feat(lab): publish E37 R0 acceptance result
This commit is contained in:
@@ -0,0 +1,134 @@
|
||||
# LAB E37 — RAVNOVES00 acceptance contract R0
|
||||
|
||||
## Result identity
|
||||
|
||||
Status: `accepted-r0-source-scoped-contract`.
|
||||
|
||||
Result:
|
||||
`e37-ravnoves-acceptance-01b1efd586f747341c712d82f0907b39436a6f91ae92b1dfae987eca05fd8344`.
|
||||
|
||||
Worker package:
|
||||
`e37-worker-package-960b804df639dd2937895377aa67d5dca4b325f293ab48868d16049da2ff28bc`.
|
||||
|
||||
Physical execution node: `DESKTOP-OPJ8J04` / Worker 006.
|
||||
|
||||
## Question
|
||||
|
||||
E37 does not ask whether K1 has already reached 90% quality. It asks which
|
||||
immutable RAVNOVES00 cases, task ontology, split and metrics will be used to
|
||||
make that claim honestly in the next gate.
|
||||
|
||||
The prior E30 evidence review supplied a useful engineering-reviewed
|
||||
substrate, but it was not yet an acceptance contract. Without freezing the
|
||||
denominator and validation holdout, later tuning could silently move the
|
||||
evaluation set or collapse presence, point ownership and freshness into one
|
||||
ambiguous score.
|
||||
|
||||
## Frozen source and provenance
|
||||
|
||||
Source recording:
|
||||
|
||||
- display name: `RAVNOVES00`;
|
||||
- session: `20260720T065719Z_viewer_live`;
|
||||
- classification: immutable private physical recording.
|
||||
|
||||
Inputs:
|
||||
|
||||
- E30 materialization:
|
||||
`e30-materialization-841af926d8d28ab93538c46d8f31278a2234c4d1c12c7dc4dc296b249d59735a`;
|
||||
- E30 engineering generation:
|
||||
`e30-engineering-generation-62a4fea10dea9b77f69ceac1af5bf0e4928d9c7716083c22258a03670fe5bd4f`;
|
||||
- E30 human exception generation:
|
||||
`e30-review-generation-7982a882558d0be690b4c7092e328c080bfcbf52478a220452be7e887a588250`.
|
||||
|
||||
The result binds the exact manifest and decision-file SHA-256 values for all
|
||||
three inputs. Of 486 terminal labels, 484 come from the accepted engineering
|
||||
generation and two from the saved user exception decisions. These labels are
|
||||
not described as independent ground truth.
|
||||
|
||||
## Method
|
||||
|
||||
Profile: `e37-ravnoves00-r0-acceptance/v1`.
|
||||
|
||||
The deterministic builder:
|
||||
|
||||
1. verifies the exact E30 identities and artifact digests;
|
||||
2. requires one engineering decision for every materialized review item;
|
||||
3. requires human decisions to match the declared exception set exactly;
|
||||
4. projects each reviewed item into three independent dimensions:
|
||||
presence, geometry association and freshness;
|
||||
5. assigns severity without changing the reviewed outcome;
|
||||
6. creates a deterministic 30% validation holdout inside each source-stratum
|
||||
and range bucket using seed `ravnoves00-r0-validation-v1`;
|
||||
7. closes terminal accounting and rejects any free-space claim;
|
||||
8. publishes four immutable artifacts atomically.
|
||||
|
||||
The worker package contains only the runtime projection, profile and the six
|
||||
required input manifest/data files. Execution used the pinned Triton image
|
||||
with no network, a read-only container filesystem, dropped capabilities,
|
||||
`no-new-privileges`, a bounded PID limit and no GPU allocation.
|
||||
|
||||
## Contract
|
||||
|
||||
| Measure | Result |
|
||||
| --- | ---: |
|
||||
| reviewed items | 486 |
|
||||
| development items | 340 |
|
||||
| validation items | 146 |
|
||||
| terminal outcomes | 486 |
|
||||
| accounting | 100% |
|
||||
| engineering-reviewed labels | 484 |
|
||||
| human exception labels | 2 |
|
||||
| false-free claims | 0 |
|
||||
|
||||
Frozen distributions:
|
||||
|
||||
- presence: 300 `object-present`, 104 `occupied-environment`,
|
||||
82 `background-or-noise`;
|
||||
- geometry association: 108 `object-associated`,
|
||||
104 `independent-occupied`, 100 `insufficient-support`,
|
||||
92 `unknown`, 82 `rejected-nonobject`;
|
||||
- freshness: 292 `current`, 102 `unavailable`, 92 `stale`;
|
||||
- severity: 290 `standard`, 71 `medium`, 125 `high`.
|
||||
|
||||
The following targets are declared separately for R1 and later gates:
|
||||
|
||||
- presence quality: at least 90%;
|
||||
- geometry-association quality: at least 90%;
|
||||
- freshness quality: at least 90%;
|
||||
- accounting: 100%;
|
||||
- false-free claims: zero.
|
||||
|
||||
## Acceptance
|
||||
|
||||
All seven predeclared R0 checks pass:
|
||||
|
||||
- source identity is frozen;
|
||||
- reviewed denominator is complete;
|
||||
- development and validation split is complete;
|
||||
- every item has a terminal label in all three dimensions;
|
||||
- human exception accounting is complete;
|
||||
- false-free claims are zero;
|
||||
- authority remains diagnostic.
|
||||
|
||||
## What this proves
|
||||
|
||||
RAVNOVES00 now has one reproducible source-scoped evaluation contract. Future
|
||||
algorithm changes can use the 340-item development set without changing the
|
||||
146-item validation holdout, label ontology or denominator. Presence,
|
||||
geometry association and freshness can no longer be reported as one blended
|
||||
success score.
|
||||
|
||||
## What this does not prove
|
||||
|
||||
E37 does not evaluate or pass any 90% quality target. It does not provide
|
||||
independent ground truth, prove another route, camera, rig or mount, validate
|
||||
free space, or grant navigation, command or safety authority.
|
||||
|
||||
## Decision
|
||||
|
||||
Accept R0 as the only RAVNOVES00 measurement basis. The next planned K1
|
||||
laboratory gate is R1: compute the first source-scoped perception-quality
|
||||
baseline on the frozen validation holdout, diagnose the separate deficits and
|
||||
tune only against the development partition.
|
||||
|
||||
Reference in New Issue
Block a user