@mailwoman/corpus 6.0.0 → 6.2.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
@@ -7,7 +7,7 @@
7
7
  * surpass-v1.5.0). Teaches the model to SPLIT the département out of the locality on
8
8
  * bare/space/comma-delimited French place rows — the admin-deciding failure class the pre-GPU
9
9
  * self-validation proved moves the resolved coordinate (collision communes −61%; see
10
- * docs/articles/evals/2026-06-19-fr-admin-split-prevalidation.md). Ported from
10
+ * docs/articles/evals/experiments/2026-06-19-fr-admin-split-prevalidation.md). Ported from
11
11
  * scripts/build-fr-admin-split-shard.mjs.
12
12
  *
13
13
  * Failure shapes (the model currently mis-handles all of these):
@@ -7,7 +7,7 @@
7
7
  * surpass-v1.5.0). Teaches the model to SPLIT the département out of the locality on
8
8
  * bare/space/comma-delimited French place rows — the admin-deciding failure class the pre-GPU
9
9
  * self-validation proved moves the resolved coordinate (collision communes −61%; see
10
- * docs/articles/evals/2026-06-19-fr-admin-split-prevalidation.md). Ported from
10
+ * docs/articles/evals/experiments/2026-06-19-fr-admin-split-prevalidation.md). Ported from
11
11
  * scripts/build-fr-admin-split-shard.mjs.
12
12
  *
13
13
  * Failure shapes (the model currently mis-handles all of these):
@@ -12,7 +12,7 @@
12
12
  *
13
13
  * ORDER ROBUSTNESS (2026-06-06): mixing the two renderings stops a native-only shard from teaching
14
14
  * German order so well it reads the US/feed-order eval as a "collapse". See
15
- * docs/articles/evals/2026-06-06-anchor-pilot.md (the order-artifact correction).
15
+ * docs/articles/evals/resolver-geo/2026-06-06-anchor-pilot.md (the order-artifact correction).
16
16
  */
17
17
  import { type ShardRecipe } from "./scaffold.ts";
18
18
  export declare const germanRecipe: ShardRecipe;
@@ -12,7 +12,7 @@
12
12
  *
13
13
  * ORDER ROBUSTNESS (2026-06-06): mixing the two renderings stops a native-only shard from teaching
14
14
  * German order so well it reads the US/feed-order eval as a "collapse". See
15
- * docs/articles/evals/2026-06-06-anchor-pilot.md (the order-artifact correction).
15
+ * docs/articles/evals/resolver-geo/2026-06-06-anchor-pilot.md (the order-artifact correction).
16
16
  */
17
17
  import { spawnSync } from "node:child_process";
18
18
  import { stableSourceID } from "../adapter.js";
@@ -46,8 +46,8 @@ export interface LocaleSynthesisOpts {
46
46
  * house-AFTER-street, postcode-BEFORE-city). `"international"` renders house-FIRST, postcode-AFTER-city — the US/GB
47
47
  * layout that international feeds, US-centric systems, and our own OpenAddresses de-sample impose on non-US
48
48
  * addresses. Training both teaches the model that a German address can arrive either way, so the eval's US-order
49
- * rendering stops reading as a collapse. See `docs/articles/evals/2026-06-06-anchor-pilot.md` (the order-artifact
50
- * correction).
49
+ * rendering stops reading as a collapse. See `docs/articles/evals/resolver-geo/2026-06-06-anchor-pilot.md` (the
50
+ * order-artifact correction).
51
51
  */
52
52
  order?: "native" | "international";
53
53
  /**
@@ -9,8 +9,8 @@
9
9
  * region, postcode, country}.
10
10
  *
11
11
  * Rationale: the [2026-05-28 night-2
12
- * postmortem](../../docs/articles/evals/2026-05-28-night-2-postmortem.md) and the [layer-1
13
- * eval](../../docs/articles/evals/2026-05-28-layer-1-morphology-fst.md) showed that synth-street
12
+ * postmortem](../../docs/articles/evals/night-shifts/2026-05-28-night-2-postmortem.md) and the [layer-1
13
+ * eval](../../docs/articles/evals/experiments/2026-05-28-layer-1-morphology-fst.md) showed that synth-street
14
14
  * pushed the model into a high-confidence "decompose mode" that leaked into `dependent_locality`.
15
15
  * Per DeepSeek's turn-2 recipe, the model needs explicit counter-examples: addresses where the
16
16
  * model should NOT emit street labels. This synthesizer is that source.
@@ -9,8 +9,8 @@
9
9
  * region, postcode, country}.
10
10
  *
11
11
  * Rationale: the [2026-05-28 night-2
12
- * postmortem](../../docs/articles/evals/2026-05-28-night-2-postmortem.md) and the [layer-1
13
- * eval](../../docs/articles/evals/2026-05-28-layer-1-morphology-fst.md) showed that synth-street
12
+ * postmortem](../../docs/articles/evals/night-shifts/2026-05-28-night-2-postmortem.md) and the [layer-1
13
+ * eval](../../docs/articles/evals/experiments/2026-05-28-layer-1-morphology-fst.md) showed that synth-street
14
14
  * pushed the model into a high-confidence "decompose mode" that leaked into `dependent_locality`.
15
15
  * Per DeepSeek's turn-2 recipe, the model needs explicit counter-examples: addresses where the
16
16
  * model should NOT emit street labels. This synthesizer is that source.
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "@mailwoman/corpus",
3
- "version": "6.0.0",
3
+ "version": "6.2.0",
4
4
  "description": "Mailwoman corpus pipeline: BIO-labeled dataset builder for the neural classifier.",
5
5
  "license": "AGPL-3.0-only OR LicenseRef-Commercial",
6
6
  "repository": {
@@ -82,9 +82,9 @@
82
82
  },
83
83
  "dependencies": {
84
84
  "@dsnp/parquetjs": "1.8.8",
85
- "@mailwoman/codex": "6.0.0",
86
- "@mailwoman/core": "6.0.0",
87
- "@mailwoman/formatter": "6.0.0",
85
+ "@mailwoman/codex": "6.2.0",
86
+ "@mailwoman/core": "6.2.0",
87
+ "@mailwoman/formatter": "6.2.0",
88
88
  "csv-parse": "^7.0.1",
89
89
  "fast-glob": "^3.3.3",
90
90
  "fastest-levenshtein": "^1.0.16",