From e94bd80bdeb5d2672a3b7a33ee4d9153855718fc Mon Sep 17 00:00:00 2001 From: xtkuang <87661715@qq.com> Date: Mon, 27 Jul 2026 18:33:37 +0800 Subject: [PATCH] Record locked v3 validation outcome --- README.md | 3 + ...IRMATION_AND_NETWORK_STAGE_B_2026-07-27.md | 121 ++++++++++++++++++ 2 files changed, 124 insertions(+) diff --git a/README.md b/README.md index d2b1b22..15ca6da 100644 --- a/README.md +++ b/README.md @@ -98,6 +98,9 @@ manuscript Results. The companion formula-linked workbook is The expanded v3 confirmation, network endpoint semantics, Stage-B calibration outcome, and frozen locked thresholds are recorded in `docs/calibration/V3_CONFIRMATION_AND_NETWORK_STAGE_B_2026-07-27.md`. +The first locked execution passed stable contact and the active H4 challenge +20/20, but the network final gate passed 199/200 because one 2% loss instance +exceeded the frozen maximum source-age limit. The audit preserves that failure. ## Manuscript build diff --git a/docs/calibration/V3_CONFIRMATION_AND_NETWORK_STAGE_B_2026-07-27.md b/docs/calibration/V3_CONFIRMATION_AND_NETWORK_STAGE_B_2026-07-27.md index 907a1ec..a4681ce 100644 --- a/docs/calibration/V3_CONFIRMATION_AND_NETWORK_STAGE_B_2026-07-27.md +++ b/docs/calibration/V3_CONFIRMATION_AND_NETWORK_STAGE_B_2026-07-27.md @@ -154,3 +154,124 @@ The proposed-only locked protocol is about the proposed loop under the registered emulator. It cannot support superiority over mapping baselines; that requires a separate locked comparison including `direct_energy` and `matched_wrench_energy`. + +## Locked execution outcome + +All three locked studies were executed from clean commit +`3329939ddec2711460da0873443584bf403720b3`. The master and slave URDF hashes +recorded in every batch manifest are respectively +`3890cd584e190a92bbffc74137dc00f49eacf6183f4e5886816eb0203a47e788` +and +`c5e0cc947122fe3ee7737852cb6cc07f014e529e46c22e17031a582c29acf5ac`. +Batch validation reported no structural errors or warnings. + +| Locked study | Registered outcome | Plan hash | Metric hash | Row hash | +|---|---:|---|---|---| +| Stable contact | 20/20 pass | `6da36a4853b227e8ca41c02589ff67a2edf90f4227dfa1aa980c20464a08988c` | `43d83c3e5482b2ed685accb5752b278fd3c7ecb60960376c4dc2a6d0b77a0e6f` | `437ae8670e8d3b4341b221835e3721a23f221e1e727e72a7ce8c9809480186fa` | +| Synthetic H4 challenge | 20/20 pass | `0c29157753b6c2fb038253e1a26ea14f83af8c36ee3a37edb8a33e14a2177d86` | `5a28a13305f124282033888420ae865e536599ab7a46ccb1ef53b8cdb1e3a8cb` | `1615026b46700a22e96a5677c734e93b36a81fc9223c45d3091b7e9d2a7d786a` | +| Proposed-only network | **199/200 final-gate pass** | `152df838b90582b3c1f4e5017e1cd0e7203e378b2be7931be70338b81b0bfd06` | `e65b738d2c05c8d146e837489d5047b102b04e22a98733b39717076f64b15c14` | `24432fcd8cde88cadc54b0e8557c5aeb3ac89851e1ed7216aeff08b747227b6e` | + +### Stage-A locked confirmation + +The stable-contact study passed all bilateral gates in all 20 disjoint-root +instances. Master and slave tracking RMSE ranges were +`0.03587–0.03615 rad` and `0.04586–0.04619 rad`; contact-active RMS was +`0.10597–0.10755 N`. Energy projection was inactive, as intended for the +stable-contact study. + +The synthetic challenge triggered the supervisor in every instance: + +- projection-intervention fraction: `0.04533–0.29500`; +- projection distortion \(D_{\mathrm{proj}}\): `0.02781–0.14921`; +- unsupervised shadow floor deficit: + `3.66e-5–1.32e-4 J`; +- raw probe work: `0.00720–0.00756 J`; +- downstream modification after energy projection: zero. + +All 15 challenge/audit outcomes passed in all 20 instances. The challenge is +synthetic and explicitly ineligible for H3; it establishes neither human nor +physical-device safety. + +### Network locked failure + +The network batch completed all 200 trials without an execution failure, and +its paired-degradation gate passed 200/200. Nevertheless, its registered final +result is a failure because the local network gate passed only 199/200. + +The sole failed trial was: + +```text +trial_id: trial-03311565ca2fa456 +trajectory: free_space +profile: symmetric_40ms_loss_2pct +replicate: 18 +failed gate: network_forward_age_gate_pass +observed max: 0.118 s +frozen limit: 0.100 s +``` + +The received forward sequence jumped from 44 to 48. Three consecutive 50 Hz +mapping updates were absent, so the held packet accumulated +`40 ms` transport delay, `60 ms` of missed updates, and `18 ms` of ordinary +mapping hold age. Its source age reached `118 ms` before the next packet +arrived. Source age exceeded `100 ms` for nine 500 Hz samples (`18 ms`). The +silence since the last packet arrival was only `78 ms`; the channel therefore +remained active, its timeout fraction was zero, and the registered arrival +timeout was `200 ms`. This is a stale-command tail event, not a disconnect. It +remains a formal failure because source age and arrival timeout were +deliberately separate endpoints. + +The failed trial still passed every tracking, paired-degradation, free-space +contact, H4, and bilateral safety gate. Across the complete network study: + +- paired tracking and paired contact gates passed in all applicable trials; +- the largest paired tracking degradation was `2.550 mrad`, below the frozen + `5 mrad` limit; +- the largest absolute contact-RMS relative change was `1.727%`, below the + frozen `10%` limit; +- H4 accounting and bilateral safety gates passed 200/200; +- all recorded joint, velocity, acceleration, actuator, haptic, and wall-force + limit fractions were zero; +- all forward and return timeout fractions were zero. + +These facts do not override the registered 199/200 final result. + +The fixed-RTT directional result was reproduced over 20 locked instances: +changing `40/40 ms` to `20/60 ms` reduced mean contact reference-lag RMSE from +`3.489 mrad` to `1.746 mrad`, while increasing mean return-feedback-lag RMSE +from `14.952 mNm` to `18.302 mNm`. + +Local locked evidence is stored under: + +- `output/experiments/bilateral-locked-v3-stable-20260727`; +- `output/experiments/bilateral-locked-v3-energy-20260727`; +- `output/experiments/bilateral-network-v3-locked-20260727`. + +The network seed-provenance hash is +`a7ea2be5f89ec0150c7dc3a78fc2e1cc01e685dd68f090a086eb411fd066bb08`. + +### Consequence for the next protocol + +The v3 network result must remain 199/200. The `0.100 s` limit must not be +changed to `0.120 s` and retroactively applied to this batch. + +A subsequent protocol may be designed only as a new, explicitly post-v3 study +with fresh disjoint root seeds. Before those seeds are generated, it should: + +1. separate a distributional freshness endpoint, such as predeclared p99 + source age, from a hard maximum-age or timeout safety endpoint; +2. justify the hard age limit from the intended command rate and real system + safety requirement rather than from the observed `118 ms` value; +3. include registered burst-loss conditions or measured network traces, since + independent 2% loss can still produce consecutive losses; +4. keep availability, freshness, and tracking degradation as separate + outcomes; +5. if `100 ms` is intended as a safety boundary, implement and register an + online source-age guard with explicit degraded-mode, hold/deceleration, and + recovery behavior instead of relying only on the current `200 ms` + last-arrival timeout. + +Only after that proposed-loop protocol is settled should a new three-method +locked comparison be registered. Such a comparison needs fresh seeds, +method-specific nominal references, predeclared primary contrasts, and +multiple-comparison control; it cannot be appended to the present batch.