cells in Phase 6b). No recording gap; the substrate-validity and s0-nest hard gates are demonstrated and gating. Lock 4-rung map: | Cell | Structure | Class | Bayes median ttt | Bayes n_failed | Lead | | --- | --- | --- | ---: | ---: | ---: | | `nominal_structure_s0` | correct optics | `bayes_particle_dominant` | 11.0 | 0 | 177.0 | | `nominal_structure_s1` | no-reflection (partial) | `bayes_particle_dominant` | 11.0 | 0 | 177.0 | | `nominal_structure_s2` | target-only (partial) | `bayes_particle_dominant` | 11.0 | 0 | 177.0 | | `nominal_structure_s3` | model-less (`predict == 0`) | `photometric_dominant` | 500.0 | 30 | 312.0 | Phase 6b closeout interpretation, applying the pre-frozen constraints: The raw exit string `response_edge_transfers` is emitted **only because the `s3` cell flips**. Per frozen constraint 2, `s3` is the degenerate, observation-independent, model-less limit (`predict_s3 == 0`); a response controller beating a model-less agent is a near-tautological transfer, not evidence that the response edge reaches the core task. It is recorded here as the `s3`-only degenerate result and is **not** the headline. The substantive, non-degenerate result is `s1` and `s2`: under partial wrong likelihood structure (no-reflection, target-only) the particle Bayes lane is **undegraded** - median `time_to_threshold` 11.0, zero failures, identical to `s0`, same 177-step lead. Partial structural misspecification produces no measurable convergence-speed degradation on the core photometric task. This is robust, not a non-biting artifact: the ladder demonstrably bites at `s3` (500.0, 30/30 failed) and `s0` nests the Phase 6 lock exactly, so the structure mechanism is active across all rungs. Therefore: the Phase 1-5b synthetic response-control edge **does not transfer** to the core photometric task under any non-degenerate wrong-structure regime. It appears only when Bayes is stripped to a model-less control. This confirms and strengthens the Phase 6 parameter-axis conclusion on the true structural axis (the Phase 5b analog): the synthetic Sundog/response edge is substrate-specific, not a core-task property. "The response edge transfers to the core task" is **not supported** and must not appear in public copy. Recorded as-run. No grid recut, no structure-axis change, no Phase 6c. Exit criterion outcome: **resolved.** The last live scientific ambiguity - whether wrong likelihood structure bridges synthetic to core - is closed: non-transfer under partial misspecification, degenerate-only flip at `s3`. The Bayes-vs-Sundog experimental arc (Phases 1-6b) is complete. ### Phase 7 — Public Visualization and Motion Rail Card Goal: make the comparison legible on `sundog.cc` without flattening it into marketing. Deliverables: - `bayes.html` interactive visualization. - Split-screen mode: - left: posterior heatmap, entropy, most likely hypothesis; - right: response trace, scan/seek/track phase, best observed signal. - Round selector: Known Model, Warped Field, Decoy Field, Hybrid. - Rail card content for the homepage / applications gallery. - Short video clips for the highlight rail. Suggested rail card: ```text Title: Posterior vs. Halo Status: Comparative Benchmark Bayes: evidence -> posterior -> action. Sundog: probe -> response -> control. Question: must the agent infer the hidden world, or can it act from the trace? Stamp sequence: MODEL KNOWN / FIELD READABLE / SIGNAL ALIASED / HYBRID NEEDED ``` Exit criterion: the public artifact says "complementary" more clearly than it says "competition." ### Phase 8 — Writeup, Claim Ratchet, and Site Integration Goal: update the research docs with the strongest claim actually earned. Immediate work after the Pressure Mines applied lane: - Backfill [`BAYESIAN_FLOOR_PROFILE_TEMPLATE.md`](BAYESIAN_FLOOR_PROFILE_TEMPLATE.md) with the fields that proved reusable in Balance and Mines: admission profile, same-observation seed policy, claim-gate fields, reported-only lanes, posterior-cell slices, and public interpretation copy. *(landed 2026-05-17)* - Keep `alignment.html` as the public Bayes/Sundog side-by-side. The Mines row now has the explicit middle claim: stronger than "Sundog beat naive," weaker than "Sundog beat Bayes." - Phase 1 exact Bayes-Correct reference task has landed: small hidden-source grid, correct likelihood, exact posterior, and Sundog response-control comparator. The 128-seed lock passes the known-model gate. - Phase 2 mismatch slate has landed: five field variants, fixed clean `bayes_misspecified`, finite-mixture `bayes_adaptive`, `hc_sundog`, and `sundog_memory`. The lock finds one real separation (`anisotropic`) plus boundary cells, so the claim is complementary and narrow. - Add `bayes.html` only after the standalone benchmark has real run artifacts. Do not publish a decorative Bayes page before Phase 1/2 receipts exist. - Add a public rail card only if Round 3 or Phase 5 includes a real failure boundary. - Update `docs/PROMO_HIGHLIGHTS.md` with a cautious comparison paragraph. - Update `docs/SCIENTIFIC_CRITERIA.md` if the comparison introduces a stronger target-unaware baseline for the core photometric claim. Claim ratchet candidates: > *If Bayes-Correct wins Round 1 and Sundog wins at least one mis-specified > round:* Bayes and Sundog are complementary. In the tested hidden-source > family, posterior inference wins when the likelihood is right; response > control remains useful when the model is wrong but the field is readable. > > *If aliased fields break Sundog:* Sundog requires the response trace to be > meaningfully coupled to the hidden target. When the trace improves for the > wrong reason, posterior or hybrid hypothesis management is required. > > *If Bayes dominates broadly:* Sundog should be framed as a lightweight > controller for cases where maintaining a posterior is too expensive or > unnecessary, not as a general partial-observation replacement. > > *If Hybrid dominates:* The next program direction should be hybrid: use > Sundog for local response control and Bayes for ambiguity, reset, and > hidden-cause disambiguation. Exit criterion: docs, rail copy, and public language all agree on the same boundary. ## Implementation Notes ### Suggested files ```text sundog/ ├── docs/ │ └── SUNDOG_V_BAYES.md ├── bayes.html ├── js/ │ ├── bayes-core.mjs │ ├── bayes-controllers.mjs │ ├── bayes-visualizer.mjs │ └── bayes-replay.mjs ├── scripts/ │ └── bayes-phase1-reference.mjs └── results/ └── bayes/ ├── phase1-reference/ ├── phase2-mismatch/ ├── phase3-aliasing/ ├── phase4-hybrid/ └── phase5-envelope/ ``` ### Suggested npm commands ```bash npm run bayes:phase1 npm run bayes:phase2 npm run bayes:phase3 npm run bayes:phase4 npm run bayes:phase5 npm run bayes:replay -- --phase phase3-aliasing --seed 42 ``` ### Minimal result schema Each trial row should include: ```text run_id seed phase scenario controller_family controller_variant hidden_target_id model_status # correct / misspecified / adaptive / none field_variant # gaussian / warped / clipped / decoy / alias / drift sensor_variant # full-local / sparse-local / delayed / noisy action_budget compute_budget_ms terminal_response time_to_threshold cumulative_response wrong_lock reacquired_after_drift posterior_entropy_final posterior_mass_true sundog_best_observed failure_label ``` ## Public Copy Bank Use: - Bayes turns evidence into belief. Sundog turns response into control. - The posterior draws a map. The halo moves the hand. - When the model is right, Bayes should win. - When the map lies but the response still answers, Sundog can hold. - When the trace improves for the wrong reason, Sundog fails. - The hybrid is not a compromise. It is the boundary made useful. Avoid: - Sundog beats Bayes. - Bayes needs direct sight. - Inference is obsolete. - The halo is the posterior. - The signal is always truthful. ## Open Questions 1. Phase 1 first substrate is resolved: discrete grid exact posterior first. Continuous particles wait for the later continuous-field extension. 2. Should Bayes-Adaptive use Gaussian processes, particles with hyperparameter learning, or a small mixture model? Recommendation: mixture first, GP second. 3. Should the public failure card be called **SIGNAL ALIASED**, **BOUNDARY FOUND**, or **HYBRID NEEDED**? Recommendation: use SIGNAL ALIASED as the Round 3 stamp and HYBRID NEEDED as the concluding stamp. 4. Does the Bayes comparison live as a core research benchmark or as a workbench application? Recommendation: core research benchmark until Phase 5, then operating-envelope workbench if the browser page ships. 5. Should the photometric paper be updated immediately to mention Bayesian baselines? Recommendation: not until Phase 6 produces a real photometric Bayes comparator. ## Closing Note This roadmap should make Sundog smaller and stronger. If Bayes wins, the program learns when explicit belief is required. If Sundog wins, the program earns a cleaner statement about response control. If both win in different cells, the program gets the best possible public posture: not a universal claim, but a map. The most important card is the failure card. Sundog should fail when the signal is aliased, because that failure teaches the theorem's boundary: indirect signal is only useful when the trace remains coupled to the thing that matters.