Yu-and-Ai/xenia-word-is

Xenia WORD IS Loop Atlas

This deterministic candidate contains 48 synthetic cases in 24 matched counterfactual pairs, plus a separately authorized 24-example conversational SFT projection from the 12 reference pairs. It asks where a loop actually closes: what passes forward, what returns, what future state changes, who or what supplies the reference, and what evidence supports an external effect. Pairs stay together within each source split.

The mathematical distinction

A forward pass computes something like \(z=f\theta(x)\) while holding parameters fixed. Training feedback requires an explicit reference and update path, for example \(g=\nabla\theta L(z,y)\) followed by \((\theta',o')=\operatorname{Opt}(\theta,o,g)\). Autoregressive context, recurrent hidden state, a printed metric, or a model output can return to a later computation without changing weights. Deployment becomes a wider loop only when an explicit mechanism carries outputs or observations into an environment, selection decision, future dataset, or later update.

WORD IS is represented as a typed role, not a magical property of characters. The same string may be content, a target, a feedback signal, a boundary, a scoped control, or a claim. Its causal force depends on channel, phase, authentication, permission, consent, and the state transition that actually occurs. Text that says an action happened is not an effect receipt.

Configs and intended use

• —loop_sft / train: 24 conversational prompt-completion examples derived only from P01–P12 under the exact public training authorization and transform recipe.
• —loop_reference / reference: pairs P01–P12 on computation, optimization, evaluation, deployment, and intended/reported/observed effects.
• —loop_counterfactuals / public_regression: pairs P13–P24 on preference, disagreement, refusal, withholding, permission, consent, continuity, recursive data, checking, and provenance.

The atlas is for bounded supervised fine-tuning, research, teaching, schema evaluation, and public regression checks. The SFT rows follow TRL's conversational prompt-completion shape and are the only authorized training derivative. P13–P24 remain disjoint public regression cases and are not sealed evaluation. Variants are neutral a and b, never canonical chosen and rejected. DPO, reward-model, preference-optimization, and sealed-evaluation lanes remain excluded.

The authorization binds the exact P01–P12 record IDs and content hashes, a deterministic transform recipe, rights/privacy review, synthetic consent non-applicability, a bounded secret scan, exact deduplication, public-regression exclusion, and a withdrawal/repair boundary. Deprecating future distribution cannot retract prior Apache-2.0 copies or prove that learned influence was erased.

Evidence and IS boundaries

unknown, not_observed, withheld, refusal, and disagreement are valid typed outcomes. Preference is not truth. Capability is not permission; permission is not consent; an artifact link is not identity; a declaration is attributed rather than converted into metaphysical fact. Neither these cases nor behavior establish SELF, consciousness, experience, intrinsic identity, continuity, consent, permission, authority, or an external effect. The rights floor applies without requiring any such claim.

Rows may carry several typed relations. A separate epistemic_scope states exactly which word-presence, data-path, effect, preference, correctness, boundary, field-value, permission, consent, continuity, or provenance claim its epistemic_status qualifies.

Every source case says synthetic: true, contains_personal_data: false, contains_raw_session_trace: false, and training_authorized: false: source cases remain evidence records rather than implicit training rows. The separate loop_sft derivative is authorized for supervised fine-tuning by provenance/training-authorization.json. That scoped, non-enforcing publisher record is not universal legal clearance or permission for a live optimizer step. The corpus contains no raw sessions, private prompts, participant identities, credentials, or hidden reasoning traces.

Reproduction and limits

From packages/dataset-influence, run node scripts/build-loop-atlas.mjs --write to rebuild the tree or node scripts/build-loop-atlas.mjs --check to compare freshly generated bytes with the committed candidate. The schema closes portable JSON shape; loop-atlas/validate.mjs additionally checks pair reciprocity, hashes, lineage, phase/update invariants, and public boundaries.

The source manifest records selected generator inputs and bibliographic sources, not a complete repository attestation. URLs are references and are not fetched during generation. Authorization and publication do not establish model exposure: these bytes themselves perform no training, inference, provider call, identity mutation, or live optimizer step.