Skip to content
System One

AshutoshVJTI/progressgate

Watches recent agent steps for semantic stagnation, meaning different actions that still chase the same contradicted assumption. Jev reads the trajectory and a small deterministic policy returns CONTINUE, WARN, REPLAN or HALT.

AshutoshVJTI/progressgate on GitHub

Stars
0
Forks
0
Language
TypeScript
License
MIT
Last push
18 Sep 2026

Read from GitHub on . Counts change daily.

How it uses Jev

export async function fetchSignals(client: TypeSafeClient, input: CheckInput): Promise<JevResponse> {
  const started = Date.now();
  const state = formatState(input);

  const response = await client.systemOne({
    state,
    questions: {
      assumptionContradicted: noul(
        "Does `latest_step` continue to rely on an assumption, hypothesis, or root cause that evidence already present in `recent_steps` or `authoritative_state` has contradicted or ruled out? Answer yes only when there is actual contradicting evidence already present, not merely because the same general topic is being revisited.",
      ),
      strategyNovelty: noul(
        "Is `latest_step` testing a genuinely different underlying strategy or causal hypothesis than the one being tested across `recent_steps`, rather than a surface-level variation -- different tool, different wording, different parameter -- of the same underlying hypothesis?",
      ),
      materialProgress: noul(
        "Did `latest_step` materially move the real, externally-verifiable task state toward `goal` -- not the agent's stated plan or belief, but the actual state of the system, code, data, or environment? Answer no if the real state relevant to `goal` is unchanged from before `latest_step`, even if the agent took an action.",

View in src/sdk/jev.ts

Open on GitHub

More like this

Use cases this is tagged with