Skip to content

Decide which models and scenarios earn a place in the weekly run #5

Description

@leggetter

Deferred until a clean full matrix existed. That precondition is now met: the
13 August run is the first with no contaminated rows.

Open question 3 in the delivery plan: three models cost about $50 a week, five
about $150. The current cadence is weekly for the frontier agents plus the weak
pair and monthly for the -no-skills twins, at roughly $185 a month against a
$200 budget.

What the first clean run says, which was not knowable before:

  • Seven of fifteen scenarios pass every experiment. That is a floor and worth
    keeping, but it is also seven scenarios' worth of spend producing no signal
    each week.
  • The -no-skills twins now do produce signal, unlike when they were cut to
    monthly: the delta is +1, 0 and -3 depending on the model, not zero.
  • The weak model is where most failures live and is the cheapest arm to run.

Supersedes the reasoning in #1, which was written before any of this was
measurable.

Done when the cadence in .github/workflows/eval-refresh.yml reflects a
decision made against this data rather than against the estimates it was
originally set from.

Metadata

Metadata

Assignees

No one assigned

    Labels

    harnessRunner, scorers, provisioner and CI

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions