Decision record · assembled from platform records

Run 234 · Bodily Injury Frequency

v12 → v13 · approved by reviewer on 10 Aug 2026, 21:22 UTC

Generated 05 Sep 2026, 11:35 UTC directly from the platform's stored records. Nothing below was written for this document; each section says when it was recorded.

Status

No longer in force. The v13 this approval created was later replaced by run 252's approval of its own v13. The winner below still won this run; only the resulting version was replaced. Everything below is kept exactly as it was signed.

Checked when this record was generated — the only line in this document that is not a stored record.

The decision

EXP-07 created v13 from v12: +0.013 Gini on train, +0.011 on held-out data. 3 guardrails held. 26 agent actions recorded, 4 refused by the platform's own limits.

Weakest point named at sign-off: The one number to look at before approving: the 3+ prior accidents level carries 1.7% of exposure and rides partial credibility, so its relativity is the least settled part of this change.

Question to decision: the run took 9.3s from "Improve lift on Bodily Injury Frequency v12 without adding more than two rating factors. Keep territory relativities within filed tolerance." to a package a human could sign.

Frozen inside the approval transaction — recorded at sign-off, not assembled later.

Decision evidence

Driver age relativity
123420406080Driver ageRelativity, age 45 = 1.00v12 bands (solid) · Fitted spline (muted) · Knots (dotted)
Driver agev12 bands±2 SE rangeFitted spline±2 SE rangeKnotsEarned exposure
182.6452.452 to 2.8533.5133.202 to 3.8553.5131218.4
192.6452.452 to 2.8533.2302.968 to 3.515492.7
202.6452.452 to 2.8532.9722.748 to 3.213644.1
212.6452.452 to 2.8532.7372.543 to 2.946719.7
222.6452.452 to 2.8532.5262.353 to 2.712738.4
232.6452.452 to 2.8532.3362.178 to 2.506842.9
242.6452.452 to 2.8532.1682.020 to 2.327930.9
251.5211.418 to 1.6322.0191.880 to 2.1702.019956.7
261.5211.418 to 1.6321.8891.757 to 2.0311052.9
271.5211.418 to 1.6321.7741.649 to 1.9091108.2
281.5211.418 to 1.6321.6731.556 to 1.8001172.3
291.5211.418 to 1.6321.5841.474 to 1.7021152.9
301.5211.418 to 1.6321.5061.404 to 1.6151221.4
311.5211.418 to 1.6321.4361.342 to 1.5361238.5
321.5211.418 to 1.6321.3751.289 to 1.4661268.6
331.5211.418 to 1.6321.3201.243 to 1.4021342.2
341.5211.418 to 1.6321.2721.202 to 1.3451397.2
3511.2291.168 to 1.2931426.4
3611.1911.138 to 1.2471494.3
3711.1571.112 to 1.2051476.4
3811.1281.089 to 1.1681496.9
3911.1021.070 to 1.1341568.5
4011.0791.054 to 1.1041633.6
4111.0581.039 to 1.0781540.2
4211.0411.027 to 1.0541606.0
4311.0251.017 to 1.0341539.7
4411.0121.008 to 1.0161586.6
451111616.9
4610.9900.987 to 0.9931527.1
4710.9810.975 to 0.9881606.2
4810.9740.965 to 0.9841592.8
4910.9690.957 to 0.9811625.8
5010.9640.950 to 0.9791495.2
5110.9620.944 to 0.9801501.1
5210.9600.939 to 0.9811471.2
5310.9600.936 to 0.9841447.7
5410.9610.933 to 0.9891416.2
550.9560.890 to 1.0280.9630.932 to 0.9961305.6
560.9560.890 to 1.0280.9670.931 to 1.0041312.1
570.9560.890 to 1.0280.9720.932 to 1.0131220.2
580.9560.890 to 1.0280.9780.934 to 1.0241209.8
590.9560.890 to 1.0280.9850.937 to 1.0371155.2
600.9560.890 to 1.0280.9940.941 to 1.0501159.3
610.9560.890 to 1.0281.0050.947 to 1.0661181.7
620.9560.890 to 1.0281.0160.954 to 1.0821233.1
630.9560.890 to 1.0281.0290.962 to 1.1001195.0
640.9560.890 to 1.0281.0430.972 to 1.1201209.9
650.9560.890 to 1.0281.0590.984 to 1.1411142.1
660.9560.890 to 1.0281.0770.997 to 1.1631129.5
670.9560.890 to 1.0281.0951.012 to 1.1861021.0
680.9560.890 to 1.0281.1161.029 to 1.2111035.3
690.9560.890 to 1.0281.1381.048 to 1.237919.3
701.4131.309 to 1.5261.1621.069 to 1.2641.162830.6
711.4131.309 to 1.5261.1881.092 to 1.293774.9
721.4131.309 to 1.5261.2161.117 to 1.323750.7
731.4131.309 to 1.5261.2451.145 to 1.354658.4
741.4131.309 to 1.5261.2761.174 to 1.388664.8
751.4131.309 to 1.5261.3101.206 to 1.423627.8
761.4131.309 to 1.5261.3451.239 to 1.460629.4
771.4131.309 to 1.5261.3831.274 to 1.500559.6
781.4131.309 to 1.5261.4221.311 to 1.543525.3
791.4131.309 to 1.5261.4641.349 to 1.588492.4
801.4131.309 to 1.5261.5071.388 to 1.637422.7
811.4131.309 to 1.5261.5531.428 to 1.690342.1
821.4131.309 to 1.5261.6011.468 to 1.746344.5
831.4131.309 to 1.5261.6521.510 to 1.807301.9
841.4131.309 to 1.5261.7041.551 to 1.872259.3
851.4131.309 to 1.5261.7591.593 to 1.942267.6
861.4131.309 to 1.5261.8161.636 to 2.016247.5
871.4131.309 to 1.5261.8751.679 to 2.095240.9
881.4131.309 to 1.5261.9371.722 to 2.179213.6
891.4131.309 to 1.5262.0011.766 to 2.267180.3
901.4131.309 to 1.5262.0671.810 to 2.3602.067379.1

Fitted peak 3.51 at age 18, and the curve reaches 2.07 at age 90 · Knots are the join points of the smooth pieces, not breaks in the price

Actual frequency by risk decile
00.1000.20012345678910Decile of predicted rate, equal exposureClaims per car yearActual, this model (solid) · Predicted (muted) · Actual, v12 deciles (dotted)
Decile of predicted rate, equal exposureActual, this modelPredictedActual, v12 decilesEarned exposure
10.0430.0430.0467531.5
20.0530.0530.0557530.6
30.0620.0590.0627531.2
40.0640.0660.0677531.1
50.0740.0730.0727530.3
60.0780.0800.0807530.9
70.0940.0900.0897531.4
80.1030.1030.1027530.5
90.1180.1250.1207530.8
100.1990.1950.1947530.8

Top decile against bottom decile: 4.67x here, 4.23x on v12

Rendered from the run's frozen evidence inside the approval transaction at sign-off (10 Aug 2026, 21:22 UTC) — recorded at sign-off, not reconstructed at read time.

What the reviewer was told

EXP-07 combines the driver_age spline with the capped prior accidents count. Train lift is +0.013 and the out of time holdout comes in at +0.011, shrinkage in the expected range.

The one number to look at before approving: the 3+ prior accidents level carries 1.7% of exposure and rides partial credibility, so its relativity is the least settled part of this change.

Written when the review was opened, before approval.

Guardrails

LimitHow it held
Factor budget2 of 2 used. The spline replaces the banded age factor but still counts against the budget; prior accidents is the second. Net new factors in the plan: one, so the merged model carries 10.
Territory stabilityFiled territory relativities stayed frozen in every fit except EXP-03, which tested unfreezing them and was scrapped for the 14.4% swing. With relativities pinned, the largest average rate movement inside any one territory, driven by the new factors, is 3.4%, on T-127, within the 5% limit.
Lift validationLift held on 5 of 5 random folds during the run, then one out of time holdout on 2025 H2. Holdout lift +0.011 against train +0.013.

The limits the agent worked under, as shown at review time.

Experiment ledger

All 7 experiments and why each lived or died.

CodeDispositionWhy
EXP-07Winnerpromoted in this review
EXP-01Absorbedinto EXP-07, the spline carried forward
EXP-04Absorbedinto EXP-07, the capped count carried forward
EXP-02Scrappedgain does not survive fold jitter
EXP-03Scrappedterritory relativities move 14.4%, outside the 5% guardrail
EXP-05ScrappedAIC prefers it, but lift is flat and the family change adds filing complexity
EXP-06Scrapped22% of mileage is missing and the gaps correlate with region, needs data work first

Written as each experiment landed.

Agent action record

Every action the platform recorded for this run, in order, starting 10 Aug 2026, 21:22 UTC — including what the agent refused to do and the one irreversible human action at the end.

#Time (UTC)WhoActionTargetDetail
121:22:20AgentreadPolicy datasetProfiled 100,000 policy rows; Bodily Injury claim counts with earned exposure as the target.
221:22:20Agentreadv12 filed territory relativitiesRecomputed the 2023 filing table from its own procedure. The table stays frozen for every experiment except EXP-03's blend.
321:22:20AgentfitBaseline v12 on trainRefit the model in force on 83,332 train rows. Baseline Gini 0.2362; every experiment is judged against this fit.
421:22:20Agentfit5-fold baseline cross-validationFit the baseline once per fold so every experiment's fold deltas compare against the same held-out scores.
521:22:20AgentchangeDriver age termEXP-01 changes the specification on the run branch only; v12 itself is untouched.
5 coarse bands → natural cubic spline
621:22:21AgentfitEXP-01 GLM fit and fold CVTrain Gini 0.2423 (+0.0061 vs baseline), scored across 5 folds.
721:22:21AgentchangeDriver age × vehicle ageEXP-02 changes the specification on the run branch only; v12 itself is untouched.
absent → interaction term
821:22:22AgentfitEXP-02 GLM fit and fold CVTrain Gini 0.2363 (+0.0001 vs baseline), scored across 5 folds.
921:22:22AgentrefuseEXP-02 promotionA guardrail stops promotion. The agent cannot carry this change forward.
Refused: lift held in only 3 of 5 folds
1021:22:22AgentrevertEXP-02 specificationChange not carried. The run branch keeps v12 plus surviving candidates only.
1121:22:22AgentchangeTerritory relativitiesEXP-03 changes the specification on the run branch only; v12 itself is untouched.
2023 filed table → credibility blend toward recent experience
1221:22:24AgentfitEXP-03 GLM fit and fold CVTrain Gini 0.2399 (+0.0037 vs baseline), scored across 5 folds.
1321:22:24AgentrefuseEXP-03 promotionA guardrail stops promotion. The agent cannot carry this change forward.
Refused: territory movement 14.4% in T-103 beyond the filed 5% tolerance; lift held in only 2 of 5 folds
1421:22:24AgentrevertEXP-03 specificationChange not carried. The run branch keeps v12 plus surviving candidates only.
1521:22:24AgentchangePrior accidentsEXP-04 changes the specification on the run branch only; v12 itself is untouched.
absent → count capped at 3
1621:22:24AgentfitEXP-04 GLM fit and fold CVTrain Gini 0.2434 (+0.0072 vs baseline), scored across 5 folds.
1721:22:24AgentchangeError familyEXP-05 changes the specification on the run branch only; v12 itself is untouched.
Poisson → negative binomial
1821:22:28AgentfitEXP-05 GLM fit and fold CVTrain Gini 0.2362 (+0.0000 vs baseline), scored across 5 folds.
1921:22:28AgentrefuseEXP-05 promotionA guardrail stops promotion. The agent cannot carry this change forward.
Refused: lift held in only 2 of 5 folds
2021:22:28AgentrevertEXP-05 specificationChange not carried. The run branch keeps v12 plus surviving candidates only.
2121:22:28AgentrefuseEXP-06 fitDeclined to fit on the platform's data profile. The refusal keeps its artifact: the missingness evidence below.
Refused: 22% of mileage is missing and the gaps correlate with region, needs data work first
2221:22:28AgentchangeAge spline + capped accidentsEXP-07 changes the specification on the run branch only; v12 itself is untouched.
two separate candidates → one combined specification
2321:22:29AgentfitEXP-07 GLM fit and fold CVTrain Gini 0.2494 (+0.0132 vs baseline), scored across 5 folds.
2421:22:29AgentchangeRun winnerPromoted the best candidate whose guardrails all held, then confirmed the gain on the 2025H2 holdout (+0.0110 Gini).
no winner → EXP-07
2521:22:29AgenthandoffHuman reviewWrote the review summary and requested human review. The agent cannot approve; creating a model version is the human's action alone.
2621:22:31Humanapprove · irreversibleModel version v13Approved EXP-07 and created v13. The run's only irreversible action, and only a human can take it.

Recorded as each action happened — a byproduct of the work, not a reconstruction.

Provenance