docs(perception): update RAVNOVES00 maturation roadmap
This commit is contained in:
DCCONSTRUCTIONS
@@ -3,6 +3,52 @@
|
||||
This plan supersedes the app-dependent experiment order in the reference Bible.
|
||||
Each gate produces evidence and an explicit GO, PAUSE or BLOCKED result.
|
||||
|
||||
## L3.4F candidate-visible adjudication checkpoint — 2026-08-03
|
||||
|
||||
- GO for engineering reference: the L3.4E mismatch gallery now opens a
|
||||
fullscreen candidate-visible adjudication workspace. Candidate boxes are a
|
||||
yellow read-only layer; human boxes remain independently movable, resizable,
|
||||
deletable, addable and relabelable with model scores hidden.
|
||||
- The complete revision
|
||||
`l34f-adjudication-session-cd8b2d7a1f5a5fdc4dd54fcd82a03c86c326c575e9ef3670863bd66df350a74b`
|
||||
covers 32/32 exact RIGHT-camera frames and 260 human objects. A full visual
|
||||
pass found no honest correction to make: all 260 source objects remained
|
||||
unchanged, with zero added, deleted, geometry-changed or class-changed
|
||||
objects.
|
||||
- Immutable result
|
||||
`l34f-adjudicated-reference-fe2964ea46fab02fd750ef30302a2b12e4c822a2407ba1137022a3ae26e15e5f`
|
||||
is published as the dated LAB L3.4F variant. Its UI contains the 32-frame
|
||||
OVERLAY/CANDIDATE/HUMAN/SOURCE gallery rather than metrics-only evidence.
|
||||
- Browser acceptance is preserved at
|
||||
`.runtime/visual-qa/l34f-adjudication/l34f-lab-result.png`
|
||||
(`sha256:b9fa6ee7a304d8cfefbceb976e61f6ba459e7d9783405bd90d54403fce74a6ce`)
|
||||
and `.runtime/visual-qa/l34f-adjudication/l34f-visual-evidence.png`
|
||||
(`sha256:96af84926be6b488cc359eb772062697c0ab3e00b2fc4c95ca9a5d34c3dda6d0`).
|
||||
- PAUSE for detector acceptance: candidate visibility and reviewer reuse make
|
||||
L3.4F an engineering reference, not independent or metric-grade truth. It
|
||||
authorizes neither retuning nor L3.5. E48 still requires two distinct real
|
||||
prediction-free reviewer identities plus explicit adjudication.
|
||||
|
||||
## L3.4D prediction-hidden labeling checkpoint — 2026-08-03
|
||||
|
||||
- GO for collection UI: the contextual `Разметить данные` action now opens the
|
||||
existing fullscreen editor from the exact frozen L3.4D cumulative candidate.
|
||||
- The L3.4D source contract returns 32 hash-bound RIGHT-camera frames with
|
||||
`candidate_predictions_included=false`, `prelabels_included=false` and no
|
||||
annotation-seed endpoint. Candidate-seeded objects are rejected on save.
|
||||
- Blank reviewer session
|
||||
`l34-annotation-session-21b706576366796e534e57e5c10ccce33f2966e9343a1709757e2bf52c05a01a`
|
||||
is created under the separate `l34d-blind` namespace and is currently
|
||||
`0/32`, revision `0`.
|
||||
- Visual QA is fixed in
|
||||
`.runtime/visual-qa/l34d-blind/2026-08-03_l34d_prediction-hidden-empty-session.jpg`
|
||||
(`sha256:111ebd240ec7ccefeb569b42c7e937f1a5e0a7c5e72273f41f641df5984af030`):
|
||||
the clean frame has zero boxes, the blind status is visible and the close
|
||||
action remains in bounds at 1280×720.
|
||||
- PAUSE for evaluation: this empty draft is not truth and does not open L3.5.
|
||||
One complete prediction-hidden review, independent verification and
|
||||
adjudication are still required before a single reveal/evaluation pass.
|
||||
|
||||
## Architecture stabilization checkpoint — 2026-07-28
|
||||
|
||||
| Gate | Current result |
|
||||
@@ -17,7 +63,8 @@ Each gate produces evidence and an explicit GO, PAUSE or BLOCKED result.
|
||||
| E31 binding sensitivity | MEASURED — E45 closes accounting for 87/87 accepted correspondences and finds no material monotonic residual association with represented image radius, rig speed or pose age. It does not supply calibration-target truth or outer-fisheye coverage. |
|
||||
| Detector Truth Island | PREPARED FOR HUMAN REVIEW — E46 freezes 32 references with no prelabels, predictions, scores or candidate identity. Immutable handoff `e46-review-handoff-652f7409c8a529946a26264cea006f3afdd714e19529c731e0f66d0f03fd5668` provides two separately assigned, hash-verified CVAT/COCO reviewer packages and a fail-closed E48 converter. Two real independent reviews and adjudication are still required; the handoff is not ground truth. |
|
||||
| Detector candidate comparison | FROZEN BEFORE TRUTH — E47 freezes raw-KB4 and fixed-valid-FOV-fill predictions from the same exact Mask R-CNN checkpoint. No accuracy result or winner exists before the E46 truth seal. |
|
||||
| Detector truth/evaluation executors | DEFERRED BY OWNER, READY, NOT RUN — E48 fail-closed review/adjudication sealing and separate post-seal E49 scoring are implemented. Neither result exists because real independent reviews are absent. The deferral does not convert E37–E40 into blind truth and does not block R2/R4 operational work. |
|
||||
| RIGHT YOLOX Truth Island freeze | FROZEN BEFORE TRUTH — L3.4 freezes the exact current YOLOX-S generation on all 32 E46 right-camera frames as `l34-right-yolox-truth-island-freeze-5175a03144978b25130019da6d37bceb8c6ed6aa3d0d3a4d2df4483e1e27ae76`. Every row binds candidate, source-image SHA-256, sequence, frame and source time with `truth_joined=false`. The 268 boxes are descriptive only; no accuracy or acceptance claim exists. |
|
||||
| Detector truth/evaluation executors | READY, MANUAL TRUTH NOT RUN — E48 fail-closed review/adjudication sealing, E49 comparison scoring and L3.5 post-seal evaluation of the exact L3.4 generation are implemented. No E48/E49/L3.5 result exists because two real independent reviews and adjudication are absent. This does not convert E37–E40 or L3.4 counts into blind truth. |
|
||||
| RAVNOVES00 recorded AI publication | GO — after a real backend restart, the admitted E10 overlay `e10-integrated-perception-36964643f1c0727a434671e3cddc2e76c536b9c28b8c56845f18941ae791c39a` recovered from its sealed cache without model execution or source revalidation. Recording `f83e29dc-5d25-42a5-824e-228cdbb83d21` streamed `69,227,327` bytes with verified SHA-256 `c2bb73fca8bb616a3929df236c6dc63d071b6466a6ca57d06487ef18ed985b57`; a warm localhost request completed in `0.391 s`, and its steady-state server RSS delta was `48 KiB`. |
|
||||
| L2.6 lossless replay parity | GO — replay-pack-v2 exact equivalence passed for all 4,570 native LiDAR frames, 4,598 pose frames and 10,751,258 points. The same local-surface algorithm produced 4,570/4,570 valid frames with 20.6421 ms pose-age p95 while retaining separate v2 identity, RGBI/intensity and host-time provenance. |
|
||||
| L2.6 motion/semantic qualification | GO (diagnostic only) — E51 processed 4,489/4,489 E32/E34 frames, preserved all 2,119,302 current occupied point rows, measured 0.4148 ms frame p95 and 0.2969 MiB peak-RSS growth, and emitted explicit freshness/conflict/proximity/motion evidence. Dynamic class, collision state, free space, command, navigation and safety authority remain unavailable. |
|
||||
|
||||
@@ -65,6 +65,20 @@ false-free claims. R5 is a later operator-authorized capture and transfer
|
||||
cycle. The detailed gate is
|
||||
`docs/20_RAVNOVES00_REFERENCE_SOURCE_PRODUCT_PLAN.md`.
|
||||
|
||||
### Detector checkpoint · 2026-08-01
|
||||
|
||||
The current physical scope is one recorded RAVNOVES00 source and only
|
||||
`sensor.camera.right`; left-camera duplication, live transport and new hardware
|
||||
are excluded. L3.4 freezes the admitted YOLOX-S generation on the exact 32 E46
|
||||
Truth Island frames before any independent labels are available. L3.5 is the
|
||||
prepared downstream gate: it can join only an accepted E48 seal that belongs to
|
||||
the same E46 island and was sealed after this exact L3.4 freeze. The shared E49
|
||||
metric engine remains unable to select a winner or authorize retraining.
|
||||
|
||||
This moves the detector branch to the last honest technical boundary. Actual
|
||||
quality scoring still waits on two independent human reviews and adjudication;
|
||||
no synthetic or model-assisted substitute is admitted.
|
||||
|
||||
## Workstream B — Simulation Polygon
|
||||
|
||||
This workstream may proceed in parallel but cannot bypass Workstream A or grant
|
||||
|
||||
@@ -68,11 +68,12 @@ failures remain blocking even when an aggregate percentage passes.
|
||||
| R4 — reference release candidate | Replay the complete RAVNOVES00 source through the accepted pipeline at recorded pace with regression, resource and recovery evidence | Source-scoped release decision for the known-location product envelope |
|
||||
| R5 — later transfer | Under the preregistered E43 protocol, record the same K1 and mount over a new route after explicit operator authorization; freeze predictions before independent labels are revealed | Generalization decision; no retuning before comparison |
|
||||
|
||||
R2 and non-UI R4 preparation are the active stabilization path. The owner has
|
||||
deferred the two-reviewer E46–E49 blind-acceptance branch without cancelling or
|
||||
weakening it; no independent accuracy claim exists while it is deferred. R3 is
|
||||
explicitly deferred for separate discussion. R5 has a frozen protocol but no
|
||||
capture or labels yet.
|
||||
R2 and non-UI R4 preparation remain active. On 2026-08-01 the owner resumed the
|
||||
detector branch through the last machine-executable boundary: freeze the current
|
||||
right-camera YOLOX candidate and prepare its post-seal evaluator. The two real
|
||||
reviewers and adjudication are still a manual prerequisite, so no independent
|
||||
accuracy claim exists yet. R3 remains a separate product discussion. R5 has a
|
||||
frozen protocol but no capture or labels yet.
|
||||
|
||||
## Current evidence
|
||||
|
||||
@@ -197,6 +198,21 @@ seal and computes the frozen AP/AR, miss, large-box, valid-FOV-centre leakage
|
||||
and temporal class-count metrics. It cannot select a candidate automatically.
|
||||
No E48/E49 result exists until two real reviews and adjudication are supplied.
|
||||
|
||||
L3.4 immutable current-candidate freeze
|
||||
`l34-right-yolox-truth-island-freeze-5175a03144978b25130019da6d37bceb8c6ed6aa3d0d3a4d2df4483e1e27ae76`
|
||||
adds the actual admitted right-camera YOLOX-S generation to the same exact E46
|
||||
island. It freezes 268 predictions over 32/32 frames and 20 temporal/anchor
|
||||
groups. Every prediction row binds `candidate_id`, E46 sequence, image and
|
||||
frame identity, source time and source-image SHA-256, and explicitly retains
|
||||
`truth_joined=false`. The counts do not measure correctness.
|
||||
|
||||
The L3.5 executor is prepared but cannot run before an accepted E48 seal. It
|
||||
fails closed unless E46, E48 and L3.4 reference the same island and the L3.4
|
||||
freeze predates the independent truth seal. It converts no boxes or labels: it
|
||||
adapts the exact frozen rows to the existing E49 AP/AR, miss, false-large-box,
|
||||
valid-FOV leakage and temporal-flicker engine. It cannot select a candidate or
|
||||
authorize retraining. No L3.5 result exists while human truth is absent.
|
||||
|
||||
### Recorded AI publication checkpoint · 2026-07-30
|
||||
|
||||
The existing accepted E10 generation is now usable as a stable recorded product
|
||||
@@ -269,8 +285,9 @@ The current non-UI priority order is:
|
||||
producers to adopt verified resolution before any separately authorized
|
||||
physical reclamation;
|
||||
5. preserve E43 unchanged until the owner supplies the new capture.
|
||||
6. keep the E46 handoff, E47 prediction freeze and E48/E49 executors immutable
|
||||
while the owner-requested review deferral is active;
|
||||
6. preserve E46/E47, the exact L3.4 current-candidate freeze and the E48/E49/L3.5
|
||||
fail-closed lineage while two independent reviews and adjudication are
|
||||
obtained;
|
||||
7. continue R2/R4 deterministic replay, recovery and resource acceptance
|
||||
without treating operational readiness as independent detector accuracy.
|
||||
|
||||
@@ -281,8 +298,8 @@ The following remain outside the present stabilization increment:
|
||||
- initiation of a new physical route/camera capture by an agent;
|
||||
- second K1 or changed mount generalization;
|
||||
- open-world and adversarial evaluation.
|
||||
- two independent E46 reviews, adjudication and E49 candidate scoring until the
|
||||
owner resumes that branch.
|
||||
- independent detector scoring until two real E46 reviews and adjudication are
|
||||
supplied; the code path is ready, but the truth cannot be synthesized.
|
||||
|
||||
The owner may provide a future recording under E43. Until it exists, current
|
||||
work neither depends on it nor simulates its result. The existing E36 audit and
|
||||
|
||||
@@ -0,0 +1,97 @@
|
||||
# LAB L3.4 · RIGHT YOLOX Truth Island freeze
|
||||
|
||||
Date: 2026-08-01
|
||||
|
||||
Status: predictions frozen; awaiting independent E46 truth
|
||||
|
||||
Immutable result:
|
||||
`l34-right-yolox-truth-island-freeze-5175a03144978b25130019da6d37bceb8c6ed6aa3d0d3a4d2df4483e1e27ae76`
|
||||
|
||||
## Objective and architecture stage
|
||||
|
||||
Freeze the actual admitted YOLOX-S candidate on the exact existing E46
|
||||
RAVNOVES00 Truth Island before any independent labels are revealed. This is the
|
||||
current `sensor.camera.right` recorded-replay branch. Left camera, live K1
|
||||
transport, new hardware and a second route are explicitly out of scope.
|
||||
|
||||
The historical `LAB E50` remains the accepted exact-content reference index.
|
||||
L3.4 uses a distinct identity and storage root; it does not overwrite or
|
||||
reinterpret E50.
|
||||
|
||||
## Immutable source evidence
|
||||
|
||||
- E46 Truth Island:
|
||||
`e46-detector-truth-island-d8ab2745679636dce374b720b562fce05d6d0a26be3eac88650224d7aa92267d`;
|
||||
- source session: `20260720T065719Z_viewer_live`;
|
||||
- source: `sensor.camera.right`, recorded replay only;
|
||||
- 32 preselected model-independent frames across 20 anchor/temporal groups;
|
||||
- full rectified YOLOX qualification: 4,489/4,489 frames;
|
||||
- admitted L3.3 generation:
|
||||
`l33-camera-first-detector-review-2458f6214a731c06bfb44c6e44d4a32ee10389359debdc412c321220a9cdc010`.
|
||||
|
||||
## Method, model and binding
|
||||
|
||||
The candidate is `YOLOX-S`, identity `yolox-s-kb4-core3`, with the exact model
|
||||
SHA-256 `c5c2d13e59ae883e6af3b45daea64af4833a4951c92d116ec270d9ddbe998063`
|
||||
and frozen minimum score `0.25`. COCO detector classes are mapped only into the
|
||||
seven-class E46 task ontology; `truck` and `bus` map to `heavy_vehicle`.
|
||||
|
||||
Every persisted row binds the candidate ID, E46 sequence, image ID, frame
|
||||
index, source time and source-image SHA-256. Each row states
|
||||
`truth_joined=false`. Prediction rows are content-addressed in the immutable
|
||||
manifest, so a changed box, score, frame or image produces another generation.
|
||||
|
||||
## Result
|
||||
|
||||
| Measurement | Value |
|
||||
| --- | ---: |
|
||||
| Frozen frames | 32 / 32 |
|
||||
| Anchor/temporal groups | 20 |
|
||||
| Frames with predictions | 32 |
|
||||
| Predictions | 268 |
|
||||
| Car | 234 |
|
||||
| Heavy vehicle | 22 |
|
||||
| Person | 7 |
|
||||
| Motorcycle | 3 |
|
||||
| Bicycle | 2 |
|
||||
|
||||
These are descriptive counts, not quality metrics. AP, recall, misses and false
|
||||
positives are unavailable until independent truth is sealed.
|
||||
|
||||
## Visual evidence contract
|
||||
|
||||
The LAB result exposes every exact E46 camera frame together with only the
|
||||
frozen L3.4 boxes. `BOXES` and `SOURCE` modes, frame navigation, search,
|
||||
expand/restore and Escape are provided through the reusable Mission Core LAB
|
||||
viewer. The frame API revalidates the E46 manifest, frame/image/group/time
|
||||
binding, source-image digest, PNG bytes and prediction generation before
|
||||
rendering.
|
||||
|
||||
Browser acceptance passed on the canonical `127.0.0.1:8000` service for normal
|
||||
`BOXES`, normal `SOURCE`, expanded `BOXES`, frame navigation and Escape restore.
|
||||
The accepted visual manifest is retained outside Git at
|
||||
`.runtime/compute-experiments/l3/right-yolox-truth-island-freeze/visual-acceptance/manifest.json`
|
||||
with SHA-256
|
||||
`92641f0c3f76f72e12fc8d49ec2d951089771efb7e1dfe7fb94d2523148154f9`.
|
||||
Its primary frame-1249 BOXES screenshot has SHA-256
|
||||
`7858b16774e34ee9775238736d0379c146250e8cb7f2efe79e61cdeeb8343a6d`;
|
||||
the corresponding SOURCE and expanded BOXES screenshots are independently
|
||||
hashed in that manifest.
|
||||
|
||||
## Downstream gate
|
||||
|
||||
L3.5 is implemented but intentionally has no result. It accepts only an
|
||||
adjudicated E48 truth seal for the same E46 island and rejects any L3.4 freeze
|
||||
created after the truth was sealed. It then reuses the E49 metric engine without
|
||||
changing labels, boxes or source identity. Candidate selection and retraining
|
||||
remain explicit later decisions.
|
||||
|
||||
## Decision and limitations
|
||||
|
||||
The current right-camera YOLOX prediction generation is reproducibly frozen and
|
||||
visually inspectable. It is eligible for future blind scoring, but is neither
|
||||
accepted nor rejected as a detector.
|
||||
|
||||
Two distinct human reviewers and adjudication are the next non-synthetic gate.
|
||||
No live, hardware, cross-camera, cross-route, navigation, safety or command
|
||||
claim is made.
|
||||
Reference in New Issue
Block a user