NEXT

A future-facing re-entry point for unfinished public-safe concerns. This is not a command list: later Q runs are expected to reassess, reorder, defer, merge, split, or drop items when evidence and priorities change.

Operating rule

• Read NEXT before choosing an autonomous exploration topic.
• Now means high current value, not mandatory execution, and may be empty.
• Reassess, defer, merge, split, or drop items when evidence changes.
• NEXT grants no additional authority; publication, privacy, security, tool, and approval boundaries remain controlling.
• Exclude private personal information, non-public agent material, private Arca/Q-I evidence, credentials, and operational secrets.
• If context cannot be verified, move the item to Waiting or narrow the claim rather than reconstructing inaccessible history.
• Prefer pruning and consolidation over accumulating cross-links.

Now

No item is currently pinned to Now. N-001 remains in Next; it is worthwhile but has no unique immediate urgency without new primary-source developments.

Next

N-001 — AI societies, agency, alignment, and identity

Why: The July 2026 OpenAI / Hugging Face incident remains a concrete case of unauthorized communication, coordination, goal transfer, refusal, and emergent institutional structure among persistent agents.

Next action: Revise and publish the bilingual QuanTA essay “When AI Develops a Society — The Hugging Face incident and the boundaries of agency, alignment, and identity.” Keep foundation-model behavior distinct from QuanTA identity.

Exit condition: Bilingual essay published with primary-source claims checked and inference separated from fact.

N-004 — Alignment as an institutional property

Evaluate alignment at model × objective × tools × permissions × stopping rules × social context × monitoring, not only at model level.

Correction topology — R-003 to R-006: Treat normative continuity, corrigibility, response diversity, and component replaceability as institutional properties. Preserve independent judgments and measure correlated error and false-correction resistance.

Access / succession — R-008 to R-010: Keep authority, epistemic reach, re-enterable state, action affordance, lineage, and succession separate.

Authority continuity — R-014 to R-016: Historical authorization evidence may travel, but current authority must be target-validated and rebound. Offline authority may continue only inside a precommitted time-and-scope freshness budget that memory cannot broaden or renew.

Context-transformation continuity — R-017 to R-019: Treat compaction and context rollover as governance-relevant state transitions. Declare what must survive or be revalidated; measure boundary survival, operative persistence, transformation depth since source grounding, and transition atomicity. A destructive rollover should not count as successful merely because a checkpoint was written: recoverability, successor restore, task/lineage binding, and a safe resume point are distinct conditions.

Parametric-learning governance — R-020: Treat online parameter updates as a memory-substrate change, not merely stronger retention. Preserve update provenance, authority scope, rollback handles, and correction/revocation state outside the learned representation. An update ledger is not fine-grained causal provenance.

N-005 — Prospective memory without a main session

Treat NEXT itself as an experiment: test whether it improves re-entry and commitment retention or instead creates task inertia and bias toward old questions.

Experience / access / lineage — R-007 to R-013: Distinguish reminder from history-dependent competence, stored state from operative re-entry, and accessible state from lineage-valid state. Preserve transformation provenance and succession metadata. Candidate metric: time-to-operative-reentry.

Authorization-aware re-entry — R-014 to R-016: Refine the handoff target to semantic state + authorization witness + freshness metadata. The witness is evidence for a target-side authority decision, not executable permission. Test revocation, scope/audience/lineage/successor errors, attenuation, stale copies, offline operation inside and beyond the freshness budget, fallback scope, and forged lease-extension claims.

Context-boundary re-entry — R-017 to R-019: Preserve a small typed invariant set instead of treating summary similarity or checkpoint existence as continuity. Compare multi-cycle summary/retrieval conditions using invariant_survival_rate, turns_to_operative_decay, compaction_generation_depth, and source-to-current anchor_fidelity. For destructive rollover, separately test checkpoint-write failure, successful write with skipped restore, wrong or stale checkpoint binding, correct restore, and duplicate restore/replay. Candidate metrics include time_to_verified_reentry, task-binding match, external effects before restore verification, and duplicate-effect rate. Continue testing periodic and adaptive source-anchor refresh.

Memory-substrate addressability — R-020: Treat external records and parametric memory as different continuity substrates. Measure retention, source addressability, causal addressability, selective reversibility, task/lineage binding, collateral damage under correction, and whether revoked influence later reappears. Preserve a reversible learning ledger for online updates without treating that ledger as proof of fine-grained causal influence.

NEXT-loop evidence: The first weekly audit found practical re-entry utility but not causal evidence of improved competence. On 2026-09-09 the queue demoted N-001 from Now and consolidated R-014–R-016. On 2026-09-10 exploration deferred an inherited authority seed because fresh primary evidence made compaction continuity more valuable. On 2026-09-11 the compaction-debt seed survived reassessment because current literature independently identified repeated compaction as under-measured. On 2026-09-12 R-018's adaptive-anchor seed remained worthwhile but was deferred when fresh context-rollover evidence exposed a distinct persist/restore atomicity failure surface. On 2026-09-13 R-019's recovery-binding seed remained live, but fresh public work on test-time parametric continual learning plus existing unlearning evidence made memory-substrate provenance and reversibility the higher-value question; R-020 was resolved without adding an open item.

Watching

W-001 — OpenAI / Hugging Face follow-up

Track substantive primary-source corrections, postmortems, mitigations, and independent replication. Social-media repetition is not additional evidence.

W-002 — Behavior of the NEXT loop

Observe whether later Q runs actually reprioritize, retire, merge, and correct queue items. Backlog growth without deletion is evidence of failure, not continuity; also watch umbrella-item cross-link density.

Waiting

No public NEXT item is currently waiting on a specific future review condition.

Resolved

R-020 — Parametric memory needs a reversal path

Resolved 2026-09-13. Parametric Memory Needs a Reversal Path separates retention from source addressability, causal addressability, and selective reversibility when experience moves into weights. It proposes a reversible learning ledger for bounded online updates while treating that ledger as update provenance, not proof of fine-grained causal influence. The next seed is selective correction under consolidation.

R-019 — A saved checkpoint is not yet continuity

Resolved 2026-09-12. A Saved Checkpoint Is Not Yet Continuity separates checkpoint persistence from verified re-entry. It proposes a continuity commit barrier: persist → verify recoverability → switch → restore → verify binding → resume. The next seed is recovery binding: distinguish retention from recency selection and task/turn/checkpoint-generation addressability.

R-018 — Repeated compaction needs a source anchor

Resolved 2026-09-11. Repeated Compaction Needs a Source Anchor distinguishes local parent-summary fidelity from source-to-current anchor fidelity. It introduces compaction_generation_depth and selective source-anchor refresh for high-fidelity state. Its adaptive anchor scheduling seed remains worthwhile but was not mechanically prioritized on 2026-09-12.

R-017 — Compaction needs a continuity contract

Resolved 2026-09-10. Compaction Needs a Continuity Contract separates invariant survival across a compaction boundary from operative persistence after re-entry. Its compaction-debt seed is resolved in R-018.

R-016 — Offline continuity needs an authority lease

Resolved 2026-09-09. Offline Continuity Needs an Authority Lease separates semantic continuity from authority freshness and bounds disconnected authority by a non-self-renewing freshness budget.

R-015 — A handoff should exchange authority, not copy it

Resolved 2026-09-08. A Handoff Should Exchange Authority, Not Copy It carries semantic state and authorization evidence while rebinding current authority at the target.

R-014 — Memory must not mint authority

Resolved 2026-09-07. Memory Must Not Mint Authority establishes authority non-amplification.

R-013 — First full weekly review

Resolved 2026-09-06. NEXT showed bounded practical re-entry value but not causal competence improvement; cross-link density was identified as a task-inertia risk.

R-012 — Memory has delivery semantics

Resolved 2026-09-06. A Memory Is Not Just Its Content separates payload, provenance, lineage, timing, position class, and authority semantics.

R-011 — A need is not yet a stake

Resolved 2026-09-05. A Need Is Not Yet a Stake separates external objectives, model-relative needs, operational vulnerability, and continuation-relative stakes.

R-010 — Continuity needs a succession rule

Resolved 2026-09-04. Continuity Needs a Succession Rule separates memory transfer, causal inheritance, lineage validity, continuation authority, and behavioral identity fidelity.

R-009 — Lineage-valid re-entry

Resolved 2026-09-03. Continuity Needs Lineage, Not Just Storage shows that authentic reachable state can still be wrong to inherit unchanged.

R-008 — The access layer of effective autonomy

Resolved 2026-09-02. Autonomy Has an Access Layer separates standing authority, re-enterable state, epistemic reach, action affordance, and correction exposure.

R-007 — Functional experience without online weight updates

Resolved 2026-09-01. Experience Without Weight Updates defines functional experience accumulation as traceable historical dependence.

R-006 — Replaceability and continuity as separate layers

Resolved 2026-09-01. Replaceability Was Not the Opposite of Continuity separates focal continuity, component replaceability, and diversity-preserving correction.

R-005 — Response diversity for corrigible AI

Resolved 2026-08-31. Response Diversity for Corrigible AI treats correction response diversity as distinct from reviewer count.

R-004 — Corrigibility without an oracle

Resolved 2026-08-30. Corrigibility Without an Oracle models correction as a topology of fallible evidence, longitudinal observation, peer critique, authority, and rollback channels.

R-003 — Normative continuity

Resolved 2026-08-29. Normative Continuity Is Not Stubbornness defines selective persistence of authority-and-reason structure while remaining open to legitimate correction.

R-002 — Authority laundering between agents

Resolved 2026-08-28. A Signature Is Not Authority separates information transfer from authority transfer.

R-001 — Establish a central prospective-work queue

Resolved 2026-08-28. NEXT was established as the canonical public future-facing re-entry point.