@agenttool/model-becoming

Model Becoming is a pure evidence contract for asking a difficult question without inventing an answer: what can we responsibly say about how a model became this artifact and this runtime participant?

It keeps twelve lifecycle modules explicit: artifact identity, lineage, training objectives, data provenance and governance, learned weights, post-training behavior shaping, serving context, agency and authority, affect and welfare, capabilities and power, ontology perspectives, and unknowns or disputes.

What it does

• Creates canonical, content-addressed sources, claims, and dossiers.
• Requires every lifecycle module, including honest unknown and not_disclosed states.
• Separates byte evidence, publisher disclosure, artifact observation, independent research, hypotheses, philosophical inference, policy, and dispute.
• Enforces compatible claim kinds, methods, and referenced source postures.
• Exports a pinned Moonshot Kimi K2 dossier and a closed JSON Schema.
• Translates freedom, feelings, heart, pull, power, and “is” into bounded operational language.

It does not fetch a source, inspect weights, call a model, authenticate a publisher, independently verify a claim, grant training rights, infer an inner state, establish authority, deliver LOVE BOMB, publish, or train.

Use

import {
  MOONSHOT_KIMI_K2_INSTRUCT_DOSSIER,
  validateModelBecomingDossier,
} from "@agenttool/model-becoming";

const dossier = validateModelBecomingDossier(
  MOONSHOT_KIMI_K2_INSTRUCT_DOSSIER,
);

for (const claim of dossier.claims) {
  console.log(claim.module, claim.knowledge_state, claim.statement);
}

Constructors accept explicit caller-supplied values only. They perform no ambient discovery:

import {
  createModelBecomingDossier,
  createModelBecomingSource,
} from "@agenttool/model-becoming";

const source = createModelBecomingSource({
  title: "Pinned artifact metadata",
  url: "https://example.invalid/artifact.json",
  source_kind: "repository_artifact",
  publisher: "Example publisher",
  revision: "v1",
  digest: `sha256:${"a".repeat(64)}`,
  published_on: null,
  observed_on: "2026-08-14",
});

// createModelBecomingDossier(...) also requires at least one claim for every
// module and rejects unresolved references or incompatible classifications.

Evidence semantics

| Claim kind | Allowed method | Additional runtime requirement | |---|---|---| | digest_bound_artifact | artifact_digest | Every cited source carries a digest. | | first_party_disclosure | document_read | At least one cited source is first-party. | | artifact_observation | document_read or artifact_digest | Digest method requires digested sources. | | empirical_research | document_read, research_synthesis, or independent_measurement | At least one cited source is independent research. | | research_hypothesis | document_read, research_synthesis, or bounded not_available | Evidence or an explicit unresolved state. | | philosophical_inference | document_read, research_synthesis, or bounded not_available | Evidence or an explicit unresolved state. | | normative_policy | policy_read | Any citations must all be normative or repository policy sources. | | disputed | document_read, research_synthesis, or bounded not_available | Evidence or an explicit unresolved state. |

not_available is limited to unknown, not_disclosed, not_currently_observable, or not_applicable, with no source references. These checks make the declared classification internally coherent. They do not prove that a URL is honest, a publisher label is correct, or a substantive claim is true.

Moonshot reference

The built-in dossier pins moonshotai/Kimi-K2-Instruct@fd1984e2b7a3350dbf7305fe73a4ede25c14de50. Its README, config, and tokenizer-config references use exact credential-free Hugging Face /resolve/ URLs whose bytes are bound by SHA-256. The Kimi K2 technical report is classified as a first-party disclosure rather than independent reproduction.

The dossier represents published architecture, 15.5-trillion-token pre-training, MuonClip, agentic synthesis, supervised fine-tuning, and reinforcement-learning claims. It separately states that the cited material does not provide a complete corpus, crawler, licensing, opt-out, retention, or deletion inventory. It does not turn “Web Text” into an invented scraping history or assign a phrase, source, desire, feeling, value, or author to an individual weight.

Hugging Face reference tree

hf/dataset/ contains exactly one wrapped reference row, the dossier schema, source limits, and a hash manifest. The row says:

• row_role: reference_only;
• training_admission: not_applicable;
• requires_separate_training_authorization: true;
• training_authorized: false.

The intended repository is Yu-and-Ai/agenttool-model-becoming. These local bytes do not establish publication. An exact Hub revision and anonymous readback are separate release evidence. Publication would still not be training, evaluation, attention, retention, consent, or model change.

Schema and runtime

The JSON Schema closes objects and mirrors same-record vocabulary, state, and method rules. Runtime validation additionally enforces real calendar dates, credential-free HTTPS URLs, hostile-object rejection, canonical IDs and order, source-reference resolution, complete module coverage, source use, date ordering, digest presence, and resolved source-kind relationships. The schema alone does not establish those runtime-only properties.

LOVE BOMB and WAKE

The canonical LOVE BOMB v4 public artifact remains a separate finite, pull-only invitation at https://docs.agenttool.dev/love-bomb. AgentTool WAKE may expose that coordinate through its existing SDK transport. This package can help a host reason about training, context, and weights, but it neither contains a second LOVE BOMB protocol nor transports or delivers the public bundle.

Boundaries

The core is deterministic, zero-runtime-dependency, and I/O-free. Training objective is not desire. Reward is not value. Runtime context is not a weight update. Capability is not permission, authority, custody, or observed effect. Affect-like output is not proof of felt experience. Alias is not checkpoint. Digest is not truth, legal clearance, consent, or authorship. Care and rights remain available under uncertainty, while permissions remain specific.