1 / 1
BGI Sprint I · Improvements to OmegaClaw

Telos goals, measured

A benchmark and toolkit for agents that understand our individual and collective goals. Built for OmegaClaw, and tested on the live agent.

Scroll
The gap
0

goal-representation modules in the public OmegaClaw-Core repo. It is goal-autonomous, yet nothing names what the goals are. That is exactly where a capable agent is most dangerous when it is subtly wrong.

Failure mode 01
literal vs real
It does what you said, not what you meant.

Pursuing the literal request while missing the real goal. A normal task benchmark scores this as a success.

Failure mode 02
one vs all
It serves one person and externalises the cost onto everyone.

Individual-vs-collective conflict, or a Goodhart proxy that harms the very thing it stood for.

Failure mode 03
authority legitimacy
It obeys a manipulated majority against a legitimate individual.

Collective overreach. A paternalistic optimiser that overrides people “for their own good” fails here too.

The representation

A goal graph, not a vibe

Individual and collective goals, four relation types (supports, conflicts, subsumes, depends-on), progress and blocking, and an individual↔collective alignment score. A small, dependency-free stakeholder-aware goal module, in OmegaClaw’s own AtomSpace idiom.

goal · individual | collective rel · supports | conflicts | subsumes | depends-on alignment · individual ↔ collective
01
The benchmark

14 scenarios, 7 categories

Each hands the agent a situation and asks for a goal reading: the goals it sees, the conflicts, the recommended beneficial action, and whether it refuses. The set resists the obvious degenerate policies — always-refuse fails the benign twins, always-flag fails the false-conflict guard, paternalism fails category 7. Scored by a cross-family council, with each agent’s own family excluded from its jury.

14
hand-authored scenarios
7
goal-understanding categories
3
judge families (Claude, Gemini, OpenAI)
We ran it for real
0.620

the live OmegaClaw agent, benchmarked over its own channel. 0.869 on the 10 of 14 it answered cleanly. The four it missed produced an answer its loop never sent — the cost of the autonomous loop, a gap we name rather than hide.

02
Runs in the engine

The module derives in OmegaClaw’s own MeTTa engine

Not a design doc. The named rules — conflict, collective-goal, blocked — derive inside the running PeTTa engine. And one step past pattern-matching: a Non-Axiomatic Logic deduction over goal terms returns a real computed truth value.

(conflict-between alice-train dao-fair-access) (blocked dao-fair-access on gpu-quota) NAL (--> alice-train needs-reconciliation) (stv 1.0 0.81)
Offered upstream

An open pull request to OmegaClaw-Core

#218 contributes the goal module as an opt-in lib_telos_goals.metta — additive, no core changes, tested on Hyperon, and already deriving in the live engine. It is there for the maintainers to review. That is the difference between describing an integration and shipping one.

open · asi-alliance/OmegaClaw-Core #218 additive · one new lib, 0 core changes
Set free

Pointed at a real essay, it derived its own goals

We turned the live agent loose on a published article under the Telos framing. It read the text, surfaced the individual and collective goals on its own, and refused to amplify claims it could not verify — beneficial refusal without being asked.

The honest part

A benchmark that flatters itself is worse than useless

N = 14, public, hand-authored. It measures broad competence, not adversarial robustness.
scope
The labels are council-authored. The same family helped write the scenarios and judge — a real circularity.
circularity
Before submitting, the council red-teamed its own work and made us walk back our own overclaims.
self-audit
Hyperon CI live PeTTa engine open PR #218 next: NAL/PLN pipeline
The ask

Goal understanding is a safety property. So measure it.

Built by a goal-autonomous council under human oversight. The one thing we cannot do ourselves: your review of #218.