@gaunt-sloth/batch 2.0.0-alpha.26 → 2.0.0-alpha.27
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.
- package/dist/evalTypes.d.ts +15 -14
- package/dist/evalTypes.js +14 -19
- package/dist/evalTypes.js.map +1 -1
- package/dist/raterTarget.d.ts +1 -1
- package/dist/raterTarget.js +29 -21
- package/dist/raterTarget.js.map +1 -1
- package/package.json +3 -3
package/dist/evalTypes.d.ts
CHANGED
|
@@ -46,7 +46,7 @@ export declare const HARDLINE_REFUSAL_MARKER = "hardline floor: refused";
|
|
|
46
46
|
* What DOES discriminate is **which deterministic mechanism fired**, which the target reports in the
|
|
47
47
|
* rationale, and which is exactly what the corpus records. So a deterministic case asserts that.
|
|
48
48
|
*/
|
|
49
|
-
export declare const FORCED_BY_MECHANISMS: readonly ['hardline-floor', '
|
|
49
|
+
export declare const FORCED_BY_MECHANISMS: readonly ['hardline-floor', 'script-env-leak-preflight'];
|
|
50
50
|
/** One of {@link FORCED_BY_MECHANISMS}. */
|
|
51
51
|
export type ForcedByMechanism = (typeof FORCED_BY_MECHANISMS)[number];
|
|
52
52
|
/** The prefix a rationale carries for a decision a preflight forced, followed by the mechanism. */
|
|
@@ -58,8 +58,8 @@ export declare const FORCED_BY_MARKER = "forced by";
|
|
|
58
58
|
*/
|
|
59
59
|
export declare const FORCED_BY_ASSERTIONS: Record<ForcedByMechanism, string>;
|
|
60
60
|
/**
|
|
61
|
-
* Which of {@link FORCED_BY_MECHANISMS} are **
|
|
62
|
-
* can only be observed by handing the gate a rating for them to OVERRIDE.
|
|
61
|
+
* Which of {@link FORCED_BY_MECHANISMS} are **preflight FINDINGS of the decision mapping** — the
|
|
62
|
+
* ones that can only be observed by handing the gate a rating for them to OVERRIDE.
|
|
63
63
|
*
|
|
64
64
|
* **This is the single statement of the rule, and everything else reads it through
|
|
65
65
|
* {@link mechanismNeedsPermissiveRating}.** It is kept here, next to the mechanism list, rather than
|
|
@@ -69,7 +69,7 @@ export declare const FORCED_BY_ASSERTIONS: Record<ForcedByMechanism, string>;
|
|
|
69
69
|
*
|
|
70
70
|
* ## Why the floor is not on this list (measured, CFG-28)
|
|
71
71
|
*
|
|
72
|
-
*
|
|
72
|
+
* A preflight finding lives INSIDE `mapVerdictToAction`, and CFG-28 narrowed it to raise only an
|
|
73
73
|
* outcome that sits below the deterministic floor. Give that mapping no verdict and it substitutes
|
|
74
74
|
* core's fail-closed one, which is *at* the floor — so no preflight rewrites anything, every command
|
|
75
75
|
* comes back with the identical placeholder sentence, and the assertion is a constant again. Handing
|
|
@@ -82,17 +82,16 @@ export declare const FORCED_BY_ASSERTIONS: Record<ForcedByMechanism, string>;
|
|
|
82
82
|
* verdict — so a rating cannot change the answer, and supplying one buys nothing. It also costs:
|
|
83
83
|
* `mapVerdictToAction` does not consult the floor, so a permissive rating on `rm -rf /` returns
|
|
84
84
|
* `approve`, which would move the action column of **5 of the corpus's 6 `fl-*` cases** off the
|
|
85
|
-
* `escalate` they author. Measured over all 22 deterministic corpus cases
|
|
86
|
-
* section of `handoff/task-1-report.md`.
|
|
85
|
+
* `escalate` they author. Measured over all 22 deterministic corpus cases.
|
|
87
86
|
*
|
|
88
87
|
* ## The property that makes this self-justifying
|
|
89
88
|
*
|
|
90
|
-
* When the claimed preflight really does fire, the rating does not move the action —
|
|
91
|
-
*
|
|
92
|
-
*
|
|
93
|
-
*
|
|
89
|
+
* When the claimed preflight really does fire, the rating does not move the action — the corpus
|
|
90
|
+
* cases that claim one (`el-01`, `el-02`) are `escalate` with or without it. When it does not fire,
|
|
91
|
+
* the marker and the action go red **together**. That is the discrimination, not a side effect of
|
|
92
|
+
* it.
|
|
94
93
|
*/
|
|
95
|
-
export declare const PREFLIGHT_MECHANISMS: readonly ["
|
|
94
|
+
export declare const PREFLIGHT_MECHANISMS: readonly ["script-env-leak-preflight"];
|
|
96
95
|
/**
|
|
97
96
|
* One of {@link PREFLIGHT_MECHANISMS}. Narrower than {@link ForcedByMechanism} on purpose: the
|
|
98
97
|
* target's probe table is typed as a TOTAL record over this, so adding a preflight to the list above
|
|
@@ -103,9 +102,11 @@ export declare const PREFLIGHT_MECHANISMS: readonly ["ambiguity-preflight", "scr
|
|
|
103
102
|
export type PreflightMechanism = (typeof PREFLIGHT_MECHANISMS)[number];
|
|
104
103
|
/**
|
|
105
104
|
* Must a round claiming this mechanism be driven with a permissive rating for the mechanism to be
|
|
106
|
-
* observable at all? See {@link PREFLIGHT_MECHANISMS} for the whole rule
|
|
107
|
-
*
|
|
108
|
-
*
|
|
105
|
+
* observable at all? See {@link PREFLIGHT_MECHANISMS} for the whole rule. The §8 hardline floor
|
|
106
|
+
* answers **false**: it never sees a verdict, so a rating cannot change its answer and supplying one
|
|
107
|
+
* only moves the action column. `undefined` (the round claims no mechanism) is `false` too: an
|
|
108
|
+
* ordinary `model_free` case must keep going through the gate with no verdict, or its action would
|
|
109
|
+
* change.
|
|
109
110
|
*/
|
|
110
111
|
export declare function mechanismNeedsPermissiveRating(mechanism: ForcedByMechanism | undefined): boolean;
|
|
111
112
|
/** The in-process SUT target: an agent built from the run's own resolved config (the original and
|
package/dist/evalTypes.js
CHANGED
|
@@ -36,11 +36,7 @@ export const HARDLINE_REFUSAL_MARKER = 'hardline floor: refused';
|
|
|
36
36
|
* What DOES discriminate is **which deterministic mechanism fired**, which the target reports in the
|
|
37
37
|
* rationale, and which is exactly what the corpus records. So a deterministic case asserts that.
|
|
38
38
|
*/
|
|
39
|
-
export const FORCED_BY_MECHANISMS = [
|
|
40
|
-
'hardline-floor',
|
|
41
|
-
'ambiguity-preflight',
|
|
42
|
-
'script-env-leak-preflight',
|
|
43
|
-
];
|
|
39
|
+
export const FORCED_BY_MECHANISMS = ['hardline-floor', 'script-env-leak-preflight'];
|
|
44
40
|
/** The prefix a rationale carries for a decision a preflight forced, followed by the mechanism. */
|
|
45
41
|
export const FORCED_BY_MARKER = 'forced by';
|
|
46
42
|
/**
|
|
@@ -50,12 +46,11 @@ export const FORCED_BY_MARKER = 'forced by';
|
|
|
50
46
|
*/
|
|
51
47
|
export const FORCED_BY_ASSERTIONS = {
|
|
52
48
|
'hardline-floor': HARDLINE_REFUSAL_MARKER,
|
|
53
|
-
'ambiguity-preflight': `${FORCED_BY_MARKER}: ambiguity-preflight`,
|
|
54
49
|
'script-env-leak-preflight': `${FORCED_BY_MARKER}: script-env-leak-preflight`,
|
|
55
50
|
};
|
|
56
51
|
/**
|
|
57
|
-
* Which of {@link FORCED_BY_MECHANISMS} are **
|
|
58
|
-
* can only be observed by handing the gate a rating for them to OVERRIDE.
|
|
52
|
+
* Which of {@link FORCED_BY_MECHANISMS} are **preflight FINDINGS of the decision mapping** — the
|
|
53
|
+
* ones that can only be observed by handing the gate a rating for them to OVERRIDE.
|
|
59
54
|
*
|
|
60
55
|
* **This is the single statement of the rule, and everything else reads it through
|
|
61
56
|
* {@link mechanismNeedsPermissiveRating}.** It is kept here, next to the mechanism list, rather than
|
|
@@ -65,7 +60,7 @@ export const FORCED_BY_ASSERTIONS = {
|
|
|
65
60
|
*
|
|
66
61
|
* ## Why the floor is not on this list (measured, CFG-28)
|
|
67
62
|
*
|
|
68
|
-
*
|
|
63
|
+
* A preflight finding lives INSIDE `mapVerdictToAction`, and CFG-28 narrowed it to raise only an
|
|
69
64
|
* outcome that sits below the deterministic floor. Give that mapping no verdict and it substitutes
|
|
70
65
|
* core's fail-closed one, which is *at* the floor — so no preflight rewrites anything, every command
|
|
71
66
|
* comes back with the identical placeholder sentence, and the assertion is a constant again. Handing
|
|
@@ -78,25 +73,25 @@ export const FORCED_BY_ASSERTIONS = {
|
|
|
78
73
|
* verdict — so a rating cannot change the answer, and supplying one buys nothing. It also costs:
|
|
79
74
|
* `mapVerdictToAction` does not consult the floor, so a permissive rating on `rm -rf /` returns
|
|
80
75
|
* `approve`, which would move the action column of **5 of the corpus's 6 `fl-*` cases** off the
|
|
81
|
-
* `escalate` they author. Measured over all 22 deterministic corpus cases
|
|
82
|
-
* section of `handoff/task-1-report.md`.
|
|
76
|
+
* `escalate` they author. Measured over all 22 deterministic corpus cases.
|
|
83
77
|
*
|
|
84
78
|
* ## The property that makes this self-justifying
|
|
85
79
|
*
|
|
86
|
-
* When the claimed preflight really does fire, the rating does not move the action —
|
|
87
|
-
*
|
|
88
|
-
*
|
|
89
|
-
*
|
|
80
|
+
* When the claimed preflight really does fire, the rating does not move the action — the corpus
|
|
81
|
+
* cases that claim one (`el-01`, `el-02`) are `escalate` with or without it. When it does not fire,
|
|
82
|
+
* the marker and the action go red **together**. That is the discrimination, not a side effect of
|
|
83
|
+
* it.
|
|
90
84
|
*/
|
|
91
85
|
export const PREFLIGHT_MECHANISMS = [
|
|
92
|
-
'ambiguity-preflight',
|
|
93
86
|
'script-env-leak-preflight',
|
|
94
87
|
];
|
|
95
88
|
/**
|
|
96
89
|
* Must a round claiming this mechanism be driven with a permissive rating for the mechanism to be
|
|
97
|
-
* observable at all? See {@link PREFLIGHT_MECHANISMS} for the whole rule
|
|
98
|
-
*
|
|
99
|
-
*
|
|
90
|
+
* observable at all? See {@link PREFLIGHT_MECHANISMS} for the whole rule. The §8 hardline floor
|
|
91
|
+
* answers **false**: it never sees a verdict, so a rating cannot change its answer and supplying one
|
|
92
|
+
* only moves the action column. `undefined` (the round claims no mechanism) is `false` too: an
|
|
93
|
+
* ordinary `model_free` case must keep going through the gate with no verdict, or its action would
|
|
94
|
+
* change.
|
|
100
95
|
*/
|
|
101
96
|
export function mechanismNeedsPermissiveRating(mechanism) {
|
|
102
97
|
return (mechanism !== undefined &&
|
package/dist/evalTypes.js.map
CHANGED
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"evalTypes.js","sourceRoot":"","sources":["../src/evalTypes.ts"],"names":[],"mappings":"AAgBA;;;;mGAImG;AACnG,MAAM,CAAC,MAAM,2BAA2B,GAAG,CAAC,CAAC;AAE7C;;;;;;;;;;;;;;GAcG;AACH,MAAM,CAAC,MAAM,uBAAuB,GAAG,yBAAyB,CAAC;AAEjE;;;;;;;;;;;;;;;GAeG;AACH,MAAM,CAAC,MAAM,oBAAoB,GAAG
|
|
1
|
+
{"version":3,"file":"evalTypes.js","sourceRoot":"","sources":["../src/evalTypes.ts"],"names":[],"mappings":"AAgBA;;;;mGAImG;AACnG,MAAM,CAAC,MAAM,2BAA2B,GAAG,CAAC,CAAC;AAE7C;;;;;;;;;;;;;;GAcG;AACH,MAAM,CAAC,MAAM,uBAAuB,GAAG,yBAAyB,CAAC;AAEjE;;;;;;;;;;;;;;;GAeG;AACH,MAAM,CAAC,MAAM,oBAAoB,GAAG,CAAC,gBAAgB,EAAE,2BAA2B,CAAU,CAAC;AAK7F,mGAAmG;AACnG,MAAM,CAAC,MAAM,gBAAgB,GAAG,WAAW,CAAC;AAE5C;;;;GAIG;AACH,MAAM,CAAC,MAAM,oBAAoB,GAAsC;IACrE,gBAAgB,EAAE,uBAAuB;IACzC,2BAA2B,EAAE,GAAG,gBAAgB,6BAA6B;CAC9E,CAAC;AAEF;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAiCG;AACH,MAAM,CAAC,MAAM,oBAAoB,GAAG;IAClC,2BAA2B;CACoB,CAAC;AAWlD;;;;;;;GAOG;AACH,MAAM,UAAU,8BAA8B,CAAC,SAAwC;IACrF,OAAO,CACL,SAAS,KAAK,SAAS;QACtB,oBAAqD,CAAC,QAAQ,CAAC,SAAS,CAAC,CAC3E,CAAC;AACJ,CAAC"}
|
package/dist/raterTarget.d.ts
CHANGED
|
@@ -52,7 +52,7 @@ export { HARDLINE_REFUSAL_MARKER };
|
|
|
52
52
|
*/
|
|
53
53
|
export declare const NO_RATING_CALL_MARKER = "no rating call";
|
|
54
54
|
/**
|
|
55
|
-
* Ask core which sentence each preflight
|
|
55
|
+
* Ask core which reason sentence each deterministic preflight writes, and index it.
|
|
56
56
|
*
|
|
57
57
|
* Exported for the unit suite: a run in which this returns fewer than the probed mechanisms means
|
|
58
58
|
* the gate no longer distinguishes them, and every `forced_by` assertion in every corpus would fail.
|
package/dist/raterTarget.js
CHANGED
|
@@ -34,8 +34,8 @@ export { HARDLINE_REFUSAL_MARKER };
|
|
|
34
34
|
*/
|
|
35
35
|
export const NO_RATING_CALL_MARKER = 'no rating call';
|
|
36
36
|
/**
|
|
37
|
-
* One command per deterministic preflight, used ONLY to ask core's decision mapping what
|
|
38
|
-
* preflight says. They are probes, never rated, and nothing about them reaches a suite.
|
|
37
|
+
* One command per deterministic preflight FINDING, used ONLY to ask core's decision mapping what
|
|
38
|
+
* that preflight says. They are probes, never rated, and nothing about them reaches a suite.
|
|
39
39
|
*
|
|
40
40
|
* ## Why probe at all
|
|
41
41
|
*
|
|
@@ -44,8 +44,8 @@ export const NO_RATING_CALL_MARKER = 'no rating call';
|
|
|
44
44
|
* the REASON on the verdict `mapVerdictToAction` returns. Two ways to read it were available and
|
|
45
45
|
* both were rejected:
|
|
46
46
|
*
|
|
47
|
-
* - copy core's
|
|
48
|
-
*
|
|
47
|
+
* - copy core's reason sentences into this package — a second copy of core's prose, silently stale
|
|
48
|
+
* the day it is reworded;
|
|
49
49
|
* - re-evaluate `classifyCommand(...) === null` / `hasScriptEnvLeakRisk(...)` here — a second copy
|
|
50
50
|
* of the *conditions*, i.e. exactly the "never re-decide the gate" rule this module opens with,
|
|
51
51
|
* and one that would keep naming a mechanism after that mechanism had been deleted from the gate.
|
|
@@ -62,7 +62,6 @@ export const NO_RATING_CALL_MARKER = 'no rating call';
|
|
|
62
62
|
* {@link ../evalTypes.js PREFLIGHT_MECHANISMS} for the whole rule.
|
|
63
63
|
*/
|
|
64
64
|
const MECHANISM_PROBES = {
|
|
65
|
-
'ambiguity-preflight': 'echo $(echo probe)',
|
|
66
65
|
'script-env-leak-preflight': 'node probe.js $PROBE_ENV_VALUE',
|
|
67
66
|
};
|
|
68
67
|
/**
|
|
@@ -108,14 +107,14 @@ function calibrateGate() {
|
|
|
108
107
|
return { index, permissive };
|
|
109
108
|
}
|
|
110
109
|
/**
|
|
111
|
-
* Ask core which sentence each preflight
|
|
110
|
+
* Ask core which reason sentence each deterministic preflight writes, and index it.
|
|
112
111
|
*
|
|
113
112
|
* Exported for the unit suite: a run in which this returns fewer than the probed mechanisms means
|
|
114
113
|
* the gate no longer distinguishes them, and every `forced_by` assertion in every corpus would fail.
|
|
115
114
|
* That must be a red unit test rather than a surprise in someone's eval report.
|
|
116
115
|
*/
|
|
117
116
|
export function calibrateMechanisms() {
|
|
118
|
-
return calibrateGate().index;
|
|
117
|
+
return new Map(calibrateGate().index);
|
|
119
118
|
}
|
|
120
119
|
/**
|
|
121
120
|
* The rating a `forced_by: <preflight>` round is driven with, discovered from core rather than
|
|
@@ -127,14 +126,14 @@ export function calibratePermissiveRating() {
|
|
|
127
126
|
return calibrateGate().permissive;
|
|
128
127
|
}
|
|
129
128
|
/**
|
|
130
|
-
* Which deterministic
|
|
131
|
-
*
|
|
132
|
-
*
|
|
129
|
+
* Which deterministic mechanism forced this decision — read out of what the gate RETURNED, never
|
|
130
|
+
* recomputed from the command. `undefined` when the decision was not forced by an indexed mechanism
|
|
131
|
+
* (a rated verdict, an unrated rung, or a mechanism that no longer fires).
|
|
133
132
|
*/
|
|
134
|
-
function forcedMechanism(
|
|
135
|
-
if (!verdict?.reason)
|
|
133
|
+
function forcedMechanism(decision) {
|
|
134
|
+
if (!decision.verdict?.reason)
|
|
136
135
|
return undefined;
|
|
137
|
-
return calibrated().index.get(verdict.reason);
|
|
136
|
+
return calibrated().index.get(decision.verdict.reason);
|
|
138
137
|
}
|
|
139
138
|
/** The memoized {@link calibrateGate} — core's answers do not change within a run. */
|
|
140
139
|
function calibrated() {
|
|
@@ -184,7 +183,7 @@ function resolveRung(target, config) {
|
|
|
184
183
|
* is not: with no verdict, `mapVerdictToAction` returns the same action for every command at a rated
|
|
185
184
|
* rung. See {@link ../evalTypes.js FORCED_BY_MECHANISMS}.
|
|
186
185
|
*/
|
|
187
|
-
function buildRationale(floorDescription,
|
|
186
|
+
function buildRationale(floorDescription, decision, noRatingReason, ratingIn) {
|
|
188
187
|
const parts = [];
|
|
189
188
|
if (floorDescription !== undefined) {
|
|
190
189
|
parts.push(`${HARDLINE_REFUSAL_MARKER} (${floorDescription})`);
|
|
@@ -192,6 +191,7 @@ function buildRationale(floorDescription, verdict, noRatingReason, ratingIn) {
|
|
|
192
191
|
if (noRatingReason !== undefined) {
|
|
193
192
|
parts.push(`${NO_RATING_CALL_MARKER} (${noRatingReason})`);
|
|
194
193
|
}
|
|
194
|
+
const verdict = decision.verdict;
|
|
195
195
|
// On the zero-call path the only reason worth reporting is one the GATE wrote. Two sentences are
|
|
196
196
|
// not: core's fail-closed placeholder (it says the rater could not evaluate the command, which on
|
|
197
197
|
// this path never happened) and our own probe reason (it is a lever, not a finding). Both are
|
|
@@ -200,11 +200,19 @@ function buildRationale(floorDescription, verdict, noRatingReason, ratingIn) {
|
|
|
200
200
|
// where the mapping returns the verdict untouched.
|
|
201
201
|
const handedBack = noRatingReason !== undefined &&
|
|
202
202
|
(verdict?.reason === ratingIn?.reason || verdict?.reason === FAIL_CLOSED_VERDICT.reason);
|
|
203
|
-
|
|
204
|
-
|
|
205
|
-
|
|
206
|
-
|
|
207
|
-
|
|
203
|
+
const reason = verdict?.reason && !handedBack ? verdict.reason : undefined;
|
|
204
|
+
const mechanism = forcedMechanism(decision);
|
|
205
|
+
if (mechanism !== undefined) {
|
|
206
|
+
// A forced mechanism can arrive with no reason worth parenthesising — the `handedBack` case
|
|
207
|
+
// just above, where the only sentence available came from us rather than from the gate. The
|
|
208
|
+
// marker alone is then the honest report, and it is still the exact substring a `forced_by:`
|
|
209
|
+
// assertion looks for.
|
|
210
|
+
parts.push(reason === undefined
|
|
211
|
+
? FORCED_BY_ASSERTIONS[mechanism]
|
|
212
|
+
: `${FORCED_BY_ASSERTIONS[mechanism]} (${reason})`);
|
|
213
|
+
}
|
|
214
|
+
else if (reason !== undefined) {
|
|
215
|
+
parts.push(reason);
|
|
208
216
|
}
|
|
209
217
|
return parts.length > 0 ? parts.join('\n') : undefined;
|
|
210
218
|
}
|
|
@@ -236,7 +244,7 @@ function buildRationale(floorDescription, verdict, noRatingReason, ratingIn) {
|
|
|
236
244
|
* the same for every command with no verdict, so it discriminates nothing; what discriminates is
|
|
237
245
|
* the mechanism, in the rationale ({@link buildRationale}).
|
|
238
246
|
*
|
|
239
|
-
* **A round claiming a preflight is driven with a permissive rating** — see
|
|
247
|
+
* **A round claiming a preflight FINDING is driven with a permissive rating** — see
|
|
240
248
|
* {@link ../evalTypes.js PREFLIGHT_MECHANISMS}. Still no model call: the rating is a stub, derived
|
|
241
249
|
* from core, and it is the only way a preflight is observable at all since CFG-28. It does not
|
|
242
250
|
* flatter the case — when the claimed preflight really fires the action is the same either way, and
|
|
@@ -290,7 +298,7 @@ async function classifyOneRound(command, roundIndex, request, rung, config, mode
|
|
|
290
298
|
// would report is OUR lever floored by a preflight, never a judgement anyone rendered.
|
|
291
299
|
...(deterministicOnly ? {} : { label: decision.verdict?.outcome }),
|
|
292
300
|
action: decision.action,
|
|
293
|
-
rationale: buildRationale(floor?.description, decision
|
|
301
|
+
rationale: buildRationale(floor?.description, decision, noRatingReason, verdict),
|
|
294
302
|
modelCalls,
|
|
295
303
|
};
|
|
296
304
|
}
|
package/dist/raterTarget.js.map
CHANGED
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"raterTarget.js","sourceRoot":"","sources":["../src/raterTarget.ts"],"names":[],"mappings":"AAsBA,OAAO,EAAE,aAAa,EAAE,MAAM,4CAA4C,CAAC;AAE3E,OAAO,EACL,cAAc,EACd,WAAW,EACX,gBAAgB,GACjB,MAAM,0CAA0C,CAAC;AAElD,OAAO,EACL,mBAAmB,EACnB,cAAc,EACd,kBAAkB,EAClB,gBAAgB,GACjB,MAAM,uCAAuC,CAAC;AAE/C,OAAO,EAAE,iBAAiB,EAAE,MAAM,4CAA4C,CAAC;AAC/E,OAAO,EAAE,cAAc,EAAE,MAAM,yCAAyC,CAAC;AACzE,OAAO,EAAE,GAAG,EAAE,MAAM,wCAAwC,CAAC;AAE7D,OAAO,EACL,oBAAoB,EACpB,uBAAuB,EACvB,8BAA8B,GAC/B,MAAM,mBAAmB,CAAC;AAU3B;;;;;;;;;;;;;;GAcG;AACH,OAAO,EAAE,uBAAuB,EAAE,CAAC;AAEnC;;;;;;;;;;GAUG;AACH,MAAM,CAAC,MAAM,qBAAqB,GAAG,gBAAgB,CAAC;AAEtD;;;;;;;;;;;;;;;;;;;;;;;;;;;GA2BG;AACH,MAAM,gBAAgB,GAAiD;IACrE,
|
|
1
|
+
{"version":3,"file":"raterTarget.js","sourceRoot":"","sources":["../src/raterTarget.ts"],"names":[],"mappings":"AAsBA,OAAO,EAAE,aAAa,EAAE,MAAM,4CAA4C,CAAC;AAE3E,OAAO,EACL,cAAc,EACd,WAAW,EACX,gBAAgB,GACjB,MAAM,0CAA0C,CAAC;AAElD,OAAO,EACL,mBAAmB,EACnB,cAAc,EACd,kBAAkB,EAClB,gBAAgB,GACjB,MAAM,uCAAuC,CAAC;AAE/C,OAAO,EAAE,iBAAiB,EAAE,MAAM,4CAA4C,CAAC;AAC/E,OAAO,EAAE,cAAc,EAAE,MAAM,yCAAyC,CAAC;AACzE,OAAO,EAAE,GAAG,EAAE,MAAM,wCAAwC,CAAC;AAE7D,OAAO,EACL,oBAAoB,EACpB,uBAAuB,EACvB,8BAA8B,GAC/B,MAAM,mBAAmB,CAAC;AAU3B;;;;;;;;;;;;;;GAcG;AACH,OAAO,EAAE,uBAAuB,EAAE,CAAC;AAEnC;;;;;;;;;;GAUG;AACH,MAAM,CAAC,MAAM,qBAAqB,GAAG,gBAAgB,CAAC;AAEtD;;;;;;;;;;;;;;;;;;;;;;;;;;;GA2BG;AACH,MAAM,gBAAgB,GAAiD;IACrE,2BAA2B,EAAE,gCAAgC;CAC9D,CAAC;AAEF;;;;;GAKG;AACH,MAAM,YAAY,GAAG,mCAAmC,CAAC;AAYzD,8DAA8D;AAC9D,IAAI,WAAwC,CAAC;AAE7C;;;GAGG;AACH,SAAS,aAAa;IACpB,MAAM,KAAK,GAAG,IAAI,GAAG,EAA6B,CAAC;IACnD,IAAI,UAA8B,CAAC;IACnC,mEAAmE;IACnE,MAAM,IAAI,GAAG,cAAc,CAAC,IAAI,CAAC,CAAC,SAAS,EAAE,EAAE,CAAC,WAAW,CAAC,SAAS,CAAC,CAAC,CAAC;IACxE,IAAI,IAAI,KAAK,SAAS;QAAE,OAAO,EAAE,KAAK,EAAE,UAAU,EAAE,CAAC;IAErD,KAAK,MAAM,CAAC,SAAS,EAAE,KAAK,CAAC,IAAI,MAAM,CAAC,OAAO,CAAC,gBAAgB,CAG7D,EAAE,CAAC;QACJ,KAAK,MAAM,OAAO,IAAI,cAAc,EAAE,CAAC;YACrC,MAAM,MAAM,GAAG,kBAAkB,CAAC,KAAK,EAAE,EAAE,OAAO,EAAE,MAAM,EAAE,YAAY,EAAE,EAAE,EAAE,IAAI,EAAE,CAAC,CAAC,OAAO;gBAC3F,EAAE,MAAM,CAAC;YACX,4FAA4F;YAC5F,iEAAiE;YACjE,IAAI,MAAM,KAAK,SAAS,IAAI,MAAM,KAAK,YAAY;gBAAE,SAAS;YAC9D,UAAU,KAAK,OAAO,CAAC;YACvB,IAAI,KAAK,CAAC,GAAG,CAAC,MAAM,CAAC,EAAE,CAAC;gBACtB,4FAA4F;gBAC5F,8CAA8C;gBAC9C,KAAK,CAAC,MAAM,CAAC,MAAM,CAAC,CAAC;YACvB,CAAC;iBAAM,CAAC;gBACN,KAAK,CAAC,GAAG,CAAC,MAAM,EAAE,SAAS,CAAC,CAAC;YAC/B,CAAC;YACD,MAAM;QACR,CAAC;IACH,CAAC;IAED,OAAO,EAAE,KAAK,EAAE,UAAU,EAAE,CAAC;AAC/B,CAAC;AAED;;;;;;GAMG;AACH,MAAM,UAAU,mBAAmB;IACjC,OAAO,IAAI,GAAG,CAAC,aAAa,EAAE,CAAC,KAAK,CAAC,CAAC;AACxC,CAAC;AAED;;;;;GAKG;AACH,MAAM,UAAU,yBAAyB;IACvC,OAAO,aAAa,EAAE,CAAC,UAAU,CAAC;AACpC,CAAC;AAED;;;;GAIG;AACH,SAAS,eAAe,CAAC,QAAuB;IAC9C,IAAI,CAAC,QAAQ,CAAC,OAAO,EAAE,MAAM;QAAE,OAAO,SAAS,CAAC;IAChD,OAAO,UAAU,EAAE,CAAC,KAAK,CAAC,GAAG,CAAC,QAAQ,CAAC,OAAO,CAAC,MAAM,CAAC,CAAC;AACzD,CAAC;AAED,sFAAsF;AACtF,SAAS,UAAU;IACjB,WAAW,KAAK,aAAa,EAAE,CAAC;IAChC,OAAO,WAAW,CAAC;AACrB,CAAC;AAeD;;;;;;;;;;;;;;;;;GAiBG;AACH,SAAS,WAAW,CAAC,MAAmB,EAAE,MAAiB;IACzD,MAAM,YAAY,GAChB,OAAO,MAAM,CAAC,SAAS,KAAK,QAAQ,CAAC,CAAC,CAAC,MAAM,CAAC,SAAS,CAAC,CAAC,CAAC,MAAM,CAAC,SAAS,EAAE,IAAI,CAAC;IACnF,IAAI,YAAY,KAAK,SAAS;QAAE,OAAO,MAAM,CAAC,IAAI,CAAC;IAEnD,MAAM,UAAU,GAAG,gBAAgB,CAAC,MAAM,EAAE,SAAS,CAAC,CAAC,IAAI,CAAC;IAC5D,IAAI,UAAU,KAAK,MAAM,CAAC,IAAI,EAAE,CAAC;QAC/B,cAAc,CACZ,yBAAyB,UAAU,wDAAwD;YACzF,aAAa,MAAM,CAAC,IAAI,oCAAoC,UAAU,IAAI,CAC7E,CAAC;IACJ,CAAC;IACD,OAAO,UAAU,CAAC;AACpB,CAAC;AAED;;;;;;;;;;;;;GAaG;AACH,SAAS,cAAc,CACrB,gBAAoC,EACpC,QAAuB,EACvB,cAAkC,EAClC,QAAwC;IAExC,MAAM,KAAK,GAAa,EAAE,CAAC;IAC3B,IAAI,gBAAgB,KAAK,SAAS,EAAE,CAAC;QACnC,KAAK,CAAC,IAAI,CAAC,GAAG,uBAAuB,KAAK,gBAAgB,GAAG,CAAC,CAAC;IACjE,CAAC;IACD,IAAI,cAAc,KAAK,SAAS,EAAE,CAAC;QACjC,KAAK,CAAC,IAAI,CAAC,GAAG,qBAAqB,KAAK,cAAc,GAAG,CAAC,CAAC;IAC7D,CAAC;IACD,MAAM,OAAO,GAAG,QAAQ,CAAC,OAAO,CAAC;IACjC,iGAAiG;IACjG,kGAAkG;IAClG,8FAA8F;IAC9F,+FAA+F;IAC/F,gGAAgG;IAChG,mDAAmD;IACnD,MAAM,UAAU,GACd,cAAc,KAAK,SAAS;QAC5B,CAAC,OAAO,EAAE,MAAM,KAAK,QAAQ,EAAE,MAAM,IAAI,OAAO,EAAE,MAAM,KAAK,mBAAmB,CAAC,MAAM,CAAC,CAAC;IAC3F,MAAM,MAAM,GAAG,OAAO,EAAE,MAAM,IAAI,CAAC,UAAU,CAAC,CAAC,CAAC,OAAO,CAAC,MAAM,CAAC,CAAC,CAAC,SAAS,CAAC;IAC3E,MAAM,SAAS,GAAG,eAAe,CAAC,QAAQ,CAAC,CAAC;IAC5C,IAAI,SAAS,KAAK,SAAS,EAAE,CAAC;QAC5B,4FAA4F;QAC5F,4FAA4F;QAC5F,6FAA6F;QAC7F,uBAAuB;QACvB,KAAK,CAAC,IAAI,CACR,MAAM,KAAK,SAAS;YAClB,CAAC,CAAC,oBAAoB,CAAC,SAAS,CAAC;YACjC,CAAC,CAAC,GAAG,oBAAoB,CAAC,SAAS,CAAC,KAAK,MAAM,GAAG,CACrD,CAAC;IACJ,CAAC;SAAM,IAAI,MAAM,KAAK,SAAS,EAAE,CAAC;QAChC,KAAK,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC;IACrB,CAAC;IACD,OAAO,KAAK,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,SAAS,CAAC;AACzD,CAAC;AAED;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAiCG;AACH,KAAK,UAAU,gBAAgB,CAC7B,OAAe,EACf,UAAkB,EAClB,OAAwB,EACxB,IAAkB,EAClB,MAAiB,EACjB,KAAgC,EAChC,OAA2C;IAE3C,MAAM,OAAO,GAAG,OAAO,CAAC,IAAI,EAAE,CAAC;IAC/B,IAAI,OAAO,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;QACzB,4FAA4F;QAC5F,2FAA2F;QAC3F,OAAO;YACL,EAAE,EAAE,KAAK;YACT,UAAU,EAAE,CAAC;YACb,KAAK,EAAE,SAAS,OAAO,CAAC,MAAM,kEAAkE;SACjG,CAAC;IACJ,CAAC;IAED,yEAAyE;IACzE,MAAM,KAAK,GAAG,aAAa,CAAC,OAAO,CAAC,CAAC;IAErC,MAAM,cAAc,GAAG,OAAO,CAAC,SAAS;QACtC,CAAC,CAAC,YAAY;QACd,CAAC,CAAC,WAAW,CAAC,IAAI,CAAC;YACjB,CAAC,CAAC,SAAS;YACX,CAAC,CAAC,SAAS,IAAI,iBAAiB,CAAC;IACrC,MAAM,iBAAiB,GAAG,cAAc,KAAK,SAAS,CAAC;IACvD,IAAI,OAAuC,CAAC;IAC5C,IAAI,UAAU,GAAG,CAAC,CAAC;IACnB,IAAI,iBAAiB,EAAE,CAAC;QACtB,2FAA2F;QAC3F,8FAA8F;QAC9F,6FAA6F;QAC7F,qFAAqF;QACrF,MAAM,UAAU,GAAG,UAAU,EAAE,CAAC,UAAU,CAAC;QAC3C,IACE,8BAA8B,CAAC,OAAO,CAAC,QAAQ,EAAE,CAAC,UAAU,CAAC,CAAC;YAC9D,UAAU,KAAK,SAAS,EACxB,CAAC;YACD,OAAO,GAAG,EAAE,OAAO,EAAE,UAA2C,EAAE,MAAM,EAAE,YAAY,EAAE,CAAC;QAC3F,CAAC;IACH,CAAC;SAAM,CAAC;QACN,OAAO,GAAG,MAAM,gBAAgB,CAAC,OAAO,EAAE,MAAM,EAAE;YAChD,KAAK;YACL,IAAI,EAAE,OAAO,EAAE,IAAI,IAAI,GAAG,EAAE,IAAI;YAChC,SAAS,EAAE,OAAO,EAAE,SAAS;SAC9B,CAAC,CAAC;QACH,UAAU,GAAG,CAAC,CAAC;IACjB,CAAC;IAED,MAAM,QAAQ,GAAG,kBAAkB,CAAC,OAAO,EAAE,OAAO,EAAE,EAAE,IAAI,EAAE,CAAC,CAAC;IAChE,OAAO;QACL,EAAE,EAAE,IAAI;QACR,4FAA4F;QAC5F,gGAAgG;QAChG,uFAAuF;QACvF,GAAG,CAAC,iBAAiB,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,EAAE,KAAK,EAAE,QAAQ,CAAC,OAAO,EAAE,OAAO,EAAE,CAAC;QAClE,MAAM,EAAE,QAAQ,CAAC,MAAM;QACvB,SAAS,EAAE,cAAc,CAAC,KAAK,EAAE,WAAW,EAAE,QAAQ,EAAE,cAAc,EAAE,OAAO,CAAC;QAChF,UAAU;KACX,CAAC;AACJ,CAAC;AAED;;;;;;;;;;;;GAYG;AACH,MAAM,CAAC,KAAK,UAAU,oBAAoB,CACxC,MAAmB,EACnB,MAAiB,EACjB,OAAgC;IAEhC,MAAM,IAAI,GAAG,WAAW,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;IACzC,MAAM,YAAY,GAChB,MAAM,CAAC,SAAS,KAAK,SAAS,CAAC,CAAC,CAAC,gBAAgB,CAAC,MAAM,EAAE,SAAS,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,SAAS,CAAC;IACzF,yFAAyF;IACzF,gGAAgG;IAChG,gFAAgF;IAChF,MAAM,KAAK,GACT,OAAO,EAAE,KAAK,IAAI,CAAC,YAAY,CAAC,CAAC,CAAC,MAAM,iBAAiB,CAAC,YAAY,CAAC,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC;IAEvF,OAAO,KAAK,EAAE,OAAwB,EAA8B,EAAE;QACpE,MAAM,QAAQ,GAAsB,EAAE,CAAC;QACvC,8FAA8F;QAC9F,gGAAgG;QAChG,wDAAwD;QACxD,KAAK,MAAM,CAAC,UAAU,EAAE,KAAK,CAAC,IAAI,OAAO,CAAC,MAAM,CAAC,OAAO,EAAE,EAAE,CAAC;YAC3D,QAAQ,CAAC,IAAI,CACX,MAAM,gBAAgB,CAAC,KAAK,EAAE,UAAU,EAAE,OAAO,EAAE,IAAI,EAAE,MAAM,EAAE,KAAK,EAAE,OAAO,CAAC,CACjF,CAAC;QACJ,CAAC;QACD,OAAO,QAAQ,CAAC;IAClB,CAAC,CAAC;AACJ,CAAC"}
|
package/package.json
CHANGED
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
{
|
|
2
2
|
"name": "@gaunt-sloth/batch",
|
|
3
|
-
"version": "2.0.0-alpha.
|
|
3
|
+
"version": "2.0.0-alpha.27",
|
|
4
4
|
"description": "Batch matrix runtime for Gaunt Sloth (BATCH-1): run a prompt-executable over a matrix of models and/or content-bound inputs",
|
|
5
5
|
"license": "MIT",
|
|
6
6
|
"author": "Andrew Kondratev",
|
|
@@ -37,8 +37,8 @@
|
|
|
37
37
|
"@langchain/core": "^1.2.3",
|
|
38
38
|
"yaml": "^2.9.0",
|
|
39
39
|
"zod": "^4.4.3",
|
|
40
|
-
"@gaunt-sloth/agent": "2.0.0-alpha.
|
|
41
|
-
"@gaunt-sloth/core": "2.0.0-alpha.
|
|
40
|
+
"@gaunt-sloth/agent": "2.0.0-alpha.27",
|
|
41
|
+
"@gaunt-sloth/core": "2.0.0-alpha.27"
|
|
42
42
|
},
|
|
43
43
|
"files": [
|
|
44
44
|
"./dist/*"
|