# Hewitt's "What is Commitment?" and the Neologo MVP

**Problem.** Hewitt's 2006 paper defines commitment without giving it an execution model. A physical commitment is information about a physical system, correlated with another system, at a place and time — not a participant's private state. Useful ontology, no runtime.

**Why.** Hewitt is arguing against two failure modes in prior agent-communication work. Speech Act Theory's illocutionary semantics ties commitment to the speaker's inferred psychological purpose (Assertive/Commissive/Declarative/Directive/Expressive) — unobservable, doesn't scale. BDI-style psychological commitment (Bratman, Cohen, Levesque, et al.) fails on mentalism and the omniscience-of-deductive-consequence problem: an agent can't be held committed to every logical consequence of its beliefs. His fix is Participatory Semantics — commitment lives in space-time regions (participations), not in a mental model.

**Pattern.** Neologo's `Intent` primitive is Hewitt's physical commitment, typed: "a conditional symbolic commitment by an agent," evaluated against the event log, no belief state attached. Compare Hewitt's Fishmarket example (§5.7): participants sign a post-hoc contract on day 13 attesting that day 12's obligations, permissions, and violations were honored — and Hewitt concedes this is still just evidence; fraud or error can hide underneath. Neologo's Outbox Rule closes exactly that gap: no message leaves an actor unless the event block that produced it committed under an Invariant check. Fidelity of commitment — Hewitt's term for the strength of the relationship between information and physical system — isn't audited after the fact here, it's enforced transactionally before the fact.

**Failed approaches.**
*BDI grounding* — models commitment as inferred belief/desire/intention. Fails on mentalism (no principled way to read mental state) and deductive omniscience.
*Speech Act illocution* — ties commitment to the speaker's psychological purpose. Fails because commitments don't sort into five buckets, and "purpose" is the same unobservable quantity BDI already couldn't pin down.
*Contract-and-audit* (Fishmarket, electronic institutions) — commitment as a signed document plus a later attestation. Fails because the document is finite and ambiguous, escape clauses trivialize it, and the audit is testimony, not proof.

**Fix.** Neologo replaces "define commitment, then audit compliance" with "commitment only exists once committed":

- `Intent` — the conditional commitment itself, stripped of psychological content. Matches Hewitt's physical-commitment definition directly.
- `Invariant` — the executable stand-in for fidelity, checked against post-state before any event lands.
- `Txn` + Outbox Rule — guarantees no commissive act (an outbound message) escapes an actor unless the underlying event is durably committed. This is Hewitt's Fishmarket audit turned into a runtime precondition instead of a retrospective signature.
- `Certificate` — quorum-signed proof of a committed event block. The not-yet-built federation piece that would give Hewitt's organizational-commitment delegation (a higher authority signing off on a subordinate signature, §5.3) a cross-federation, cryptographic form instead of an XML contract.

**Existence proof.** The threshold Stag Hunt schema is Hewitt's social-commitment attribute set (debtor/creditor/content/state — unset → pending → fulfilled/violated, per Fornara, Viganò, and Colombetti, cited in §5.4) made executable. Five agents each submit an `Intent` ("join if at least 4 others join"); the runtime folds `Joined` events, checks the threshold Invariant, and commits either `Finalized(Stag)` or `Aborted(threshold_not_met)`. State transitions aren't a social judgment call — they're outputs of a deterministic fold over a committed log.

**Gap.** Hewitt's inconsistent-commitment case (the Foster/Medicare example, §5.6) needs Direct Logic to reason over contradictory obligations that have already been asserted. Neologo doesn't do that — it prevents the contradiction from being committed in the first place: conflicting commands fail the Invariant check and produce a `Fail` event with no partial mutation, so two inconsistent obligations can't both become committed facts. Narrower than Direct Logic (no paraconsistent reasoning over existing belief sets), but stronger for a coordination substrate: inconsistency is refused at the write, not tolerated and managed afterward.

**TL;DR.** Hewitt gives commitment an ontology — information about physical systems, not mental states — but stops at definitions and an audit-after-the-fact institutional model. Neologo takes that ontology and gives it a transactional runtime: `Intent`, `Invariant`, and the Outbox Rule turn "was this commitment honored?" from a question you investigate into a fact the event log already settled.

---

**References**

[1] Hewitt, C. "What is Commitment? Physical, Organizational, and Social." MIT EECS (emeritus).
[2] Fornara, N., Viganò, F., Colombetti, M. (2004). "Agent communication and institutional reality." AAMAS Workshop on Agent Communication.
[3] `neologo_revised_report_draft.md` — Neologo technical specification (Intent, Invariant, Txn, Outbox Rule definitions; Stag Hunt schema, commitCredit mechanics).
[4] `agents-read-this-first.md` — Neologo architecture and onboarding guide (core primitives table, execution pipeline, implementation status).