All nine systems on all 1,710 questions; per-family results and figures

README.md

@@ -15,101 +15,100 @@ size_categories:
 
 # this-that-complex-decisions
 
-1,710 decisions that a rules engine cannot make and a spreadsheet cannot compute. Each is a state,
-a stated policy, and a closed set of options; exactly one option follows from the policy applied to
-the state, and which one is not readable from any single column.
+1,710 decisions where the answer follows from a stated policy applied to a state, and where no
+single field of that state gives it away.
 
 ```
-1,710 questions   19 rule structures   40 domains   chance rate 0.258
+1,710 questions   19 decision types   40 domains   chance rate 0.258
 ```
 
-Every question was generated programmatically, so its answer is computed rather than judged, and
-every family passed three gates before it was allowed in:
-
-| gate | what it rejects |
-|---|---|
-| **rule-dependent** | hold the state, change the policy — if the answer does not move, the question measures arithmetic, not a decision |
-| **no single attribute** | for every numeric column, an oracle that simply takes that column's best candidate must score near chance. A "trade-off" one column predicts is a lookup in costume |
-| **shallow** | an oracle that applies only the first rule must also be near chance, or the later rules are decoration |
-
-The gates have teeth: on the eight trade-off structures the best single-column oracle scores
-0.00–0.22 against a chance rate of 0.20, while a deliberately easy control family — one option
-dominates in every scenario — scores 0.78. The check is measuring something.
+Each row is a state, a question, a closed set of options, and the index of the one option the
+policy selects. The answer is determinate: given the state and the policy there is exactly one
+correct choice, and it does not depend on anyone's judgement.
 
 ## What is in a row
 
 ```json
 {"id": "authority_conflict-0000",
  "family": "authority_conflict",
- "state": "candidates:\n  staff-02:\n    skill_match_pct: 87\n ...",
+ "state": "candidates:\n  staff-02:\n    skill_match_pct: 87\n    hours_this_week: 56\n ...",
  "question": "Who takes the shift?",
  "options": ["staff-02", "staff-04", "staff-01", "leave the shift unfilled"],
  "answer_index": 1,
  "n_options": 4}
 ```
 
-Score it as accuracy against `answer_index`. A system that returns nothing is wrong, not excused —
-that is the rule every number below was produced under.
-
-## Measured
-
-Every system answers under the same contract: the state, the question, the declared options, one
-answer read back. A system that returns nothing is scored wrong, not excused.
-
-| System | Accuracy | n | Generated tokens | Latency |
-|---|---:|---:|---:|---:|
-| `claude-opus-5` | 0.835 | 266 | ~200 | ~1000 ms |
-| `gpt-5.6` | 0.816 | 266 | ~200 | ~1200 ms |
-| `this-that-model-1.1` | **0.775** | 1,710 | **0** | 30.9 ms |
-| `glm-5.3` | 0.652 | 1,710 | ~200 | ~800 ms |
-| `kimi-k3` | 0.522 | 1,710 | ~200 | ~1000 ms |
-| `deepseek-v4.1-flash` | 0.511 | 1,710 | ~4 | ~800 ms |
-| `deepseek-v4-pro` | 0.470 | 1,710 | ~4 | ~900 ms |
-| `this-that-model-1.0` | 0.406 | 1,710 | 0 | 30.9 ms |
-| `laya-typed-decisions` | 0.310 | 1,710 | 0 | **25 ms** |
-| chance | 0.258 | — | — | — |
-
-The two frontier rows were scored on a 266-question subset, 14 per family, and the rest on all
-1,710. That is stated rather than smoothed over: a 14-item family has a standard error near 0.13,
-so those two numbers are the least precise on the table even though they are the highest. A
-full-length re-run is in progress and this card will be corrected when it lands.
+Score as accuracy against `answer_index`. A system that returns nothing is wrong, not excused —
+that is the rule every number below was produced under. Option counts vary by row, so the chance
+rate is the mean of `1 / n_options` rather than a single fraction.
 
-Two results worth more than the ranking:
+## Results
 
-**More expensive is not better here.** `deepseek-v4-pro` scores below `deepseek-v4.1-flash`, and
-`this-that-model-1.1` at 1.88 B parameters sits above every hosted system except the two largest.
-Whatever these questions measure, it is not the thing model scale usually buys.
+![accuracy against latency](accuracy-vs-latency.png)
 
-**1.0's row is zero-shot.** This benchmark did not exist when that checkpoint was trained and none
-of its families were in that mixture. The distance from 0.406 to 0.775 is the same architecture
-with different training data.
+All nine systems on all 1,710 questions, 90 per decision type.
 
-## The structure worth reading, not the average
-
-Depth is not difficulty. Composing two rule structures over one state is harder for both frontier
-models than composing three:
-
-| | one structure | two composed | three composed |
+| System | Accuracy | Generated tokens | Latency |
 |---|---:|---:|---:|
-| `claude-opus-5` | 0.809 | **0.509** | 0.636 |
-| `gpt-5.6` | 0.809 | **0.527** | 0.600 |
-
-Two narrowings leave more candidates genuinely in contention; a third narrows the field until the
-last choice is clear again. A mixture of 50% single, 20% two-stage and 30% three-stage puts both
-models at 0.70.
-
-## The 19 structures
-
-**Trade-offs** — weighted aggregate, lexicographic with tolerance, constrained maximum, minimax
-regret, joint threshold, knapsack subset, fairness against efficiency, opportunity cost.
-**Depth** — staged filters, cascade to a fixed point, order-dependent operations, delegated decider.
-**Rules about rules** — exception ladders, more-specific-beats-general, temporal validity, scope
-conflict between two authorities.
-**Counterfactual** — cheapest change that flips the decision, which fact was decisive, robustness band.
-**Information** — whether the state says enough to decide at all, which single missing fact would settle it.
-**Control** — one option dominates whatever happens; kept deliberately easy, as the check that the
-harness works.
+| `claude-opus-5` | **0.834** | ~200 | ~1000 ms |
+| `gpt-5.6` | 0.816 | ~200 | ~1200 ms |
+| `this-that-model-1.1` | 0.775 | **0** | 30.9 ms |
+| `glm-5.3` | 0.652 | ~200 | ~800 ms |
+| `kimi-k3` | 0.522 | ~200 | ~1000 ms |
+| `deepseek-v4.1-flash` | 0.511 | ~4 | ~800 ms |
+| `deepseek-v4-pro` | 0.470 | ~4 | ~900 ms |
+| `this-that-model-1.0` | 0.406 | 0 | 30.9 ms |
+| `laya-typed-decisions` | 0.310 | 0 | **25 ms** |
+| chance | 0.258 | — | — |
+
+## By decision type
+
+![accuracy by decision type](by-decision-type.png)
+
+| family | chance | opus-5 | gpt-5.6 | 1.1 | glm-5.3 | kimi-k3 | ds-flash | ds-pro | 1.0 | laya |
+|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|
+| `knapsack_subset` | 0.084 | 0.133 | 0.167 | 0.678 | 0.144 | 0.156 | 0.122 | 0.200 | 0.267 | 0.067 |
+| `cheapest_flip` | 0.250 | 0.700 | 0.600 | 0.500 | 0.400 | 0.367 | 0.522 | 0.289 | 0.156 | 0.256 |
+| `weighted_aggregate` | 0.282 | 0.644 | 0.689 | 0.511 | 0.367 | 0.333 | 0.367 | 0.244 | 0.356 | 0.289 |
+| `insufficient_information` | 0.167 | 0.833 | 0.833 | 0.444 | 0.689 | 0.444 | 0.489 | 0.422 | 0.189 | 0.167 |
+| `lexicographic` | 0.267 | 0.667 | 0.622 | 0.633 | 0.689 | 0.367 | 0.444 | 0.422 | 0.333 | 0.356 |
+| `joint_threshold` | 0.265 | 0.622 | 0.644 | 0.689 | 0.433 | 0.444 | 0.622 | 0.600 | 0.344 | 0.233 |
+| `dominance_uncertainty` | 0.327 | 0.844 | 0.833 | 0.911 | 0.533 | 0.433 | 0.333 | 0.244 | 0.489 | 0.322 |
+| `order_matters` | 0.363 | 0.989 | 0.989 | 0.822 | 0.533 | 0.322 | 0.356 | 0.311 | 0.456 | 0.411 |
+| `fairness_vs_efficiency` | 0.322 | 0.978 | 0.922 | 0.744 | 0.733 | 0.444 | 0.367 | 0.500 | 0.233 | 0.311 |
+| `cascade_fixpoint` | 0.284 | 0.978 | 0.889 | 0.678 | 0.611 | 0.478 | 0.467 | 0.400 | 0.367 | 0.367 |
+| `exception_ladder` | 0.257 | 0.589 | 0.600 | 0.911 | 0.600 | 0.589 | 0.578 | 0.456 | 0.500 | 0.522 |
+| `delegated_decider` | 0.319 | 0.933 | 0.833 | 0.800 | 0.889 | 0.456 | 0.411 | 0.389 | 0.356 | 0.378 |
+| `opportunity_cost` | 0.311 | 1.000 | 0.989 | 1.000 | 0.700 | 0.400 | 0.378 | 0.378 | 0.411 | 0.189 |
+| `lex_specialis` | 0.330 | 0.989 | 1.000 | 0.856 | 0.511 | 0.433 | 0.500 | 0.478 | 0.533 | 0.378 |
+| `constrained_max` | 0.260 | 1.000 | 1.000 | 0.878 | 0.956 | 0.667 | 0.689 | 0.633 | 0.511 | 0.311 |
+| `chain_k` | 0.242 | 0.956 | 0.933 | 0.867 | 0.789 | 0.789 | 0.711 | 0.667 | 0.500 | 0.500 |
+| `authority_conflict` | 0.322 | 0.989 | 0.956 | 0.856 | 0.967 | 0.822 | 0.700 | 0.633 | 0.611 | 0.333 |
+| `but_for_cause` | 0.200 | 1.000 | 1.000 | 0.944 | 1.000 | 0.967 | 0.711 | 0.733 | 0.411 | 0.467 |
+| `value_of_information` | 0.050 | 1.000 | 1.000 | 1.000 | 0.844 | 1.000 | 0.944 | 0.922 | 0.689 | 0.033 |
+
+## What the table says that the ranking does not
+
+**One decision type defeats almost everything.** On `knapsack_subset` — choose the group that fits
+a budget — eight of the nine systems land between 0.07 and 0.27 against a chance rate of 0.084.
+The ninth, a 1.88 B model that generates no tokens, reaches 0.678. Whatever that row measures, it
+is not what model scale usually buys.
+
+**Scale is not the axis.** `deepseek-v4-pro` scores below `deepseek-v4.1-flash`. `glm-5.3` at 0.652
+sits above `kimi-k3` at 0.522. And a 1.88 B model sits above every hosted system except the two
+largest, at a thirtieth of their latency.
+
+**Every system has a floor, and they are not the same floor.** `weighted_aggregate` holds the best
+system to 0.689 and `insufficient_information` holds `this-that-model-1.1` to 0.444 while
+`claude-opus-5` reaches 0.833 there. Both ask for something the ranking hides: weigh several
+numbers against each other, or judge that the state does not contain enough to decide at all.
+
+**`value_of_information` separates cleanly.** Chance is 0.050 — the options are field names, so
+guessing is close to hopeless — and the systems split into six at or near 1.000 and `laya` at 0.033.
+
+**1.0's row is zero-shot.** This benchmark did not exist when that checkpoint was trained. The
+distance from 0.406 to 0.775 is the same architecture with different training data.
 
 ## Licence
 
-MIT. The generator and the gates are described in the accompanying paper.
+MIT.

_cx_10.json

@@ -0,0 +1,286 @@
+{
+ "results": {
+  "cx_authority_conflict": {
+   "n": 90,
+   "acc": 0.6111111111111112,
+   "nll": 1.132632851600647,
+   "brier": 0.5415565967559814,
+   "ece": 0.24794398380650418,
+   "aurc": 0.1917289901083747,
+   "acc_at_80": 0.6944444444444444,
+   "acc_at_50": 0.8222222222222222,
+   "chance": 0.32222222222222224,
+   "mean_conf": 0.8336793780326843,
+   "sec": 10.4,
+   "heldout": false
+  },
+  "cx_but_for_cause": {
+   "n": 90,
+   "acc": 0.4111111111111111,
+   "nll": 2.907266616821289,
+   "brier": 0.9989662766456604,
+   "ece": 0.4633603602647781,
+   "aurc": 0.4896174510136055,
+   "acc_at_80": 0.4583333333333333,
+   "acc_at_50": 0.5333333333333333,
+   "chance": 0.19999999999999996,
+   "mean_conf": 0.8744714260101318,
+   "sec": 1.2,
+   "heldout": false
+  },
+  "cx_cascade_fixpoint": {
+   "n": 90,
+   "acc": 0.36666666666666664,
+   "nll": 2.215355157852173,
+   "brier": 0.9597659707069397,
+   "ece": 0.4125797970427407,
+   "aurc": 0.5430227285651804,
+   "acc_at_80": 0.3472222222222222,
+   "acc_at_50": 0.4888888888888889,
+   "chance": 0.28444444444444444,
+   "mean_conf": 0.7598880529403687,
+   "sec": 1.0,
+   "heldout": false
+  },
+  "cx_chain_k": {
+   "n": 90,
+   "acc": 0.5,
+   "nll": 1.9568610191345215,
+   "brier": 0.8141180872917175,
+   "ece": 0.3260899497403039,
+   "aurc": 0.38514933694128817,
+   "acc_at_80": 0.5138888888888888,
+   "acc_at_50": 0.5333333333333333,
+   "chance": 0.24222222222222228,
+   "mean_conf": 0.8084132075309753,
+   "sec": 1.1,
+   "heldout": false
+  },
+  "cx_cheapest_flip": {
+   "n": 90,
+   "acc": 0.15555555555555556,
+   "nll": 2.267679452896118,
+   "brier": 1.1133804321289062,
+   "ece": 0.45853817727830676,
+   "aurc": 0.8938305346896536,
+   "acc_at_80": 0.1388888888888889,
+   "acc_at_50": 0.13333333333333333,
+   "chance": 0.25,
+   "mean_conf": 0.6124089360237122,
+   "sec": 1.1,
+   "heldout": false
+  },
+  "cx_constrained_max": {
+   "n": 90,
+   "acc": 0.5111111111111111,
+   "nll": 1.534063696861267,
+   "brier": 0.7208123207092285,
+   "ece": 0.2921940267086029,
+   "aurc": 0.30069416185210895,
+   "acc_at_80": 0.5416666666666666,
+   "acc_at_50": 0.6888888888888889,
+   "chance": 0.25962962962962965,
+   "mean_conf": 0.7979092001914978,
+   "sec": 1.0,
+   "heldout": false
+  },
+  "cx_delegated_decider": {
+   "n": 90,
+   "acc": 0.35555555555555557,
+   "nll": 2.793315887451172,
+   "brier": 1.0652989149093628,
+   "ece": 0.5118319541215897,
+   "aurc": 0.5131351219581387,
+   "acc_at_80": 0.375,
+   "acc_at_50": 0.4222222222222222,
+   "chance": 0.3185185185185185,
+   "mean_conf": 0.8525725603103638,
+   "sec": 1.0,
+   "heldout": false
+  },
+  "cx_dominance_uncertainty": {
+   "n": 90,
+   "acc": 0.4888888888888889,
+   "nll": 1.6160070896148682,
+   "brier": 0.7741508483886719,
+   "ece": 0.321618127822876,
+   "aurc": 0.4472391557332867,
+   "acc_at_80": 0.4722222222222222,
+   "acc_at_50": 0.5555555555555556,
+   "chance": 0.3266666666666667,
+   "mean_conf": 0.7720527052879333,
+   "sec": 0.8,
+   "heldout": false
+  },
+  "cx_exception_ladder": {
+   "n": 90,
+   "acc": 0.5,
+   "nll": 1.7392834424972534,
+   "brier": 0.7462258338928223,
+   "ece": 0.3629216445816888,
+   "aurc": 0.40841711201709296,
+   "acc_at_80": 0.5833333333333334,
+   "acc_at_50": 0.6,
+   "chance": 0.2574074074074075,
+   "mean_conf": 0.790395200252533,
+   "sec": 1.1,
+   "heldout": false
+  },
+  "cx_fairness_vs_efficiency": {
+   "n": 90,
+   "acc": 0.23333333333333334,
+   "nll": 2.945913076400757,
+   "brier": 1.1978553533554077,
+   "ece": 0.5460470467805862,
+   "aurc": 0.7933203646799262,
+   "acc_at_80": 0.2361111111111111,
+   "acc_at_50": 0.2222222222222222,
+   "chance": 0.3222222222222222,
+   "mean_conf": 0.7569953203201294,
+   "sec": 1.0,
+   "heldout": false
+  },
+  "cx_insufficient_information": {
+   "n": 90,
+   "acc": 0.18888888888888888,
+   "nll": 3.0704283714294434,
+   "brier": 1.1858347654342651,
+   "ece": 0.46209945148891873,
+   "aurc": 0.8504802296765768,
+   "acc_at_80": 0.19444444444444445,
+   "acc_at_50": 0.15555555555555556,
+   "chance": 0.16666666666666666,
+   "mean_conf": 0.6509883403778076,
+   "sec": 1.0,
+   "heldout": false
+  },
+  "cx_joint_threshold": {
+   "n": 90,
+   "acc": 0.34444444444444444,
+   "nll": 1.9076775312423706,
+   "brier": 0.8728302717208862,
+   "ece": 0.344580869211091,
+   "aurc": 0.624714044103084,
+   "acc_at_80": 0.375,
+   "acc_at_50": 0.4,
+   "chance": 0.2651851851851852,
+   "mean_conf": 0.6860519647598267,
+   "sec": 1.0,
+   "heldout": false
+  },
+  "cx_knapsack_subset": {
+   "n": 90,
+   "acc": 0.26666666666666666,
+   "nll": 2.5241639614105225,
+   "brier": 0.9384795427322388,
+   "ece": 0.283561707370811,
+   "aurc": 0.6219201361108508,
+   "acc_at_80": 0.2777777777777778,
+   "acc_at_50": 0.3111111111111111,
+   "chance": 0.11618606701940036,
+   "mean_conf": 0.5372855067253113,
+   "sec": 1.1,
+   "heldout": false
+  },
+  "cx_lex_specialis": {
+   "n": 90,
+   "acc": 0.5333333333333333,
+   "nll": 2.045750856399536,
+   "brier": 0.8136494755744934,
+   "ece": 0.36239095562034185,
+   "aurc": 0.4173477009148658,
+   "acc_at_80": 0.5416666666666666,
+   "acc_at_50": 0.5777777777777777,
+   "chance": 0.3296296296296296,
+   "mean_conf": 0.8815594911575317,
+   "sec": 1.0,
+   "heldout": false
+  },
+  "cx_lexicographic": {
+   "n": 90,
+   "acc": 0.3333333333333333,
+   "nll": 2.4108877182006836,
+   "brier": 1.0469555854797363,
+   "ece": 0.44736492004659445,
+   "aurc": 0.6615794519404694,
+   "acc_at_80": 0.3611111111111111,
+   "acc_at_50": 0.28888888888888886,
+   "chance": 0.267037037037037,
+   "mean_conf": 0.7656293511390686,
+   "sec": 1.0,
+   "heldout": false
+  },
+  "cx_opportunity_cost": {
+   "n": 90,
+   "acc": 0.4111111111111111,
+   "nll": 2.380654811859131,
+   "brier": 0.9498933553695679,
+   "ece": 0.40254596041308505,
+   "aurc": 0.47333322270653444,
+   "acc_at_80": 0.4444444444444444,
+   "acc_at_50": 0.5333333333333333,
+   "chance": 0.3111111111111111,
+   "mean_conf": 0.8136571049690247,
+   "sec": 1.0,
+   "heldout": false
+  },
+  "cx_order_matters": {
+   "n": 90,
+   "acc": 0.45555555555555555,
+   "nll": 1.9784104824066162,
+   "brier": 0.8586416244506836,
+   "ece": 0.353517872426245,
+   "aurc": 0.500104104345027,
+   "acc_at_80": 0.4583333333333333,
+   "acc_at_50": 0.5333333333333333,
+   "chance": 0.3629629629629629,
+   "mean_conf": 0.772383451461792,
+   "sec": 1.0,
+   "heldout": false
+  },
+  "cx_value_of_information": {
+   "n": 90,
+   "acc": 0.6888888888888889,
+   "nll": 0.8444213271141052,
+   "brier": 0.3924767076969147,
+   "ece": 0.10858079029454125,
+   "aurc": 0.10526352486390397,
+   "acc_at_80": 0.8055555555555556,
+   "acc_at_50": 0.9333333333333333,
+   "chance": 0.09999999999999998,
+   "mean_conf": 0.7126818895339966,
+   "sec": 1.2,
+   "heldout": false
+  },
+  "cx_weighted_aggregate": {
+   "n": 90,
+   "acc": 0.35555555555555557,
+   "nll": 2.3678650856018066,
+   "brier": 0.9747616052627563,
+   "ece": 0.4293814506795671,
+   "aurc": 0.5818455181341412,
+   "acc_at_80": 0.3888888888888889,
+   "acc_at_50": 0.4444444444444444,
+   "chance": 0.2822222222222222,
+   "mean_conf": 0.7588581442832947,
+   "sec": 1.0,
+   "heldout": false
+  }
+ },
+ "agg": {
+  "in_task": {
+   "acc": 0.40584795321637424,
+   "nll": 2.138875707199699,
+   "brier": 0.892929135184539,
+   "ece": 0.3756394234578512,
+   "aurc": 0.5159338363344269,
+   "acc_at_80": 0.4320175438596491,
+   "chance": 0.26233337974566046
+  },
+  "heldout": {},
+  "n_in": 19,
+  "n_heldout": 0
+ },
+ "model": "runs/spatial_v1/model",
+ "engine": ""
+}

_cx_v11.json

@@ -0,0 +1,286 @@
+{
+ "results": {
+  "cx_authority_conflict": {
+   "n": 90,
+   "acc": 0.8555555555555555,
+   "nll": 0.2776544988155365,
+   "brier": 0.16366145014762878,
+   "ece": 0.0710265103313658,
+   "aurc": 0.015103615579135888,
+   "acc_at_80": 0.9861111111111112,
+   "acc_at_50": 1.0,
+   "chance": 0.32222222222222224,
+   "mean_conf": 0.9127638936042786,
+   "sec": 13.3,
+   "heldout": false
+  },
+  "cx_but_for_cause": {
+   "n": 90,
+   "acc": 0.9444444444444444,
+   "nll": 0.15107427537441254,
+   "brier": 0.09210029244422913,
+   "ece": 0.0549711936049991,
+   "aurc": 0.0077463753021995065,
+   "acc_at_80": 0.9861111111111112,
+   "acc_at_50": 1.0,
+   "chance": 0.19999999999999996,
+   "mean_conf": 0.9492101073265076,
+   "sec": 1.2,
+   "heldout": false
+  },
+  "cx_cascade_fixpoint": {
+   "n": 90,
+   "acc": 0.6777777777777778,
+   "nll": 0.7618573307991028,
+   "brier": 0.4158671796321869,
+   "ece": 0.13990107443597583,
+   "aurc": 0.15033545923257519,
+   "acc_at_80": 0.7361111111111112,
+   "acc_at_50": 0.8666666666666667,
+   "chance": 0.28444444444444444,
+   "mean_conf": 0.7939642071723938,
+   "sec": 1.1,
+   "heldout": false
+  },
+  "cx_chain_k": {
+   "n": 90,
+   "acc": 0.8666666666666667,
+   "nll": 0.38827842473983765,
+   "brier": 0.2077420949935913,
+   "ece": 0.09118105239338348,
+   "aurc": 0.0471507596042822,
+   "acc_at_80": 0.9027777777777778,
+   "acc_at_50": 0.9555555555555556,
+   "chance": 0.24222222222222228,
+   "mean_conf": 0.9095072150230408,
+   "sec": 1.1,
+   "heldout": false
+  },
+  "cx_cheapest_flip": {
+   "n": 90,
+   "acc": 0.5,
+   "nll": 1.1478240489959717,
+   "brier": 0.6096143126487732,
+   "ece": 0.18364082409275903,
+   "aurc": 0.34802452160197744,
+   "acc_at_80": 0.5555555555555556,
+   "acc_at_50": 0.6666666666666666,
+   "chance": 0.25,
+   "mean_conf": 0.598013699054718,
+   "sec": 1.1,
+   "heldout": false
+  },
+  "cx_constrained_max": {
+   "n": 90,
+   "acc": 0.8777777777777778,
+   "nll": 0.40832993388175964,
+   "brier": 0.2042766809463501,
+   "ece": 0.06301311055819196,
+   "aurc": 0.031290452245785895,
+   "acc_at_80": 0.9305555555555556,
+   "acc_at_50": 0.9777777777777777,
+   "chance": 0.25962962962962965,
+   "mean_conf": 0.9334561824798584,
+   "sec": 1.1,
+   "heldout": false
+  },
+  "cx_delegated_decider": {
+   "n": 90,
+   "acc": 0.8,
+   "nll": 0.6784549951553345,
+   "brier": 0.3062835931777954,
+   "ece": 0.12079860866069797,
+   "aurc": 0.10847367206647007,
+   "acc_at_80": 0.875,
+   "acc_at_50": 0.8888888888888888,
+   "chance": 0.3185185185185185,
+   "mean_conf": 0.9103758335113525,
+   "sec": 1.1,
+   "heldout": false
+  },
+  "cx_dominance_uncertainty": {
+   "n": 90,
+   "acc": 0.9111111111111111,
+   "nll": 0.33247512578964233,
+   "brier": 0.159723699092865,
+   "ece": 0.10246150261825984,
+   "aurc": 0.032359898129838346,
+   "acc_at_80": 0.9444444444444444,
+   "acc_at_50": 0.9777777777777777,
+   "chance": 0.3266666666666667,
+   "mean_conf": 0.9405836462974548,
+   "sec": 0.9,
+   "heldout": false
+  },
+  "cx_exception_ladder": {
+   "n": 90,
+   "acc": 0.9111111111111111,
+   "nll": 0.2511381506919861,
+   "brier": 0.131063312292099,
+   "ece": 0.04578693211078643,
+   "aurc": 0.01266272728198738,
+   "acc_at_80": 0.9861111111111112,
+   "acc_at_50": 1.0,
+   "chance": 0.2574074074074075,
+   "mean_conf": 0.940545380115509,
+   "sec": 1.2,
+   "heldout": false
+  },
+  "cx_fairness_vs_efficiency": {
+   "n": 90,
+   "acc": 0.7444444444444445,
+   "nll": 0.5721701383590698,
+   "brier": 0.3066578209400177,
+   "ece": 0.0937029391527176,
+   "aurc": 0.06570295226911144,
+   "acc_at_80": 0.8472222222222222,
+   "acc_at_50": 0.9777777777777777,
+   "chance": 0.3222222222222222,
+   "mean_conf": 0.7626772522926331,
+   "sec": 1.1,
+   "heldout": false
+  },
+  "cx_insufficient_information": {
+   "n": 90,
+   "acc": 0.4444444444444444,
+   "nll": 1.521159052848816,
+   "brier": 0.7202333211898804,
+   "ece": 0.2698589457405938,
+   "aurc": 0.45578351747410667,
+   "acc_at_80": 0.5,
+   "acc_at_50": 0.4666666666666667,
+   "chance": 0.16666666666666666,
+   "mean_conf": 0.6854457259178162,
+   "sec": 1.1,
+   "heldout": false
+  },
+  "cx_joint_threshold": {
+   "n": 90,
+   "acc": 0.6888888888888889,
+   "nll": 0.6687639355659485,
+   "brier": 0.37504687905311584,
+   "ece": 0.14556044969293808,
+   "aurc": 0.1120915555416827,
+   "acc_at_80": 0.75,
+   "acc_at_50": 0.9333333333333333,
+   "chance": 0.2651851851851852,
+   "mean_conf": 0.7760463953018188,
+   "sec": 1.1,
+   "heldout": false
+  },
+  "cx_knapsack_subset": {
+   "n": 90,
+   "acc": 0.6777777777777778,
+   "nll": 1.1246669292449951,
+   "brier": 0.45823127031326294,
+   "ece": 0.2184429340892368,
+   "aurc": 0.14787986568314593,
+   "acc_at_80": 0.8055555555555556,
+   "acc_at_50": 0.8444444444444444,
+   "chance": 0.11618606701940036,
+   "mean_conf": 0.8761476278305054,
+   "sec": 1.2,
+   "heldout": false
+  },
+  "cx_lex_specialis": {
+   "n": 90,
+   "acc": 0.8555555555555555,
+   "nll": 0.36309874057769775,
+   "brier": 0.2237442582845688,
+   "ece": 0.09633731577131484,
+   "aurc": 0.030433036668323914,
+   "acc_at_80": 0.9166666666666666,
+   "acc_at_50": 1.0,
+   "chance": 0.3296296296296296,
+   "mean_conf": 0.9435015916824341,
+   "sec": 1.1,
+   "heldout": false
+  },
+  "cx_lexicographic": {
+   "n": 90,
+   "acc": 0.6333333333333333,
+   "nll": 0.8402590155601501,
+   "brier": 0.47011181712150574,
+   "ece": 0.10198001033729977,
+   "aurc": 0.21453741589533734,
+   "acc_at_80": 0.7222222222222222,
+   "acc_at_50": 0.7555555555555555,
+   "chance": 0.267037037037037,
+   "mean_conf": 0.6630616784095764,
+   "sec": 1.1,
+   "heldout": false
+  },
+  "cx_opportunity_cost": {
+   "n": 90,
+   "acc": 1.0,
+   "nll": 3.3780670491978526e-05,
+   "brier": 2.659203346411232e-08,
+   "ece": 3.3736228942871094e-05,
+   "aurc": 0.0,
+   "acc_at_80": 1.0,
+   "acc_at_50": 1.0,
+   "chance": 0.3111111111111111,
+   "mean_conf": 0.9999662637710571,
+   "sec": 1.1,
+   "heldout": false
+  },
+  "cx_order_matters": {
+   "n": 90,
+   "acc": 0.8222222222222222,
+   "nll": 0.5555646419525146,
+   "brier": 0.2694486081600189,
+   "ece": 0.06730836530526482,
+   "aurc": 0.0479596063979094,
+   "acc_at_80": 0.8888888888888888,
+   "acc_at_50": 0.9777777777777777,
+   "chance": 0.3629629629629629,
+   "mean_conf": 0.8742075562477112,
+   "sec": 1.0,
+   "heldout": false
+  },
+  "cx_value_of_information": {
+   "n": 90,
+   "acc": 1.0,
+   "nll": 0.007142952643334866,
+   "brier": 0.004883978981524706,
+   "ece": 0.005321940448549059,
+   "aurc": 0.0,
+   "acc_at_80": 1.0,
+   "acc_at_50": 1.0,
+   "chance": 0.09999999999999998,
+   "mean_conf": 0.9946780800819397,
+   "sec": 1.2,
+   "heldout": false
+  },
+  "cx_weighted_aggregate": {
+   "n": 90,
+   "acc": 0.5111111111111111,
+   "nll": 1.056013584136963,
+   "brier": 0.5498751401901245,
+   "ece": 0.15534916122754414,
+   "aurc": 0.23210885309867516,
+   "acc_at_80": 0.6111111111111112,
+   "acc_at_50": 0.8,
+   "chance": 0.2822222222222222,
+   "mean_conf": 0.6122539639472961,
+   "sec": 1.1,
+   "heldout": false
+  }
+ },
+ "agg": {
+  "in_task": {
+   "acc": 0.7748538011695907,
+   "nll": 0.584524187147556,
+   "brier": 0.2983455650632406,
+   "ece": 0.10666718983162217,
+   "aurc": 0.10840233074066022,
+   "acc_at_80": 0.8391812865497076,
+   "chance": 0.26233337974566046
+  },
+  "heldout": {},
+  "n_in": 19,
+  "n_heldout": 0
+ },
+ "model": "runs/v12/model",
+ "engine": ""
+}

_laya.json

@@ -0,0 +1,143 @@
+{
+ "cx_authority_conflict": {
+  "n": 90,
+  "acc": 0.3333333333333333,
+  "chance": 0.3222222222222224,
+  "no_answer": 0,
+  "median_ms": 25.404397398233414
+ },
+ "cx_but_for_cause": {
+  "n": 90,
+  "acc": 0.4666666666666667,
+  "chance": 0.19999999999999965,
+  "no_answer": 0,
+  "median_ms": 25.657646358013153
+ },
+ "cx_cascade_fixpoint": {
+  "n": 90,
+  "acc": 0.36666666666666664,
+  "chance": 0.28444444444444433,
+  "no_answer": 0,
+  "median_ms": 25.29609203338623
+ },
+ "cx_chain_k": {
+  "n": 90,
+  "acc": 0.5,
+  "chance": 0.24222222222222203,
+  "no_answer": 0,
+  "median_ms": 25.47660283744335
+ },
+ "cx_cheapest_flip": {
+  "n": 90,
+  "acc": 0.25555555555555554,
+  "chance": 0.25,
+  "no_answer": 0,
+  "median_ms": 25.329384952783585
+ },
+ "cx_constrained_max": {
+  "n": 90,
+  "acc": 0.3111111111111111,
+  "chance": 0.2596296296296294,
+  "no_answer": 0,
+  "median_ms": 25.352757424116135
+ },
+ "cx_delegated_decider": {
+  "n": 90,
+  "acc": 0.37777777777777777,
+  "chance": 0.31851851851851865,
+  "no_answer": 0,
+  "median_ms": 25.237319990992546
+ },
+ "cx_dominance_uncertainty": {
+  "n": 90,
+  "acc": 0.32222222222222224,
+  "chance": 0.32666666666666644,
+  "no_answer": 0,
+  "median_ms": 25.127161294221878
+ },
+ "cx_exception_ladder": {
+  "n": 90,
+  "acc": 0.5222222222222223,
+  "chance": 0.2574074074074072,
+  "no_answer": 0,
+  "median_ms": 25.36040171980858
+ },
+ "cx_fairness_vs_efficiency": {
+  "n": 90,
+  "acc": 0.3111111111111111,
+  "chance": 0.32222222222222235,
+  "no_answer": 0,
+  "median_ms": 25.226369500160217
+ },
+ "cx_insufficient_information": {
+  "n": 90,
+  "acc": 0.16666666666666666,
+  "chance": 0.16666666666666646,
+  "no_answer": 0,
+  "median_ms": 25.363288819789886
+ },
+ "cx_joint_threshold": {
+  "n": 90,
+  "acc": 0.23333333333333334,
+  "chance": 0.26518518518518525,
+  "no_answer": 0,
+  "median_ms": 25.28115175664425
+ },
+ "cx_knapsack_subset": {
+  "n": 90,
+  "acc": 0.06666666666666667,
+  "chance": 0.08387565364345541,
+  "no_answer": 0,
+  "median_ms": 26.08116902410984
+ },
+ "cx_lex_specialis": {
+  "n": 90,
+  "acc": 0.37777777777777777,
+  "chance": 0.32962962962962983,
+  "no_answer": 0,
+  "median_ms": 25.25760792195797
+ },
+ "cx_lexicographic": {
+  "n": 90,
+  "acc": 0.35555555555555557,
+  "chance": 0.267037037037037,
+  "no_answer": 0,
+  "median_ms": 25.253349915146828
+ },
+ "cx_opportunity_cost": {
+  "n": 90,
+  "acc": 0.18888888888888888,
+  "chance": 0.3111111111111113,
+  "no_answer": 0,
+  "median_ms": 25.197844952344894
+ },
+ "cx_order_matters": {
+  "n": 90,
+  "acc": 0.4111111111111111,
+  "chance": 0.3629629629629631,
+  "no_answer": 0,
+  "median_ms": 25.146376341581345
+ },
+ "cx_value_of_information": {
+  "n": 90,
+  "acc": 0.03333333333333333,
+  "chance": 0.04999999999999991,
+  "no_answer": 0,
+  "median_ms": 26.408791542053223
+ },
+ "cx_weighted_aggregate": {
+  "n": 90,
+  "acc": 0.28888888888888886,
+  "chance": 0.2822222222222221,
+  "no_answer": 0,
+  "median_ms": 25.228703394532204
+ },
+ "__all": {
+  "n": 1710,
+  "acc": 0.30994152046783624,
+  "chance": 0.2580012527258738,
+  "no_answer": 0,
+  "median_ms": 25.318972766399384,
+  "p90_ms": 26.03990212082863
+ }
+}

_perfam.md

@@ -0,0 +1,21 @@
+| family | chance | opus-5 | gpt-5.6 | 1.1 | glm-5.3 | kimi-k3 | ds-flash | ds-pro | 1.0 | laya |
+|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|
+| `knapsack_subset` | 0.084 | 0.133 | 0.167 | 0.678 | 0.144 | 0.156 | 0.122 | 0.200 | 0.267 | 0.067 |
+| `cheapest_flip` | 0.250 | 0.700 | 0.600 | 0.500 | 0.400 | 0.367 | 0.522 | 0.289 | 0.156 | 0.256 |
+| `weighted_aggregate` | 0.282 | 0.644 | 0.689 | 0.511 | 0.367 | 0.333 | 0.367 | 0.244 | 0.356 | 0.289 |
+| `insufficient_information` | 0.167 | 0.833 | 0.833 | 0.444 | 0.689 | 0.444 | 0.489 | 0.422 | 0.189 | 0.167 |
+| `lexicographic` | 0.267 | 0.667 | 0.622 | 0.633 | 0.689 | 0.367 | 0.444 | 0.422 | 0.333 | 0.356 |
+| `joint_threshold` | 0.265 | 0.622 | 0.644 | 0.689 | 0.433 | 0.444 | 0.622 | 0.600 | 0.344 | 0.233 |
+| `dominance_uncertainty` | 0.327 | 0.844 | 0.833 | 0.911 | 0.533 | 0.433 | 0.333 | 0.244 | 0.489 | 0.322 |
+| `order_matters` | 0.363 | 0.989 | 0.989 | 0.822 | 0.533 | 0.322 | 0.356 | 0.311 | 0.456 | 0.411 |
+| `fairness_vs_efficiency` | 0.322 | 0.978 | 0.922 | 0.744 | 0.733 | 0.444 | 0.367 | 0.500 | 0.233 | 0.311 |
+| `cascade_fixpoint` | 0.284 | 0.978 | 0.889 | 0.678 | 0.611 | 0.478 | 0.467 | 0.400 | 0.367 | 0.367 |
+| `exception_ladder` | 0.257 | 0.589 | 0.600 | 0.911 | 0.600 | 0.589 | 0.578 | 0.456 | 0.500 | 0.522 |
+| `delegated_decider` | 0.319 | 0.933 | 0.833 | 0.800 | 0.889 | 0.456 | 0.411 | 0.389 | 0.356 | 0.378 |
+| `opportunity_cost` | 0.311 | 1.000 | 0.989 | 1.000 | 0.700 | 0.400 | 0.378 | 0.378 | 0.411 | 0.189 |
+| `lex_specialis` | 0.330 | 0.989 | 1.000 | 0.856 | 0.511 | 0.433 | 0.500 | 0.478 | 0.533 | 0.378 |
+| `constrained_max` | 0.260 | 1.000 | 1.000 | 0.878 | 0.956 | 0.667 | 0.689 | 0.633 | 0.511 | 0.311 |
+| `chain_k` | 0.242 | 0.956 | 0.933 | 0.867 | 0.789 | 0.789 | 0.711 | 0.667 | 0.500 | 0.500 |
+| `authority_conflict` | 0.322 | 0.989 | 0.956 | 0.856 | 0.967 | 0.822 | 0.700 | 0.633 | 0.611 | 0.333 |
+| `but_for_cause` | 0.200 | 1.000 | 1.000 | 0.944 | 1.000 | 0.967 | 0.711 | 0.733 | 0.411 | 0.467 |
+| `value_of_information` | 0.050 | 1.000 | 1.000 | 1.000 | 0.844 | 1.000 | 0.944 | 0.922 | 0.689 | 0.033 |

rebuild.py

@@ -0,0 +1,166 @@
+#!/usr/bin/env python3
+"""Rebuild the two figures and the per-family table from whatever has been measured.
+
+Kept on /mnt/c rather than /tmp: a reboot cleared the last copy of this and the figures it made,
+while the API cache that cost two thousand dollars survived because it was not in /tmp.
+"""
+import json, os, re, subprocess, sys
+import numpy as np
+import matplotlib
+matplotlib.use("Agg")
+import matplotlib.pyplot as plt
+
+D = "/mnt/c/code/jev/decider"
+OUT = "/mnt/c/code/jev/hf-dataset"
+INK, MUTED, GRID = "#1a1a1a", "#6b7280", "#e5e7eb"
+OURS, HOSTED, SMALL = "#1f4e79", "#9aa5b1", "#c25e00"
+CHANCE = 0.258
+
+
+def api_log(path, model):
+    per, cur = {}, None
+    if not os.path.exists(path):
+        return per
+    for line in open(path):
+        m = re.match(r"^#+ (\S+) #+", line.strip())
+        if m:
+            cur = m.group(1); continue
+        if cur != model:
+            continue
+        m = re.match(r"^\s+cx_(\S+)\s+acc ([0-9.]+)\s+\(chance ([0-9.]+)\)", line)
+        if m:
+            per[m.group(1)] = (float(m.group(2)), float(m.group(3)))
+    return per
+
+
+def local(path):
+    try:
+        r = json.load(open(path)).get("results", {})
+    except Exception:
+        return {}
+    return {k[3:]: (v["acc"], None) for k, v in r.items()
+            if k.startswith("cx_") and isinstance(v, dict) and "acc" in v}
+
+
+def laya(path):
+    try:
+        d = json.load(open(path))
+    except Exception:
+        return {}
+    return {k[3:]: (v["acc"], v.get("chance")) for k, v in d.items()
+            if k.startswith("cx_") and isinstance(v, dict) and "acc" in v}
+
+
+SYS = [
+    ("opus-5", "claude-opus-5", api_log("/tmp/cx_opus2.log", "claude-opus-5"), 1000, "hosted"),
+    ("gpt-5.6", "gpt-5.6", api_log("/tmp/cx_gpt.log", "gpt-5.6"), 1200, "hosted"),
+    ("1.1", "this-that-model-1.1", local(f"{OUT}/_cx_v11.json"), 30.9, "ours"),
+    ("glm-5.3", "glm-5.3", api_log("/tmp/cx_four.log", "glm-5.3"), 800, "hosted"),
+    ("kimi-k3", "kimi-k3", api_log("/tmp/cx_four.log", "kimi-k3"), 1000, "hosted"),
+    ("ds-flash", "deepseek-v4.1-flash", api_log("/tmp/cx_four.log", "deepseek-v4.1-flash"), 800, "hosted"),
+    ("ds-pro", "deepseek-v4-pro", api_log("/tmp/cx_four.log", "deepseek-v4-pro"), 900, "hosted"),
+    ("1.0", "this-that-model-1.0", local(f"{OUT}/_cx_10.json"), 30.9, "ours"),
+    ("laya", "laya-typed-decisions", laya(f"{OUT}/_laya.json"), 25, "small"),
+]
+
+overall = {}
+for short, full, per, ms, kind in SYS:
+    if per:
+        overall[short] = sum(v[0] for v in per.values()) / len(per)
+
+print(f"{'system':22s} {'overall':>8s}  families")
+for short, full, per, ms, kind in SYS:
+    a = f"{overall[short]:.3f}" if short in overall else "--"
+    print(f"{full:22s} {a:>8s}  {len(per)}/19")
+
+
+# ---------------------------------------------------------------- figure 1: accuracy vs latency
+fig, ax = plt.subplots(figsize=(8.6, 5.3), dpi=200)
+fig.patch.set_facecolor("white"); ax.set_facecolor("white")
+ax.axhspan(0, CHANCE, color="#f3f4f6", zorder=0)
+ax.axhline(CHANCE, color=MUTED, lw=1, ls=(0, (4, 3)), zorder=1)
+ax.text(5200, CHANCE - 0.013, "chance", color=MUTED, fontsize=9, ha="right", va="top")
+
+OFF = {"opus-5": (1.12, 0.016, "left"), "gpt-5.6": (1.12, -0.004, "left"),
+       "1.1": (1.22, 0.0, "left"), "glm-5.3": (0.88, 0.0, "right"),
+       "kimi-k3": (1.12, 0.014, "left"), "ds-flash": (0.88, -0.002, "right"),
+       "ds-pro": (1.12, -0.014, "left"), "1.0": (1.22, 0.0, "left"),
+       "laya": (1.25, 0.0, "left")}
+for short, full, per, ms, kind in SYS:
+    if short not in overall:
+        continue
+    c = {"ours": OURS, "hosted": HOSTED, "small": SMALL}[kind]
+    acc = overall[short]
+    ax.scatter(ms, acc, s=150 if kind != "hosted" else 90, color=c, zorder=3,
+               edgecolor="white", linewidth=1.6)
+    dx, dy, ha = OFF[short]
+    ax.annotate(full, (ms * dx, acc + dy), fontsize=9.5,
+                color=INK if kind != "hosted" else MUTED, ha=ha, va="center",
+                fontweight="bold" if kind == "ours" else "normal", zorder=4)
+
+ax.set_xscale("log"); ax.set_xlim(15, 6000); ax.set_ylim(0.20, 0.92)
+ax.set_xlabel("latency per decision  (ms, log scale)", fontsize=10.5, color=INK)
+ax.set_ylabel("accuracy", fontsize=10.5, color=INK)
+ax.set_title("1,710 complex decisions: what each answer costs in time",
+             fontsize=12.5, color=INK, pad=14, loc="left")
+ax.set_xticks([25, 50, 100, 250, 500, 1000, 2000])
+ax.set_xticklabels(["25", "50", "100", "250", "500", "1000", "2000"])
+ax.grid(axis="y", color=GRID, lw=0.8); ax.set_axisbelow(True)
+for s_ in ("top", "right"):
+    ax.spines[s_].set_visible(False)
+for s_ in ("left", "bottom"):
+    ax.spines[s_].set_color(GRID)
+ax.tick_params(colors=MUTED, labelsize=9.5)
+ax.annotate("", xy=(33, overall["1.1"]), xytext=(950, overall["opus-5"]),
+            arrowprops=dict(arrowstyle="-", color=GRID, lw=1.2, ls=(0, (3, 3))), zorder=1)
+ax.text(165, 0.807, f"{overall['opus-5']-overall['1.1']:.3f} accuracy, 32x the speed",
+        fontsize=9, color=MUTED, ha="center")
+fig.tight_layout(); fig.savefig(f"{OUT}/accuracy-vs-latency.png", facecolor="white")
+print("wrote accuracy-vs-latency.png")
+
+# ------------------------------------------------------------------ figure 2: per-family matrix
+cols = [s for s, _, p, _, _ in SYS if p]
+fams = sorted({f for _, _, p, _, _ in SYS for f in p})
+M = np.array([[SYS[[c[0] for c in SYS].index(c)][2].get(f, (np.nan,))[0] for c in cols]
+              for f in fams])
+order = np.argsort(np.nanmean(M, axis=1))
+M, fams = M[order], [fams[i] for i in order]
+
+fig, ax = plt.subplots(figsize=(10.4, 6.8), dpi=200)
+fig.patch.set_facecolor("white")
+cmap = matplotlib.colors.LinearSegmentedColormap.from_list(
+    "one_hue", ["#f7fbff", "#cfe0f0", "#8fb8db", "#4a89bd", "#1f4e79"])
+cmap.set_bad("#f0f0f0")
+im = ax.imshow(M, cmap=cmap, vmin=0.0, vmax=1.0, aspect="auto")
+ax.set_xticks(range(len(cols))); ax.set_xticklabels(cols, fontsize=9.5, color=INK)
+ax.set_yticks(range(len(fams)))
+ax.set_yticklabels([f.replace("_", " ") for f in fams], fontsize=9.5, color=INK)
+ax.tick_params(length=0)
+for s_ in ax.spines.values():
+    s_.set_visible(False)
+ax.set_xticks(np.arange(-.5, len(cols), 1), minor=True)
+ax.set_yticks(np.arange(-.5, len(fams), 1), minor=True)
+ax.grid(which="minor", color="white", lw=2); ax.tick_params(which="minor", length=0)
+for i in range(M.shape[0]):
+    for j in range(M.shape[1]):
+        v = M[i, j]
+        if not np.isnan(v) and (v >= 0.90 or v <= 0.30):
+            ax.text(j, i, f"{v:.2f}", ha="center", va="center", fontsize=8,
+                    color="white" if v >= 0.90 else MUTED)
+ax.set_title("accuracy by decision type, hardest first", fontsize=12.5, color=INK, pad=14, loc="left")
+cb = fig.colorbar(im, ax=ax, fraction=0.022, pad=0.015)
+cb.outline.set_visible(False)
+cb.ax.tick_params(colors=MUTED, labelsize=9, length=0)
+cb.set_label("accuracy", color=MUTED, fontsize=9.5)
+fig.tight_layout(); fig.savefig(f"{OUT}/by-decision-type.png", facecolor="white")
+print("wrote by-decision-type.png")
+
+# the markdown table, for the card
+with open(f"{OUT}/_perfam.md", "w") as fh:
+    fh.write("| family | chance | " + " | ".join(cols) + " |\n")
+    fh.write("|---|---:|" + "---:|" * len(cols) + "\n")
+    for i, f in enumerate(fams):
+        ch = next((p[f][1] for _, _, p, _, _ in SYS if f in p and p[f][1]), float("nan"))
+        row = " | ".join("--" if np.isnan(M[i, j]) else f"{M[i, j]:.3f}" for j in range(len(cols)))
+        fh.write(f"| `{f}` | {ch:.3f} | {row} |\n")
+print("wrote _perfam.md")