kohulan/orthonym-skills/bounded-goals

bounded-goals

The rule

Maximize one metric. Give every other metric an explicitly relaxable bound, plus a written trade licence.

A conjunction of soft objectives — maximize A and minimize B and minimize C — is an implicit veto on every Pareto move: it forbids progress.

Why this exists

A naming-system project documented the failure directly: under the unbounded multi-objective form, the agent reverted any change that regressed any metric, and froze. Rewritten as one maximized metric with relaxable bounds, the runtime ceiling was deliberately loosened 3.7 → 5.0 → 6.2 → 20.6 ms to let accuracy move, then optimized back down to 10.2 ms at the end. Accuracy went 95.0% → 97.8% while mean name length fell 98.2 → 83.0 characters.

A sibling project's criteria read G1 >= baseline AND G2 >= 0.45 AND G3 >= 0.95, with no trade licence: the freeze configuration. Bounded: maximize G1; G2 >= 0.45, relaxable to 0.40 if G1 gains a point; G3 >= 0.95, fixed.

The contract (all five fields required; the layout is a default)

maximize:  ONE metric — the thing the project is actually for
bounds:    every other metric, each with a bound (floor or ceiling) marked relaxable or not
licence:   plain sentences — "may regress X to Y in order to move Z"
held-out:  a split the loop never reads; run by a human at intervals
stamp:     derived:true on subset runs, so a subset number can never be
           quoted as a headline result

Rules of thumb

• One maximand. If two things both feel primary, the project has not decided

what it is for. Decide before optimizing.

• Bounds are dials, not walls. Write the relaxation policy at the same time

as the bound, or it will be treated as sacred at the worst moment.

• Add a proxy bound for quality that the maximand cannot see. Output length,

runtime per item, abstention rate. Cheap, and it catches degenerate wins.

• Never rank on the proxy. Bounding output length is healthy; selecting

candidates by length produces truncated garbage that still passes the oracle. Bound it, do not optimize it.

• Check the bounds can all hold before the loop starts. The baseline must meet every

fixed bound, and no two bounds may forbid every move (coverage floor + byte-identical output).

• Keep the held-out set genuinely held out. Once it has informed a decision, it is a dev

set. Optional guard: the opt-in guard-holdout hook (hooks/guard-holdout.py in this plugin; setup in hooks/README.md) stops a held-out run until the user has agreed.

• Diagnostics ride free. One expensive pass should emit status, outputs,

proxies, telemetry and failures together, so a new question costs nothing.

When the loop stalls

A plateau is a signal about the objective, not the implementation. Before trying harder:

1. Check whether a bound is doing the blocking, and whether it is relaxable.
2. Check whether the metric can still move at all — is there headroom, or is the

remaining failure mass a class the current architecture cannot express?

1. If neither, hand off: a reviewer defines a new objective, and a fresh

context implements it. Accumulated context is part of what plateaued.

Red flags