Chief Scientist Perspective | 2026-09-13 Industry Review
We first place the core claims of the public paper EgoLive (2026) under the microscope: the large-scale open Ego collection specification of JoyEgoCam stereo 60fps and 200Hz IMU. This does not mean TOBI owns or operates this public dataset; this article discusses how to translate comparable specifications into procurable, auditable collection hardware and delivery language.
EgoLive's specifications are high, but a preliminary-stage pilot may not require E8. Compare E2 and E8 in terms of frame rate, trajectory, and budget during the pilot phase. For teams planning Ego datasets, VLA training, and robot transfer around 2026, this emphasizes auditable delivery, versioned calibration, and hard-sync QC, avoiding misreading public benchmarks as any single vendor's exclusive capability.
What the Paper Timeline Tells Us
Around 2026, EgoLive pushed the community discussion toward a clear question: the JoyEgoCam stereo 60fps and 200Hz IMU large-scale open egocentric capture specifications. The timeline tells us that the value of public benchmarks lies in defining "what counts as a qualified first-person/embodied sample," not in implying that any hardware vendor automatically possesses data assets of equivalent scale. TOBI cites this paper solely for specification comparison and procurement language translation.
If EgoLive is included in the 2026-09-13 procurement review, three layers should be distinguished: the timing of the paper's release and open-source link, the timing of community reproduction and secondary annotation, and the timing when enterprise-built/purchased data actually enters the repository. Between these three layers, there are often method iterations from 2026 to 2026; what hardware teams should track is field/schema changes, not chasing paper titles.
For compare-type tasks (this article's slug: news-seo-031-e2-vs-egolive-cost), the paper timeline also indicates when "failure samples and long-tail scenarios" were written into the benchmark protocol. Constraints that appear later often affect VLA training more than early scale numbers; device selection should reserve the ability to retain pre- and post-contact time windows.
Dataset Boundary Statement
EgoLive and its link https://arxiv.org/abs/2604.23570 in this article are within a public research context; TOBI does not claim ownership, hosting, or exclusive distribution of this dataset.
Chief Scientist's Perspective: Turning Paper Claims into Verifiable Metrics
Starting from EgoLive (2026), the large-scale open Ego capture specs of JoyEgoCam stereo 60fps and 200Hz IMU. Scientific readers should require vendors to provide a reproducible experiment package: fixed firmware, calibration version, recording scripts, and sync reports. TOBI does not own the public benchmark; the matrix evaluation only verifies whether the hardware fields can support similar task designs.
When comparing against third-party specifications such as JoyEgoCam (a third-party capture device described in the EgoLive paper, not a TOBI product), device ownership and licensing boundaries should be listed separately to avoid misinterpreting the paper's capture pipeline as TOBI's factory default configuration.

Acceptable requirements for capture hardware
Translating EgoLive's one-line claim into procurement language starts with defining "acceptable fields" rather than restating the JoyEgoCam stereo 60fps and 200Hz IMU large-scale open egocentric capture specifications. Hardware acceptance must be independently reproducible by third parties: given fixed firmware, a fixed calibration package, and a fixed recording script, anyone can export the same timeline semantics.
- Hard synchronization: whether each camera/IMU/gesture stream shares an acceptable clock domain (target jitter and measured distribution are recorded in the data card).
- Calibration: intrinsic and extrinsic parameters, along with model versions such as Kannala-Brandt, are frozen with the firmware; changing the camera, ring, or wrist unit requires regression testing.
- Container: frame timestamp units and dropped-frame markers in dual-stream HEVC/fMP4 or MCAP and similar formats must not be lost during transcoding.
- Sharding: shard duration for long-horizon dex/collaborative tasks, ~5 GB upload limits, and DatasetServer health checks.
- Failure samples: whether contact-invisible, occluded, and desynced segments are retained and labeled rather than silently deleted.
- Compliance: operator authorization, site permissions, and de-identification policies are included as delivery attachments; unrelated to the licensing of publicly released academic datasets.
The following sequence is suitable for a scientist team to write into an SOP (2026-09-13 version), bound to the matrix evaluation prototype:
- Freeze the matrix firmware and SDK versions, and the minimal script set for reproducing EgoLive comparison experiments.
- Bench sync testing: measure cross-path offset using LED / mechanical triggers to produce a baseline report.
- Worn-state spot check: sample 50 contact-frame segments for multi-view consistency across scenes tagged matrix, E2, EgoLive, comparison.
- Ingestion QC: outputs sync_health, drop_frame, and calib_version fields.
- Training feedback: compare contact success rates using the hard-sync subset versus the soft-alignment subset, and record the results in the procurement review.

Mapping to TOBI matrix product capabilities
TOBI's matrix capture solution emphasizes multi-camera, multi-operator, and unified timeline lab SOPs; it is often networked with E6 for collaborative egocentric capture, cross-embodiment recipes, and large-scale sharded ingestion.
In comparison with EgoLive: JoyEgoCam's large-scale open Ego capture specifications with stereo 60fps and a 200Hz IMU. On the device side, these should be translated into clickable product fields — for example, frame rate / exposure / IMU frequency, whether on-device SLAM serves as a trajectory prior, and whether gesture / ring / wrist devices share the same time domain. Readers should verify the README and delivery package at /products.html and /docs.html#device-sdk rather than assuming that paper configurations equal factory defaults.
When pillar=compare and the matrix is the primary recommendation, it is advisable to split "paper metrics" into Must (without which there is no purchase) and Stretch (R&D co-creation). Must typically covers hardware sync, calibration version, and recording container time semantics; Stretch may cover motion-capture bridging, ten-channel expansion, or cross-lab matrix standardization.
- E2 positioning: stereo global shutter, offline reconstruction friendly, lightweight human demo front station (can still serve as a networking secondary path when this article primarily recommends matrix).
- E6 specification: six channels + on-board SLAM + gesture recognition; host time-stamping and sharded upload are the primary solutions for dex/collaborative Ego.
- E8 specification: ten-channel sensor density + TOF + motion capture (R&D-phase target, subject to written confirmation). An R&D target suitable for high-density sensing and VIO comparison experiments; commercial confirmation of mass-production boundaries is required.
- Wrist: PPS + UART, in the same domain as the head-mounted unit, serving contact blind spots for ActiveUMI-type use cases.
- Light ring / finger ring: supplementary tool-side geometry for HaMeR / MANO-friendly annotation.

Engineering deep dive: turning paper metrics into gates by scenario
(Deep Dive · 31-1) For EgoLive and matrix: laboratory matrix-style capture must define three-level primary keys of operator_id, session_id, and shard_id to avoid filename collisions. Under the compare pillar, this directly relates to the delivery risk of "Matrix Selection: E under the EgoLive Cost Curve …".
(Deep Dive 31-2) Regarding EgoLive and matrix: when comparing against public benchmarks, only borrow task definitions and metrics, not data license boundaries. Under the compare pillar, this directly relates to the delivery risk of "Matrix Selection: Under the EgoLive Cost Curve E...".
(Deep Dive · 31-3) For EgoLive and matrix: Ego-to-robot transfer should have a separate coordinate system appendix: the head, wrist, world, robot_base transformation chain. Under the compare pillar, this directly relates to the delivery risk of "Matrix Selection: EgoLive Cost Curve E…".
(Deep dive 31-4) For EgoLive and matrix: if HEVC transcoding rewrites timestamps, the original_pts field must be preserved in metadata. Under the compare pillar, this directly relates to the delivery risk of "matrix selection: EgoLive cost curve under E...".
(Deep dive 31-5) For EgoLive and matrix: ring-based supplemental points are suitable for sparse geometric constraints and should not replace high-resolution hand pixel supervision. Under the compare pillar, this directly relates to the delivery risk of "Matrix selection: under the EgoLive cost curve E…".
(Deep dive · 31-6) For EgoLive and matrix: if the device has on-device SLAM (such as E6), the trajectory can serve as a prior and should be disableable on the training side; E2 defaults to offline reconstruction, and offline reconstruction (not on-device SLAM) must not be written down as an E2 factory capability. Under the compare pillar, this directly relates to the delivery risk of "Matrix selection: EgoLive cost curve E…".
(Deep dive · 31-7) Regarding EgoLive and matrix: Collaborative Ego scenarios require declaring the speech/instruction synchronization strategy to avoid audio-visual causal misalignment. Under the compare pillar, this directly relates to the delivery risk of "matrix selection: EgoLive cost curve under E…".
(Deep dive · 31-8) Regarding EgoLive and matrix: QA spot checks should re-run the same script on a Linux host to ensure auditable delivery. Under the compare pillar, this directly relates to the delivery risk of "Matrix selection: EgoLive cost curve under E…".
(Extension · 31-B1) 2026-09-13 Retrospective: There may be a gap between the E8 ten-channel and TOF channels in R&D integration versus mass-production specifications; the contract should clearly state the delivery tier. If your team tags=matrix,E2,EgoLive,compare, it is recommended to add this to the project risk register.
(Extension · 31-B2) 2026-09-13 Retrospective: Motion-capture bridging is a Stretch capability and requires a separate milestone and acceptance script. If your team tags=matrix,E2,EgoLive,comparison, it is recommended to record this in the project risk register.
(Extension · 31-B3) 2026-09-13 retrospective: Global shutter reduces geometric distortion compared to rolling shutter in fast lateral motion scenarios, benefiting stereo matching. If your team tags=matrix, E2, EgoLive, comparison, it is recommended to record this in the project risk register.
(Extension · 31-B4) 2026-09-13 Retrospective: The choice of a 65mm-class baseline affects depth accuracy and wearing volume, and should be consistent across the product page and data card. If your team tags=matrix, E2, EgoLive, comparison, it is recommended to record this in the project risk register.
(Extension · 31-B5) 2026-09-13 retrospective: fMP4 segmentation is convenient for browser preview, but the training side requires a raw or lossless backup strategy. If your team tags = matrix, E2, EgoLive, comparison, we recommend recording this in the project risk register.
Frequently Asked Questions (FAQ)
Q: Does TOBI provide the EgoLive dataset? Not provided. EgoLive is a public research/community resource; TOBI only discusses specification comparisons and capture hardware capabilities.
Q: Can using only matrix reproduce all of the paper's results? No, it cannot be automatically guaranteed. The papers also involve task design, annotation toolchains, and training code; the hardware is responsible for making the timeline and sensor fields verifiable and acceptable.
Q: What does the demo animation demo-eval.webp represent? Animated schematic demonstrating matrix real-device recording and wear configuration; not official EgoLive release material.
Q: What attachments should I prepare when requesting a quote? It is recommended to attach https://arxiv.org/abs/2604.23570, a target hours/scenario table, a Must-field checklist, and whether wrist/ring/matrix networking is required.
Next step: specification comparison and evaluation prototype
If you are planning matrix,E2,EgoLive,comparison scenarios against EgoLive, please contact TOBI with your Must-field checklist to apply for a matrix evaluation unit or a joint dataset delivery plan.
Read matrix product page, Developer Documentation · Device SDK, or Submit Quote Request and Technical Evaluation.
(Appendix · news-seo-031-e2-vs-egolive-cost · 1) Inconsistent versions of the wrist-mounted and head-mounted extrinsic calibration cause a systematic shift in MANO supervision. Run regression tests when the wearer changes. When assessing matrix against EgoLive (2026), archive this together with the acceptance checklist dated 2026-09-13.
(Appendix · news-seo-031-e2-vs-egolive-cost · 2) Semantic gestures can serve as weak supervision, rather than independent labels, only when the OpenXR gesture stream and video frames share the same time base. When assessing matrix against EgoLive (2026), archive this together with the acceptance checklist dated 2026-09-13.
(Appendix · news-seo-031-e2-vs-egolive-cost · 3) DatasetServer shard limits and disk flush strategy affect whether the "last frame" is lost. When assessing matrix against EgoLive (2026), archive this together with the acceptance checklist dated 2026-09-13.
(Appendix · news-seo-031-e2-vs-egolive-cost · 4) Matrix-style data capture in the lab should define a three-level primary key of operator_id, session_id, and shard_id to avoid filename collisions. When assessing matrix against EgoLive (2026), archive this together with the acceptance checklist dated 2026-09-13.
(Appendix · news-seo-031-e2-vs-egolive-cost · 5) When comparing against public benchmarks, only borrow the task definitions and metrics, not the data licensing boundaries. When assessing matrix against EgoLive (2026), archive this together with the acceptance checklist dated 2026-09-13.
(Appendix · news-seo-031-e2-vs-egolive-cost · 6) Ego-to-robot transfer should include a separate coordinate frame attachment: the transformation chain for head, wrist, world, and robot_base. When assessing matrix against EgoLive (2026), archive this together with the acceptance checklist dated 2026-09-13.
(Appendix · news-seo-031-e2-vs-egolive-cost · 7) If HEVC transcoding rewrites timestamps, the original_pts field must be preserved in the metadata. When assessing matrix against EgoLive (2026), archive this together with the acceptance checklist dated 2026-09-13.
(Appendix · news-seo-031-e2-vs-egolive-cost · 8) Ring-based point completion is suitable for sparse geometric constraints and should not replace high-resolution hand pixel supervision. When assessing matrix against EgoLive (2026), archive this together with the acceptance checklist dated 2026-09-13.
(Appendix · news-seo-031-e2-vs-egolive-cost · 9) If the device has on-board SLAM (e.g., E6), the trajectory can serve as a prior and must be disableable on the training side; E2 defaults to offline reconstruction, and offline reconstruction (not on-board SLAM) must not be written as an E2 factory capability. When assessing matrix against EgoLive (2026), archive this together with the acceptance checklist dated 2026-09-13.
(Appendix · news-seo-031-e2-vs-egolive-cost · 10) Collaborative Ego scenarios must declare the speech/instruction synchronization strategy to avoid audio-visual causal misalignment. When assessing matrix against EgoLive (2026), archive this together with the acceptance checklist dated 2026-09-13.
(Appendix · news-seo-031-e2-vs-egolive-cost · 11) QA spot checks should re-run the same script on a Linux host to ensure the delivery is auditable. When assessing matrix against EgoLive (2026), archive this together with the acceptance checklist dated 2026-09-13.
(Appendix · news-seo-031-e2-vs-egolive-cost · 12) Illumination variation in kitchen and home long-tail scenarios requires exposure strategy to be written into the SOP: fixed, manual, and adaptive modes must be consistent. When assessing matrix against EgoLive (2026), archive this together with the acceptance checklist dated 2026-09-13.
(Appendix · news-seo-031-e2-vs-egolive-cost · 13) When 60fps stereo capture is combined with a 30fps auxiliary stream, clearly specify the downsampling and phase rules. When assessing matrix against EgoLive (2026), archive this together with the acceptance checklist dated 2026-09-13.
(Appendix · news-seo-031-e2-vs-egolive-cost · 14) When aligning 200Hz IMU with video frames, it must be declared whether interpolation or nearest-neighbor is used, along with boundary frame handling. When assessing matrix against EgoLive (2026), archive this together with the acceptance checklist dated 2026-09-13.
(Appendix · news-seo-031-e2-vs-egolive-cost · 15) ORB-SLAM3-class VIO baselines emphasize multi-sensor timestamps; Ego devices should align to equivalent semantics. When assessing matrix against EgoLive (2026), archive this together with the acceptance checklist dated 2026-09-13.
(Appendix · news-seo-031-e2-vs-egolive-cost · 16) Open X-Embodiment emphasizes cross-platform action distributions; the Ego field schema should reserve an embodiment_id field. When assessing matrix against EgoLive (2026), archive this together with the acceptance checklist dated 2026-09-13.
(Appendix · news-seo-031-e2-vs-egolive-cost · 17) DROID-class robot data lacks human priors; E2's lightweight Ego is suitable as a pre-stage for human demonstration. When assessing matrix against EgoLive (2026), archive this together with the acceptance checklist dated 2026-09-13.
(Appendix · news-seo-031-e2-vs-egolive-cost · 18) The HaMeR lineage is sensitive to occlusion; ring + multi-path redundancy reduces the proportion of invisible frames. When assessing matrix against EgoLive (2026), archive this together with the acceptance checklist dated 2026-09-13.
(Appendix · news-seo-031-e2-vs-egolive-cost · 19) ActiveUMI demonstrates active perception; dual-channel acceptance of wrist-mounted plus head-mounted streams is the procurement-grade answer. When assessing matrix against EgoLive (2026), archive this together with the acceptance checklist dated 2026-09-13.
(Appendix · news-seo-031-e2-vs-egolive-cost · 20) World model training relies more on long-duration consistent timelines; random cropping of short clips destroys state continuity. When assessing matrix against EgoLive (2026), archive this together with the acceptance checklist dated 2026-09-13.
(Appendix · news-seo-031-e2-vs-egolive-cost · 21) Diffusion policies are sensitive to contact phase; the data card should annotate the start and end frames of contact_phase. When assessing matrix against EgoLive (2026), archive this together with the acceptance checklist dated 2026-09-13.
(Appendix · news-seo-031-e2-vs-egolive-cost · 22) The Sim2Real pipeline should not use different timestamp units in sim and real. When assessing matrix against EgoLive (2026), archive this together with the acceptance checklist dated 2026-09-13.
(Appendix · news-seo-031-e2-vs-egolive-cost · 23) Enterprise content funnel: paper page → spec page → prototype → small-batch data → scaled procurement, with exit criteria at each step. When assessing matrix against EgoLive (2026), archive this together with the acceptance checklist dated 2026-09-13.
(Appendix · news-seo-031-e2-vs-egolive-cost · 24) This article focuses on engineering acceptance field descriptions, so that procurement and algorithm teams can copy them directly into RFQ attachments. When assessing matrix against EgoLive (2026), archive this together with the acceptance checklist dated 2026-09-13.
(Appendix · news-seo-031-e2-vs-egolive-cost · 25) When linking to SDK documentation, fix the anchor to #device-sdk to reduce reader disorientation. When assessing matrix against EgoLive (2026), archive this together with the acceptance checklist dated 2026-09-13.
(Appendix · news-seo-031-e2-vs-egolive-cost · 26) The video poster matches the hero image, reducing first-screen CLS and trust cost. When assessing matrix against EgoLive (2026), archive this together with the acceptance checklist dated 2026-09-13.
(Appendix · news-seo-031-e2-vs-egolive-cost · 27) The FAQ must reiterate dataset boundaries to prevent sales from over-promising. When assessing matrix against EgoLive (2026), archive this together with the acceptance checklist dated 2026-09-13.
(Appendix · news-seo-031-e2-vs-egolive-cost · 28) There may be a gap between the E8 ten-channel and TOF channels in R&D integration versus mass-production specifications; the contract should specify the delivery level clearly. When assessing matrix against EgoLive (2026), archive this together with the acceptance checklist dated 2026-09-13.
(Appendix · news-seo-031-e2-vs-egolive-cost · 29) Motion-capture bridging is a Stretch capability and requires a separate milestone and acceptance script. When assessing matrix against EgoLive (2026), archive this together with the acceptance checklist dated 2026-09-13.
(Appendix · news-seo-031-e2-vs-egolive-cost · 30) Global shutter reduces geometric distortion in fast lateral motion scenarios compared to rolling shutter, benefiting stereo matching. When assessing matrix against EgoLive (2026), archive this together with the acceptance checklist dated 2026-09-13.
(Appendix · news-seo-031-e2-vs-egolive-cost · 31) The choice of 65mm-level baseline affects depth accuracy and wearing volume, and should be consistent across the product page and data card. When assessing matrix against EgoLive (2026), archive this together with the acceptance checklist dated 2026-09-13.
(Appendix · news-seo-031-e2-vs-egolive-cost · 32) fMP4 segments facilitate browser preview, but the training side requires a raw or lossless backup strategy. When assessing matrix against EgoLive (2026), archive this together with the acceptance checklist dated 2026-09-13.
(Appendix · news-seo-031-e2-vs-egolive-cost · 33) MCAP is well-suited to the ROS2 ecosystem; field naming should follow the team's schema registry. When assessing matrix against EgoLive (2026), archive this together with the acceptance checklist dated 2026-09-13.
(Appendix · news-seo-031-e2-vs-egolive-cost · 34) Scripts like linux_run_time_sync.sh should be versioned and cross-referenced with the firmware README. When assessing matrix against EgoLive (2026), archive this together with the acceptance checklist dated 2026-09-13.
(Appendix · news-seo-031-e2-vs-egolive-cost · 35) The open-source e2-sdk/e6-sdk only helps you get started; mass-production procurement still follows the commercial delivery package. When assessing matrix against EgoLive (2026), archive this together with the acceptance checklist dated 2026-09-13.
(Appendix · news-seo-031-e2-vs-egolive-cost · 36) When multiple capture operators work in parallel, mixing NTP and hardware time sources will prevent shards from being merged; a unified time source must be used. When assessing matrix against EgoLive (2026), archive this together with the acceptance checklist dated 2026-09-13.
(Appendix · news-seo-031-e2-vs-egolive-cost · 37) Health monitoring should include camera dropout, IMU saturation, remaining storage, and thermal throttling. When assessing matrix against EgoLive (2026), archive this together with the acceptance checklist dated 2026-09-13.
(Appendix · news-seo-031-e2-vs-egolive-cost · 38) Privacy: faces, house numbers, and screens require a redaction pipeline, and redaction logs must be auditable. When assessing matrix against EgoLive (2026), archive this together with the acceptance checklist dated 2026-09-13.
(Appendix · news-seo-031-e2-vs-egolive-cost · 39) Version ledger: firmware, calibration, APP, host scripts, DatasetServer — five version number categories bound to the session. When assessing matrix against EgoLive (2026), archive this together with the acceptance checklist dated 2026-09-13.
(Appendix · news-seo-031-e2-vs-egolive-cost · 40) The annotation toolchain is decoupled from the hardware; but exports must reference calib_version and sync_report. When assessing matrix against EgoLive (2026), archive this together with the acceptance checklist dated 2026-09-13.
(Appendix · news-seo-031-e2-vs-egolive-cost · 41) A VLA recipe should declare the language instruction alignment method: manual trigger vs. ASR timestamps. When assessing matrix against EgoLive (2026), archive this together with the acceptance checklist dated 2026-09-13.
(Appendix · news-seo-031-e2-vs-egolive-cost · 42) Human-to-robot transfer requires declaring the action space: joint vs. end-effector vs. bimanual. When assessing matrix against EgoLive (2026), archive this together with the acceptance checklist dated 2026-09-13.
(Appendix · news-seo-031-e2-vs-egolive-cost · 43) It is recommended that procurement contract attachments include links to public papers for task reference only, without implying data transfer. When assessing matrix against EgoLive (2026), archive this together with the acceptance checklist dated 2026-09-13.
(Appendix · news-seo-031-e2-vs-egolive-cost · 44) The algorithm team should document the observation–action causal window in the data card, including whether frames from multiple streams are retained for 120ms before contact and 200ms after contact. When assessing matrix against EgoLive (2026), archive this together with the acceptance checklist dated 2026-09-13.
(Appendix · news-seo-031-e2-vs-egolive-cost · 45) During procurement review, require the supplier to provide P50/P95 curves of sync jitter, not single-point marketing figures. When assessing matrix against EgoLive (2026), archive this together with the acceptance checklist dated 2026-09-13.
(Appendix · news-seo-031-e2-vs-egolive-cost · 46) In long-duration dex capture, thermal drift and bandwidth contention amplify soft alignment errors; hardware triggering is superior to pure software timestamping. When assessing matrix against EgoLive (2026), archive this together with the acceptance checklist dated 2026-09-13.
(Appendix · news-seo-031-e2-vs-egolive-cost · 47) Failed samples must include a timeline health score; otherwise, VLA will treat interpolated frames after repair as real contact. When assessing matrix against EgoLive (2026), archive this together with the acceptance checklist dated 2026-09-13.
(Appendix · news-seo-031-e2-vs-egolive-cost · 48) Cross-day capture requires declaring the time source switching strategy: whether CDC-ACM, PPS, or GNSS are mixed. When assessing matrix against EgoLive (2026), archive this together with the acceptance checklist dated 2026-09-13.
(Appendix · news-seo-031-e2-vs-egolive-cost · 49) Inconsistent versions of the wrist-mounted and head-mounted extrinsic calibration cause a systematic shift in MANO supervision. Run regression tests when the wearer changes. When assessing matrix against EgoLive (2026), archive this together with the acceptance checklist dated 2026-09-13.
(Appendix · news-seo-031-e2-vs-egolive-cost · 50) Semantic gestures can serve as weak supervision, rather than independent labels, only when the OpenXR gesture stream and video frames share the same time base. When assessing matrix against EgoLive (2026), archive this together with the acceptance checklist dated 2026-09-13.
(Appendix · news-seo-031-e2-vs-egolive-cost · 51) DatasetServer shard limits and disk flush strategy affect whether the "last frame" is lost. When assessing matrix against EgoLive (2026), archive this together with the acceptance checklist dated 2026-09-13.
(Appendix · news-seo-031-e2-vs-egolive-cost · 52) Matrix-style data capture in the lab should define a three-level primary key of operator_id, session_id, and shard_id to avoid filename collisions. When assessing matrix against EgoLive (2026), archive this together with the acceptance checklist dated 2026-09-13.