评测与研究

jev-broadcast-lab

@4anti2JavaScript更新于 2026-09-17可当场跑

TypeSafe Jev 的操作实验室,包含棋类竞技场、封闭式展位和仅供复盘的 Stockfish HUD。

英文原文

Operator lab for TypeSafe Jev. Chess Arena, closed-schema booths, Stockfish HUD for review only.

4anti/jev-broadcast-lab

它在哪儿调用了 Jev

UPSTREAM = os.environ.get("TYPESAFE_UPSTREAM", "https://api.typesafe.ai/v1/systemone")

server.py:50

链接指向我们抓取当天的那个 commit,行号是准的。

它问 Jev 的问题

下面是从这个项目源码里原样取出来的 question 组合。

取自 web/home.js:37

  1. what选项

    What is this page describing?

    • broadcast_labA lab for closed-schema Jev calls
    • chess_engineA chess engine like Stockfish
    • chatbotA chat model that writes prose
    • policy_gateA safety policy enforcer
    • chess_matchA live chess match

用你自己的内容跑一遍

代码
import { experimental_evaluate as evaluate } from 'ai';

const { answers } = await evaluate({
  model: 'typesafe-ai/jev',
  state,
  questions: {
    what: { type: 'choice', instructions: 'What is this page describing?',
      criteria: { broadcast_lab: 'A lab for closed-schema Jev calls', chess_engine: 'A chess engine like Stockfish', chatbot: 'A chat model that writes prose', policy_gate: 'A safety policy enforcer', chess_match: 'A live chess match' } },
  },
});

取自 web/lab/call/call.js:26

  1. fn选项

    Which function should run?

    • search_ticketsLook up tickets
    • create_refundIssue a refund
    • page_oncallPage a human
    • noneNo tool
  2. reason选项

    If refunding, which reason enum? Use none if not a refund.

    • duplicate_chargeCharged twice
    • service_creditGoodwill credit
    • chargeback_avoidDefensive refund
    • noneNot a refund

用你自己的内容跑一遍

代码
import { experimental_evaluate as evaluate } from 'ai';

const { answers } = await evaluate({
  model: 'typesafe-ai/jev',
  state,
  questions: {
    fn: { type: 'choice', instructions: 'Which function should run?',
      criteria: { search_tickets: 'Look up tickets', create_refund: 'Issue a refund', page_oncall: 'Page a human', none: 'No tool' } },
    reason: { type: 'choice', instructions: 'If refunding, which reason enum? Use none if not a refund.',
      criteria: { duplicate_charge: 'Charged twice', service_credit: 'Goodwill credit', chargeback_avoid: 'Defensive refund', none: 'Not a refund' } },
  },
});

取自 web/lab/cheat/cheat.js:188

  1. verdict选项

    Classify the act relative to this couple's setup. Use community_rule only if the setup named a religious or family boundary. Use depends_on_couple if exclusive people would split and no rule was named. Do not moralize past the labels.

    • not_cheatingNo extra-pair betrayal relative to this setup
    • microSmall extra-pair attention or concealment, short of an affair
    • emotional_affairSecret emotional primary bond with someone else
    • sexual_affairSexual or clearly sexual-digital contact that breaks exclusivity
    • depends_on_coupleReasonable exclusive couples would split on this
    • community_ruleA breach of a named religious, family, or subculture rule more than of sex-as-such
  2. secrecy是/否

    Concealment from the partner (or from the named family rule-holders) is doing real work in this case.

  3. harm打分

    Typical trust harm if this stays hidden, for this setup. Not a legal finding.

    • 0None
    • 1Sting
    • 2Break
    • 3Blow-up

用你自己的内容跑一遍

代码
import { experimental_evaluate as evaluate } from 'ai';

const { answers } = await evaluate({
  model: 'typesafe-ai/jev',
  state,
  questions: {
    verdict: { type: 'choice', instructions: 'Classify the act relative to this couple\'s setup. Use community_rule only if the setup named a religious or family boundary. Use depends_on_couple if exclusive people would split and no rule was named. Do not moralize past the labels.',
      criteria: { not_cheating: 'No extra-pair betrayal relative to this setup', micro: 'Small extra-pair attention or concealment, short of an affair', emotional_affair: 'Secret emotional primary bond with someone else', sexual_affair: 'Sexual or clearly sexual-digital contact that breaks exclusivity', depends_on_couple: 'Reasonable exclusive couples would split on this', community_rule: 'A breach of a named religious, family, or subculture rule more than of sex-as-such' } },
    secrecy: { type: 'boolean', instructions: 'Concealment from the partner (or from the named family rule-holders) is doing real work in this case.' },
    harm: { type: 'score', instructions: 'Typical trust harm if this stays hidden, for this setup. Not a legal finding.',
      criteria: ['None', 'Sting', 'Break', 'Blow-up'] },
  },
});

取自 web/lab/debate/debate.js:124

  1. target选项

    What did the counter aim at?

    • claimThe stated claim
    • premiseA supporting reason
    • personThe speaker
    • motiveWhy they might say it
    • definitionThe meaning of a word
    • side_issueA different topic
    • evidenceThe data or source
  2. quality打分

    How well does the counter actually answer the original claim?

    • 0Miss
    • 1Glance
    • 2Partial
    • 3Meets it
  3. on_point是/否

    The counter addresses the original claim rather than the speaker or a side issue.

用你自己的内容跑一遍

代码
import { experimental_evaluate as evaluate } from 'ai';

const { answers } = await evaluate({
  model: 'typesafe-ai/jev',
  state,
  questions: {
    target: { type: 'choice', instructions: 'What did the counter aim at?',
      criteria: { claim: 'The stated claim', premise: 'A supporting reason', person: 'The speaker', motive: 'Why they might say it', definition: 'The meaning of a word', side_issue: 'A different topic', evidence: 'The data or source' } },
    quality: { type: 'score', instructions: 'How well does the counter actually answer the original claim?',
      criteria: ['Miss', 'Glance', 'Partial', 'Meets it'] },
    on_point: { type: 'boolean', instructions: 'The counter addresses the original claim rather than the speaker or a side issue.' },
  },
});

同类的其他项目