Edit-cost sensitivity and generalization

This experiment evaluates the normalized 234-point delete/insert/replace/match cost grid against deterministic 10%–100% dictionary-knowledge curves for five frozen splits and all 20 default models present in the checked campaign. Its primary generalization outcome is exact stemming of changed forms whose normalized surface was absent from training. Pairwise F0.5, over-stemming, and under-stemming are independent guards against a superficially good exact-root result.

The retained results are the frozen historical 20-language standard cohort. The current publication contract admits only a complete 31-model standard/default-package campaign; optional and filtered models remain outside this experiment.

A cost label uses D<delete>I<insert>R<replace>M<match>. D, I, and R are the relative costs of deleting a source character, inserting a target character, and replacing a source character. M is the cost of keeping an equal source/target character unchanged (the match or skip step). Thus D2I5R3M0 means costs 2, 5, 3, and 0 respectively; these numbers are edit-path costs, not counts of generated or retained commands.

Campaign identity

Protocol: radixor-cost-sensitivity-v4
Source identity: 4.2.0-6-g84e57fb
Git revision: 84e57fb27ae40913569e826858a8ebb07cf2ea01 (recorded pre-release source state)
Generator SHA-256: fc6aec8f08c3596a33e7f167ababfd2e583aed0671869bc4a208d75191ccc1df
Analysis script SHA-256: 36fb4d8fafa8f4daef11308c574f786d05112add47341c1f99cb58ddf7c1085c
Raw CSV SHA-256: 13ec2dfe0af799c3cf6a4cf9948903a02497f2e7de540627a507ba2950d7f738
Exact classes across languages: 334
Physically measured scenarios: 16,700
Expanded logical observations: 234,000

The baseline median unseen changed-form exactness across language, seed, and 10%–90% knowledge observations is 69.478%. This macro-style statement gives each scenario equal weight; it is not a token-weighted production estimate.

Key findings

The 4,680 language × normalized-cost combinations collapse to 334 exact dictionary-specific command classes, a 14.01× reduction. The range is 5 classes for Danish through 44 for German.
Baseline macro unseen changed-form exactness rises from 62.256% at 10% knowledge to 70.881% at 90%, a +8.625 pp gain.
The predeclared language-specific selections reduce median command count for 17/20 dictionaries. Across the macro knowledge curve the reduction is 5.01%–6.11% while the largest absolute exactness change is 0.003 pp. The largest language-level median reduction is 26.64% for French.
Retained-command ratio has a stable positive association with trie nodes and value references, so it is a useful structural-size proxy. No cost or representation predictor keeps one association sign across the central 95% of strata for any unseen-form quality outcome; quality effects are language- and knowledge-dependent.
Exact-class count is more associated with baseline command vocabulary (Spearman +0.510) than with raw dictionary rows (+0.260). This supports command-aware, not merely size-based, resource classification.

Dictionary sensitivity census

Common positive scaling has already been removed from the 234-point normalized grid. The remaining points collapse only when their generated command is identical for every full-dictionary pair, verified command by command after fingerprint bucketing. More classes therefore mean that the dictionary exposes more cost-dependent encoder decisions; this is sensitivity, not automatically better quality.

Language	Rows	Mean family size	Changed forms	Baseline commands	Exact classes	Grid reduction	Largest class
Danish	4,179	7.72	74.09%	138	5	46.80×	110
Yiddish	802	5.36	64.56%	202	21	11.14×	54
Ukrainian	1,493	10.54	81.03%	286	7	33.43×	76
Norwegian Nynorsk	4,688	4.19	69.01%	299	11	21.27×	77
Norwegian Bokmål	17,929	5.06	63.22%	325	14	16.71×	54
Swedish	12,371	8.93	77.61%	332	10	23.40×	59
English	396,939	2.53	20.80%	337	15	15.60×	54
Portuguese	4,001	53.86	96.29%	369	9	26.00×	77
Hungarian	19,406	48.22	95.86%	462	12	19.50×	54
Czech	5,113	11.07	82.25%	537	12	19.50×	54
Dutch	4,992	6.30	68.28%	538	27	8.67×	54
Persian	69	54.64	96.34%	704	18	13.00×	54
Italian	10,009	33.72	94.07%	750	19	12.32×	54
Polish	9,990	13.24	84.92%	846	19	12.32×	54
Spanish	65,059	14.24	87.03%	1,496	18	13.00×	54
Russian	37,410	21.55	90.72%	1,840	21	11.14×	54
Finnish	57,027	32.71	94.07%	2,683	7	33.43×	110
French	59,240	8.00	77.19%	2,730	32	7.31×	54
Hebrew	2,358	25.90	92.28%	4,150	13	18.00×	59
German	54,092	6.16	72.82%	6,986	44	5.32×	54

Dictionary-level associations

The following across-language coefficients describe whether dictionaries with a given property expose more distinct cost-dependent command sequences. There are only 20 language observations, so these are exploratory effect descriptions, not causal estimates or evidence that unlisted properties are irrelevant.

Dictionary property	Outcome	Spearman ρ	Pearson r	Languages
`dictionary_rows`	`exact_equivalence_classes`	+0.260	+0.072	20
`dictionary_forms`	`exact_equivalence_classes`	+0.088	-0.063	20
`mean_family_size`	`exact_equivalence_classes`	-0.176	-0.241	20
`changed_form_share`	`exact_equivalence_classes`	-0.231	-0.132	20
`baseline_patch_commands`	`exact_equivalence_classes`	+0.510	+0.639	20

Language-specific observed optima

Here, optimal always means optimal for the stated objective inside the measured normalized 234-point grid. It does not mean that the same setting is optimal for an unmeasured corpus, a different dictionary, or a cost ratio outside the grid.

Smallest full-dictionary patch-command vocabulary

This table minimizes the number of distinct patch commands retained by the compiled trie at 100% dictionary knowledge. All five seeds compile the same complete dictionary and must produce the same count. When several cost configurations reach that minimum, the displayed representative is chosen by higher median unseen F0.5 over the 10%–90% observations, then lower median over- and under-stemming; Tied minima records how many grid configurations reach the same minimum count. The exactness delta and F0.5 columns describe the displayed representative on partial-knowledge generalization observations and expose the quality consequence of pursuing structural size alone.

Language	Minimum-command costs	Commands	Baseline	Reduction	Unseen exact Δ	Unseen F0.5	Tied minima
Czech	`D1I1R1M1`	509	537	5.21%	+0.000 pp	0.927175	25
Danish	`D1I1R2M0`	123	138	10.87%	+0.000 pp	0.878150	54
German	`D2I5R3M0`	6,784	6,986	2.89%	+0.000 pp	0.825677	2
Spanish	`D1I1R1M1`	1,283	1,496	14.24%	+0.000 pp	0.962789	25
Persian	`D1I5R3M1`	660	704	6.25%	+0.000 pp	0.117096	15
Finnish	`D1I1R1M1`	2,033	2,683	24.23%	+0.005 pp	0.945461	25
French	`D1I1R1M1`	2,073	2,730	24.07%	+0.000 pp	0.915559	25
Hebrew	`D2I3R3M0`	4,023	4,150	3.06%	+0.022 pp	0.287004	8
Hungarian	`D1I1R2M0`	411	462	11.04%	+0.000 pp	0.970327	54
Italian	`D1I1R1M1`	594	750	20.80%	+0.000 pp	0.971928	25
Norwegian Bokmål	`D1I1R2M0`	313	325	3.69%	-0.004 pp	0.872959	54
Dutch	`D1I2R2M0`	502	538	6.69%	+0.005 pp	0.840309	28
Norwegian Nynorsk	`D1I3R3M0`	299	299	0.00%	+0.000 pp	0.874458	45
Polish	`D1I1R1M1`	791	846	6.50%	+0.000 pp	0.893496	25
Portuguese	`D1I1R1M1`	305	369	17.34%	+0.000 pp	0.975633	25
Russian	`D1I1R1M0`	1,840	1,840	0.00%	+0.000 pp	0.946441	9
Swedish	`D1I1R1M1`	317	332	4.52%	+0.000 pp	0.906986	25
Ukrainian	`D1I1R1M1`	272	286	4.90%	+0.000 pp	0.833220	25
English	`D1I1R1M0`	337	337	0.00%	+0.000 pp	0.896140	45
Yiddish	`D1I1R2M0`	201	202	0.50%	+0.000 pp	0.771991	90

Machine-readable values: minimum-command configurations.

Best observed pairwise generalization quality on unseen forms

This is not full-model production quality

Unseen-family OI and Unseen-family UI are medians over 45 deliberately difficult generalization scenarios: five frozen splits at each 10%–90% dictionary-knowledge level. Evaluation includes only forms from withheld dictionary families whose normalized surface never appeared in training. The values therefore measure how a partial trie connects previously unseen members of previously unseen families; they do not describe a model built from the complete dictionary.

Under-stemming is pairwise: UI = FN / (TP + FN), where every gold-related pair of forms is one positive relation. One incorrectly separated form can break several relations inside its family, so unseen-family UI is neither a per-form error rate nor the complement of exact-root accuracy. At 100% knowledge no withheld unseen-family population remains. The Full-model UI column therefore supplies a separate reference measured over the complete dictionary with the displayed cost configuration; it is the relevant column for judging full-model behavior.

Only configurations classified VIABLE in every partial-knowledge language × seed observation are eligible. A candidate must not increase either median unseen-family OI or median unseen-family UI relative to D1I1R1M0. Among those candidates, the table maximizes median unseen F0.5; ties prefer lower OI, then lower UI, and finally the production baseline. Lower OI/UI values are better; higher F0.5 is better. Deltas are against the baseline and retain six decimals because several measured effects are small.

A non-baseline configuration is selected for 7/20 dictionaries. This establishes that edit costs can alter measured stemming quality for particular dictionaries, but the magnitude and direction remain dictionary-dependent. Selection and reporting use the same matrix, so these are exploratory observed optima rather than externally validated production defaults.

Language	Quality costs	Median unseen-family OI (10%–90%)	Δ OI	Median unseen-family UI (10%–90%)	Δ UI	Full-model UI (100%)	Unseen F0.5	Δ F0.5
Czech	`D1I1R1M0`	0.000017%	+0.000000 pp	28.122302%	+0.000000 pp	0.566807%	0.927175	+0.000000
Danish	`D1I1R1M0`	0.000061%	+0.000000 pp	40.837995%	+0.000000 pp	0.703039%	0.878150	+0.000000
German	`D1I5R3M1`	0.000063%	-0.000000 pp	49.497381%	-0.001356 pp	12.031009%	0.825677	+0.000019
Spanish	`D1I1R1M0`	0.000008%	+0.000000 pp	16.101530%	+0.000000 pp	2.110205%	0.962789	+0.000000
Persian	`D1I1R1M0`	0.000000%	+0.000000 pp	97.415054%	+0.000000 pp	4.877973%	0.117096	+0.000000
Finnish	`D1I2R1M0`	0.000001%	+0.000000 pp	22.310391%	-0.000007 pp	2.355085%	0.945434	+0.000000
French	`D1I1R1M0`	0.000011%	+0.000000 pp	31.351496%	+0.000000 pp	8.248915%	0.915559	+0.000000
Hebrew	`D1I2R2M0`	0.000079%	-0.000012 pp	92.512100%	-0.042585 pp	2.749465%	0.287004	+0.001308
Hungarian	`D1I1R1M1`	0.000010%	+0.000000 pp	12.955456%	-0.000055 pp	0.736116%	0.970327	+0.000000
Italian	`D1I1R1M0`	0.000001%	+0.000000 pp	12.613898%	+0.000000 pp	0.669175%	0.971928	+0.000000
Norwegian Bokmål	`D1I1R1M0`	0.000024%	+0.000000 pp	41.628315%	+0.000000 pp	4.749500%	0.872959	+0.000000
Dutch	`D1I1R2M0`	0.000027%	+0.000000 pp	48.375349%	-0.015697 pp	2.129575%	0.840309	+0.000084
Norwegian Nynorsk	`D1I1R2M0`	0.000096%	-0.000001 pp	41.218142%	+0.000000 pp	9.675485%	0.874458	+0.000000
Polish	`D1I1R1M0`	0.000010%	+0.000000 pp	37.062231%	+0.000000 pp	1.730587%	0.893496	+0.000000
Portuguese	`D1I1R1M0`	0.000008%	+0.000000 pp	11.069494%	+0.000000 pp	0.291615%	0.975633	+0.000000
Russian	`D1I1R1M0`	0.000018%	+0.000000 pp	21.536426%	+0.000000 pp	1.957425%	0.946441	+0.000000
Swedish	`D1I1R1M0`	0.000034%	+0.000000 pp	33.331485%	+0.000000 pp	4.476263%	0.906986	+0.000000
Ukrainian	`D1I1R1M0`	0.000037%	+0.000000 pp	49.894697%	+0.000000 pp	0.836852%	0.833220	+0.000000
English	`D1I1R1M0`	0.000001%	+0.000000 pp	35.620602%	+0.000000 pp	4.615189%	0.896140	+0.000000
Yiddish	`D1I1R2M0`	0.000045%	-0.000014 pp	59.609698%	+0.000000 pp	2.184236%	0.771991	+0.000000

Machine-readable values: quality configurations.

Predeclared recommendation rule

For each language, configurations classified VIABLE at every partial-knowledge observation are retained. The rule finds the best median unseen changed-form exactness, admits settings within 0.25 percentage points, then minimizes the median retained-command ratio; remaining ties prefer unseen F0.5 and exactness. This is an exploratory operating-point selection, not an external-test estimate.

Language	Recommended costs	Unseen changed exact	Δ vs baseline	Unseen F0.5	Command ratio
Czech	`D10I10R1M1`	70.455%	+0.000 pp	0.9272	0.942×
Danish	`D1I1R10M0`	64.067%	+0.000 pp	0.8781	0.895×
German	`D2I5R3M0`	55.374%	+0.000 pp	0.8257	0.971×
Spanish	`D10I10R1M1`	87.056%	+0.000 pp	0.9628	0.820×
Persian	`D10I10R1M1`	14.906%	-0.031 pp	0.1165	0.917×
Finnish	`D10I10R1M1`	84.143%	+0.005 pp	0.9455	0.736×
French	`D10I10R1M1`	76.846%	+0.000 pp	0.9156	0.734×
Hebrew	`D10I5R10M1`	14.245%	+0.022 pp	0.2870	0.975×
Hungarian	`D1I1R10M0`	90.699%	+0.000 pp	0.9703	0.888×
Italian	`D10I10R1M1`	91.567%	+0.000 pp	0.9719	0.806×
Norwegian Bokmål	`D1I1R10M0`	67.144%	-0.004 pp	0.8730	0.965×
Dutch	`D10I1R10M0`	63.882%	+0.005 pp	0.8403	0.961×
Norwegian Nynorsk	`D1I1R1M0`	63.758%	+0.000 pp	0.8745	1.000×
Polish	`D10I10R1M1`	67.725%	+0.000 pp	0.8935	0.941×
Portuguese	`D10I10R1M1`	93.092%	+0.000 pp	0.9756	0.812×
Russian	`D10I10R1M1`	79.655%	-0.004 pp	0.9464	0.970×
Swedish	`D10I10R1M1`	75.990%	+0.000 pp	0.9070	0.953×
Ukrainian	`D10I10R1M1`	52.642%	+0.000 pp	0.8332	0.943×
English	`D1I1R1M0`	77.779%	+0.000 pp	0.8961	1.000×
Yiddish	`D1I1R1M0`	45.515%	+0.000 pp	0.7720	1.000×

Per-language conclusions

The macro result is not substituted for language evidence. Each language page now publishes its five-split knowledge curve, dictionary-specific exact cost classes, within-language association coverage, and a bounded conclusion. The compact index below exposes the main measured differences; follow the link for the supporting rows and interpretation limits.

Language	Baseline exact at 10%	Baseline exact at 90%	Knowledge effect	Selected costs	Median command change	Conclusion
Czech	64.331%	70.970%	+6.639 pp	`D10I10R1M1`	-5.83%	Evidence and conclusion
Danish	51.619%	68.000%	+16.381 pp	`D1I1R10M0`	-10.47%	Evidence and conclusion
German	48.338%	58.463%	+10.125 pp	`D2I5R3M0`	-2.94%	Evidence and conclusion
Spanish	80.556%	89.280%	+8.725 pp	`D10I10R1M1`	-18.02%	Evidence and conclusion
Persian	14.864%	16.975%	+2.112 pp	`D10I10R1M1`	-8.31%	Evidence and conclusion
Finnish	78.695%	85.707%	+7.012 pp	`D10I10R1M1`	-26.37%	Evidence and conclusion
French	69.937%	80.058%	+10.121 pp	`D10I10R1M1`	-26.64%	Evidence and conclusion
Hebrew	10.977%	15.910%	+4.933 pp	`D10I5R10M1`	-2.48%	Evidence and conclusion
Hungarian	86.780%	91.867%	+5.087 pp	`D1I1R10M0`	-11.17%	Evidence and conclusion
Italian	85.934%	93.330%	+7.397 pp	`D10I10R1M1`	-19.44%	Evidence and conclusion
Norwegian Bokmål	58.507%	70.574%	+12.068 pp	`D1I1R10M0`	-3.54%	Evidence and conclusion
Dutch	55.658%	68.281%	+12.623 pp	`D10I1R10M0`	-3.86%	Evidence and conclusion
Norwegian Nynorsk	50.756%	69.473%	+18.717 pp	`D1I1R1M0`	+0.00%	Evidence and conclusion
Polish	61.850%	70.138%	+8.288 pp	`D10I10R1M1`	-5.88%	Evidence and conclusion
Portuguese	89.204%	94.378%	+5.174 pp	`D10I10R1M1`	-18.75%	Evidence and conclusion
Russian	74.976%	81.082%	+6.106 pp	`D10I10R1M1`	-3.04%	Evidence and conclusion
Swedish	67.345%	79.922%	+12.577 pp	`D10I10R1M1`	-4.72%	Evidence and conclusion
Ukrainian	43.396%	56.474%	+13.078 pp	`D10I10R1M1`	-5.67%	Evidence and conclusion
English	75.824%	79.131%	+3.308 pp	`D1I1R1M0`	+0.00%	Evidence and conclusion
Yiddish	38.579%	50.000%	+11.421 pp	`D1I1R1M0`	+0.00%	Evidence and conclusion

Generalization across knowledge levels

Each cell is the median across 20 languages × five frozen splits at the stated knowledge level. The recommended curve applies each language's independently selected setting; it is not a single global configuration. Percentages remain macro summaries rather than token-weighted production estimates.

Knowledge	Baseline unseen changed exact	Recommended	Δ	Baseline F0.5	Recommended F0.5	Baseline commands	Recommended commands
10%	62.256%	62.256%	+0.000 pp	0.8638	0.8638	1.000×	0.939×
20%	65.946%	65.946%	+0.000 pp	0.8833	0.8833	1.000×	0.939×
30%	67.259%	67.259%	+0.000 pp	0.8881	0.8881	1.000×	0.946×
40%	68.376%	68.376%	+0.000 pp	0.8927	0.8927	1.000×	0.948×
50%	69.309%	69.309%	+0.000 pp	0.8961	0.8961	1.000×	0.947×
60%	70.014%	70.014%	+0.000 pp	0.8984	0.8984	1.000×	0.941×
70%	70.186%	70.183%	-0.003 pp	0.9021	0.9021	1.000×	0.942×
80%	70.691%	70.691%	+0.000 pp	0.9058	0.9058	1.000×	0.948×
90%	70.881%	70.881%	+0.000 pp	0.9066	0.9066	1.000×	0.950×

Dictionary properties and selected costs

These across-language coefficients connect dictionary descriptors to the exploratory language-specific selections above. With 20 dictionaries and selection on the same data, they generate hypotheses about dictionary types; they do not establish a portable cost-selection rule for an unseen resource. Predictor and selected-outcome labels are defined in the methodology glossary.

Dictionary property	Selected outcome	Spearman ρ	Pearson r	Languages
`mean_family_size`	`recommended_match_cost`	+0.673	+0.393	20
`changed_form_share`	`recommended_match_cost`	+0.673	+0.620	20
`mean_family_size`	`recommended_command_ratio`	-0.636	-0.434	20
`changed_form_share`	`recommended_insert_cost`	+0.623	+0.590	20
`changed_form_share`	`recommended_command_ratio`	-0.616	-0.487	20
`mean_family_size`	`recommended_insert_cost`	+0.615	+0.352	20
`mean_family_size`	`recommended_delete_cost`	+0.598	+0.321	20
`changed_form_share`	`recommended_delete_cost`	+0.566	+0.575	20

Associations across cost configurations

Correlations are calculated separately inside every language × seed × knowledge stratum; the table reports the median and central 95% empirical interval of those within-stratum coefficients. This avoids allowing language size or knowledge level to manufacture a pooled association. The interval is descriptive, not an independence-adjusted confidence interval. Every predictor and outcome label is defined in the methodology glossary. Only the following pairs keep the same sign throughout that central interval.

Predictor	Outcome	Median Spearman ρ	Central 95%	Median Pearson r	Strata
`patch_command_ratio`	`value_references`	+0.988	+0.263…+1.000	+0.995	900
`patch_command_ratio`	`trie_nodes`	+0.953	+0.187…+1.000	+0.970	900
`edit_cost_imbalance`	`logical_leaf_paths`	+0.177	+0.079…+0.293	+0.137	540

Quality associations are heterogeneous

No cost or representation predictor keeps one Spearman sign across the central 95% of strata for any unseen-form quality outcome. The largest absolute median for each outcome is shown below to make that heterogeneity visible; it must not be interpreted as a global effect that transfers between languages or knowledge levels.

Predictor	Outcome	Median Spearman ρ	Central 95%	Median Pearson r	Strata
`logical_leaf_paths`	`unseen_changed_exact`	-1.000	-1.000…+0.126	-1.000	502
`logical_leaf_paths`	`unseen_f05`	-0.956	-1.000…+0.331	-1.000	517
`replace_to_delete_insert`	`unseen_over_percent`	-0.766	-0.947…+0.712	-0.570	325
`logical_leaf_paths`	`unseen_under_percent`	+1.000	-0.158…+1.000	+1.000	517

Smallest observed association envelopes

These pairs have the smallest central-interval envelope in the measured matrix. They are candidates for practical insensitivity, not proof of independence; an interval that spans substantial positive and negative values instead indicates heterogeneity.

Predictor	Outcome	Median Spearman ρ	Central 95%	Median Pearson r	Strata
`delete_to_insert_ratio`	`unseen_changed_exact`	+0.000	+0.000…+0.000	+0.121	535
`delete_to_insert_ratio`	`unseen_f05`	+0.000	+0.000…+0.000	+0.118	625
`delete_to_insert_ratio`	`unseen_over_percent`	+0.000	+0.000…+0.000	+0.051	325
`delete_to_insert_ratio`	`unseen_under_percent`	+0.000	+0.000…+0.000	-0.121	602
`delete_to_insert_ratio`	`patch_command_ratio`	+0.000	+0.000…+0.000	+0.006	900
`delete_to_insert_ratio`	`trie_nodes`	+0.000	+0.000…+0.000	+0.007	900
`delete_to_insert_ratio`	`trie_edges`	+0.000	+0.000…+0.000	-0.001	797
`delete_to_insert_ratio`	`average_path_length`	+0.000	+0.000…+0.000	-0.051	540

Interpretation boundaries

The experiment establishes sensitivity and within-resource generalization, not causality or performance on unrelated domains. A weak median correlation is not described as proof of no relationship: heterogeneous signs, nonlinear effects, and repeated splits remain possible. Recommendations therefore remain language-specific until validated on an external dictionary or corpus. See the full protocol and raw-data instructions.