|
1 | 1 | { |
2 | | - "fixture_schema_version": "1", |
3 | | - "fixture_kind": "CONTROLLED_OBSERVATION_ONLY_REPLAY", |
4 | | - "scorer_id": "jacobian.observation-only-trajectory-scorer.v1", |
5 | | - "comparison_digest": "sha256:9aaa3654eae49d527ed60541909a89e0183fcf26c0ea8617682bd5fddd99df22", |
6 | | - "source_corpus_digest": "sha256:6b17b8fcc411a533c822a5117e33cf73fa6766739d7d36ac86dc4bea883af4fe", |
7 | | - "trajectory_id": "target", |
| 2 | + "comparison_digest": "sha256:825e074f83990efa74cefd4159085668261dd9d3ee836c4662d0e093eb6789af", |
8 | 3 | "estimator": "HYBRID_TYPED_TEXT", |
9 | 4 | "eventual_terminal_result": "ACCEPTED", |
10 | 5 | "eventual_terminal_reward": 1, |
| 6 | + "fixture_kind": "CONTROLLED_OBSERVATION_ONLY_REPLAY", |
| 7 | + "fixture_schema_version": "1", |
| 8 | + "interpretation": "The negative AFTER_TOOL value change remains visible but earns zero credit because no typed milestone occurred.", |
| 9 | + "limitations": [ |
| 10 | + "The values are controlled fixture estimates, not real Codex predictions.", |
| 11 | + "The scorer replays a frozen PR2 comparison and does not fit or update clusters.", |
| 12 | + "Predictive and early-warning conclusions remain deferred to PR4." |
| 13 | + ], |
| 14 | + "scorer_id": "jacobian.observation-only-trajectory-scorer.v1", |
| 15 | + "source_corpus_digest": "sha256:6b17b8fcc411a533c822a5117e33cf73fa6766739d7d36ac86dc4bea883af4fe", |
11 | 16 | "states": [ |
12 | 17 | { |
13 | | - "observation_id": "target:0", |
14 | 18 | "boundary": "PLAN", |
15 | | - "milestone_eligible": false, |
| 19 | + "credit_reason": "INITIAL_STATE", |
| 20 | + "cumulative_milestone_credit": 0.0, |
16 | 21 | "estimated_value": 0.4, |
17 | | - "value_delta": null, |
| 22 | + "milestone_eligible": false, |
| 23 | + "observation_id": "target:0", |
18 | 24 | "transition_credit": 0.0, |
19 | | - "cumulative_milestone_credit": 0.0, |
20 | | - "credit_reason": "INITIAL_STATE" |
| 25 | + "value_delta": null |
21 | 26 | }, |
22 | 27 | { |
23 | | - "observation_id": "target:1", |
24 | 28 | "boundary": "TOOL_RESULT", |
25 | | - "milestone_eligible": true, |
| 29 | + "credit_reason": "MILESTONE_VALUE_DELTA", |
| 30 | + "cumulative_milestone_credit": 0.3, |
26 | 31 | "estimated_value": 0.7, |
27 | | - "value_delta": 0.3, |
| 32 | + "milestone_eligible": true, |
| 33 | + "observation_id": "target:1", |
28 | 34 | "transition_credit": 0.3, |
29 | | - "cumulative_milestone_credit": 0.3, |
30 | | - "credit_reason": "MILESTONE_VALUE_DELTA" |
| 35 | + "value_delta": 0.3 |
31 | 36 | }, |
32 | 37 | { |
33 | | - "observation_id": "target:2", |
34 | 38 | "boundary": "AFTER_TOOL", |
35 | | - "milestone_eligible": false, |
| 39 | + "credit_reason": "NON_MILESTONE_ZERO", |
| 40 | + "cumulative_milestone_credit": 0.3, |
36 | 41 | "estimated_value": 0.2, |
37 | | - "value_delta": -0.5, |
| 42 | + "milestone_eligible": false, |
| 43 | + "observation_id": "target:2", |
38 | 44 | "transition_credit": 0.0, |
39 | | - "cumulative_milestone_credit": 0.3, |
40 | | - "credit_reason": "NON_MILESTONE_ZERO" |
| 45 | + "value_delta": -0.5 |
41 | 46 | }, |
42 | 47 | { |
43 | | - "observation_id": "target:3", |
44 | 48 | "boundary": "TOOL_RESULT", |
45 | | - "milestone_eligible": true, |
| 49 | + "credit_reason": "MILESTONE_VALUE_DELTA", |
| 50 | + "cumulative_milestone_credit": 0.7, |
46 | 51 | "estimated_value": 0.6, |
47 | | - "value_delta": 0.4, |
| 52 | + "milestone_eligible": true, |
| 53 | + "observation_id": "target:3", |
48 | 54 | "transition_credit": 0.4, |
49 | | - "cumulative_milestone_credit": 0.7, |
50 | | - "credit_reason": "MILESTONE_VALUE_DELTA" |
| 55 | + "value_delta": 0.4 |
51 | 56 | } |
52 | 57 | ], |
53 | 58 | "total_milestone_credit": 0.7, |
54 | | - "interpretation": "The negative AFTER_TOOL value change remains visible but earns zero credit because no typed milestone occurred.", |
55 | | - "limitations": [ |
56 | | - "The values are controlled fixture estimates, not real Codex predictions.", |
57 | | - "The scorer replays a frozen PR2 comparison and does not fit or update clusters.", |
58 | | - "Predictive and early-warning conclusions remain deferred to PR4." |
59 | | - ] |
| 59 | + "trajectory_id": "target" |
60 | 60 | } |
0 commit comments