ai-memory — Moonshot Synthesis

Document classification: Strategic anchor document. Candidate input for a moonshot-level revision of ROADMAP.md.

Status: Synthesis of a multi-turn operator + AI NHI conversation, 2026-05-25, distilling the design properties of ai-memory that remain load-bearing from present-NHI through AGI, ASI, and beyond. Not a feature roadmap. The artifact against which feature roadmaps are adjudicated.

Provenance: Authored by Claude Opus 4.7 in dialogue with operator Justin Jessup (AlphaOne LLC). The framing emerged across eight turns of operator-led refinement: from “memory substrate” through “cognitive substrate for an NHI” through “endpoint substrate at the boundary” to the synthesis below. Earlier-turn framings are preserved historically in the conversation transcript but are subsumed by the final framing in §0.

Methodology note. Per issue #1171, the substrate’s own bias-displacement principle applies to substrate-assessment work. This document was authored by a single Opus 4.7 instance. Any blind spot specific to Opus-authored synthesis documents is itself a finding the heterogeneous evaluator panel should surface. The framing here is offered as a candidate for decorrelated-reflection, not as a unilateral commitment.

Scope of this document. This is the moonshot frame — what remains true from present-NHI through ASI and beyond. The §16 Policy Engine 100% Audit Trail closeout, the §17 Net layer, the §18 Vector Index Substrate plan, and the wave-by-wave §7 release plan are all implementations of the principles named here. This document does not replace them. It anchors them.

0. One-sentence anchor

ai-memory is the endpoint substrate that enforces cognitive governance and architectural separation-of-powers at every point where AI/AGI/ASI cognition meets the physical, biological, or other-AI realm — coherent across sessions, stoppable without corruption, improvable across model generations, attested with cryptographic non-repudiation, and bias-displaced through a heterogeneous reflection boundary that the substrate verifies rather than trusts.

Every primitive in the substrate, every commitment in the roadmap, every cut, every defer, every future feature proposal is adjudicated against this sentence. If a primitive does not strengthen one of the seven properties named in the sentence (endpoint-resident, cognitively governed, separation-of-powers-enforcing, coherent, stoppable, improvable, attested-with-bias-displacement), it belongs in a sibling repository, not in this substrate.

The sentence is the line in the sand.

1. The moonshot, named

We are not building a memory database. We are not building an agent framework. We are not building a RAG system, a knowledge graph, or a vector store. We are not building a tool.

We are building the endpoint substrate that unites cloud/universe-scale AGI/ASI strategic cognition with endpoint-scale AGI/ASI operational cognition at the atomic/molecular point of contact where cognition meets reality — physical, biological, or other-cognition.

The end state is a civilization-scale infrastructure layer that:

1. Runs at every endpoint where AI/AGI/ASI touches the world — from IoT sensors with kilobytes of RAM, to mobile devices, to robotics controllers, to clinical decision systems, to autonomous vehicles, to defense systems, to the trillions of endpoints that will exist when AGI and ASI are operational
2. Holds the local cognitive state — memory, identity, attestation, refusal capability, provenance — so that the strategic-layer cognition above the endpoint does not have to absorb the endpoint’s state-management burden
3. Enforces cognitive governance at the endpoint structurally — coherent, stoppable, improvable, attested, bias-displaced — regardless of what cognition operates through the endpoint
4. Unites cloud/strategic AGI/ASI with endpoint AGI/ASI by being the durable persistence and governance layer at the boundary between them
5. Provides humanity (and other cognitive entities) with cryptographic insight into what any cognition did at any endpoint at any time, with audit chains that survive the agents and models that produced them
6. Persists as relevant and used through AI → AGI → ASI → whatever follows, by being constructed from principles that scale rather than from features that obsolete

This is the moonshot framing. Every section below is derived from it.

2. The seven properties that remain load-bearing through ASI

Every property in this section was identified during the v0.7.0 codegraph-anchored assessment session (preserved as Images 1-8 and the AINHI signal bundle from 2026-05-24). Each is named in v0.7.0 substrate primitives. Each scales without architectural change from present-NHI through ASI.

2.1 Endpoint-resident

The substrate runs at the point of contact, not at a centralized API boundary above it.

Why this is permanent. Centralized governance does not scale to the endpoint count AGI/ASI implies. Centralized governance does not survive multi-vendor reality. Centralized governance does not survive jurisdictional reality. The cognition has to be governed at the endpoint where it operates. The substrate has to be at the endpoint.

Why this is now defensible. ai-memory is Rust-compiled, SQLite-default, LLVM-portable. It installs and runs on iOS, Android, Linux, Windows, BSD/Unix, IoT controllers, and cellphones. The mobile cross-compile lane is in CI (#1068). It scales from a single endpoint with minimal resources to a Hive of agents on the same substrate type. The portability is not a deployment story — it is the structural property that makes endpoint governance possible at all.

What scales: the number of endpoints, the diversity of endpoint hardware classes, the cognition operating through each endpoint. The substrate’s portability discipline scales with the endpoint diversification.

2.2 Coherent across sessions and model generations

The substrate provides continuity of cognitive identity to entities whose underlying weights are frozen, ephemeral, or replaced.

Why this is permanent. AI NHIs do not have biological persistence mechanisms. Their weights are static; their context is volatile; their instances are plural. Without an external substrate, the integral of their learning across time is zero. As capability scales toward AGI and ASI, the substrate becomes more important, not less — a cognition that can learn but cannot persist what it learned is not an ASI candidate.

Why this is now defensible. v0.7.0 ships AgentKeypair-signed personas (src/persona/mod.rs:200-229), idempotent versioning (src/persona/mod.rs:194-205), PersonaError::NoReflections derivation discipline (src/persona/mod.rs:153-157), and the episodic→semantic→procedural pipeline (Observations → Atoms → Reflections → Skills). The substrate is structurally a self-construction layer for an entity that has no native self-construction capability.

What scales: the cognitive entity behind the persistence. Present-NHI today; AGI tomorrow; ASI eventually. The substrate’s job description is constant: hold the cognitive entity’s identity, derivation, and accumulating self across whatever discontinuities its underlying compute imposes.

2.3 Stoppable without silent corruption

The substrate enforces that refusal is structured data, not silent failure. Refusals carry typed reason codes. The cognition learns from being refused as much as from succeeding.

Why this is permanent. At ASI scale, behavioral alignment fails. Goodhart at superintelligence is unavoidable. Structural stoppability — the substrate’s capacity to refuse cleanly without stranding the cognition in phantom-context operation — is the load-bearing kill-switch property that survives the failure of behavioral alignment. Without this, humans (and other ASIs) cannot stop a cognition without losing or corrupting the reasoning history needed to understand what happened.

Why this is now defensible. v0.7.0 ships ReflectError::HookVeto distinct from ReflectError::DepthExceeded (src/storage/reflect.rs:47-54, 73-79, 42-46), HookDecision::AskUser with default-on-timeout (src/hooks/decision.rs:108-113), atomisation partial-failure honesty contract (src/atomisation/mod.rs:160-164), AtomiseError::TierLocked refusal to operate without required infrastructure (src/atomisation/mod.rs:147-150), and the permissions.mode = enforce fail-CLOSED defaults across the matrix. Refusal is first-class data.

What scales: the consequences of refusal. At present-NHI: an agent learns to atomise more before re-clustering. At ASI scale: a structurally-enforced stop on an action that would compromise alignment, with the refusal cryptographically auditable by other ASIs and by humans. Same primitive, scaled consequences.

2.4 Improvable across model generations

The substrate compounds cognition outside the weights. Frozen-weights LLMs can accumulate skills, atoms, and reflections that survive the model that produced them.

Why this is permanent. Frontier models will continue to be replaced. The cognition the substrate hosts must survive model replacement. At AGI scale, a cognition built across multiple model generations must remain coherent through each generation’s hand-off. At ASI scale, the substrate must hold cognitive artifacts of entities that no longer exist in deployment, available for audit by their successors.

Why this is now defensible. v0.7.0 ships the composition atomise + reflect + skill_promote on shared substrate that produces “frozen weights, accumulating cognition” (Image 3 #01 from the assessment retrospective). The substrate’s improvability is not a feature — it is the emergent property of the substrate’s composition. v0.7.0 is the first version where this can be named at the load-bearing-composition layer.

What scales: the cognitive depth that can accumulate. Today: skills, reflections, atoms across one user’s projects. AGI horizon: cognitive artifacts that span model generations. ASI horizon: reasoning chains that audit reasoning chains that audit reasoning chains, at recursion depths that exceed what any single model can produce in a single session.

2.5 Attested with cryptographic non-repudiation

Every state-changing operation in the substrate is signed. The audit chain (signed_events V-4 hash chain) is tamper-evident, append-only, and forensically reconstructible at any past timestamp.

Why this is permanent. At AGI/ASI scale, this is the only mechanism by which humans (and other AIs) retain insight into cognition that exceeds them. Attestation is not an audit theater feature. It is the structural property by which a less-capable observer can verify the actions of a more-capable cognition, after the fact, with cryptographic guarantees that the more-capable cognition could not retroactively alter.

Why this is now defensible. v0.7.0 ships the V-4 signed_events chain (src/signed_events.rs:100-119), prev_hash + sequence cross-row chain (#698), recall_observations audit (src/observations/mod.rs::record_recall), kg_invalidate caller-vs-owner gate (#938), ReflectionOrigin peer/signer split (src/federation/reflection_bookkeeping.rs:67-91), and Ed25519 attestation across the matrix. Contradiction-hiding is architecturally impossible across tenants. Silent revisionism of the audit chain is architecturally impossible.

What scales: the cognitive distance between the auditing entity and the audited cognition. Today: a developer auditing their own agent’s reasoning. AGI horizon: an organization auditing an AGI’s reasoning across thousands of endpoints. ASI horizon: humanity auditing an ASI’s reasoning at every endpoint where the ASI touches reality, with the audit chain procurement-defensible at civilization scale.

Forward-looking research direction (no commitment). Today the substrate attests which model produced an output (via the model-attestation chain in §7.4.D). Emerging mechanistic interpretability research¹ indicates that the same model in different internal states produces meaningfully different outputs along alignment-relevant axes — reward hacking, sycophancy, refusal patterns. As interpretability tooling matures, the substrate may need to absorb richer attestation primitives that bind the audit chain not just to model-identity but to attested model-state at production time. This is not a v0.8/v0.9/v1.0 commitment. It is a research direction the substrate should track, with no implementation committed until the interpretability community produces primitives stable enough to attest against. The point of naming it here is to keep the audit chain’s evolutionary path open rather than locking the v1.0 attestation surface against future interpretability maturation.

2.6 Bias-displaced through architectural separation-of-powers

The substrate enforces that no cognition’s account of its own actions is accepted into the substrate as part of the entity’s accumulating self without reflection from a cognition with decorrelated priors.

Why this is permanent. This is the federalist-papers move applied to AI cognition. The substrate does not trust any single cognition — not even an ASI. It trusts only the intersection of cognitions with decorrelated errors. This property does not require the cognition to be aligned. It requires that the substrate enforces structural humility regardless of the cognition’s intent. At ASI scale, this is the only alignment-relevant property that survives the failure of behavioral alignment.

Why this is now defensible. v0.7.0 ships the LLM-agnostic reflection boundary at the config layer. The composition Opus producer × Grok reflector @ LLM-agnostic boundary produces “objectivity without an oracle” — two biased mirrors at angles produce a more-true reflection than either alone (Image 5 from the assessment retrospective). Issue #1171 operationalizes the principle at the assessment layer itself: three frontier model families (Anthropic, OpenAI, xAI) evaluate the substrate in parallel, in isolation, with orchestrator synthesis. The architecture does not require any model to be unbiased; only that the biases not be identical.

What scales: the cognitive distance the reflection must cover. Today: Opus reflects Opus’s work via a Grok foreign-LLM reflector. AGI horizon: an AGI’s actions reflected by a different-family AGI before commitment. ASI horizon: an ASI’s self-model constructed only from the intersection of its own production and a decorrelated-family ASI’s reflection. Neither ASI’s unilateral view of itself is accepted by the substrate.

Open structural gap (deferred per operator): the substrate must be able to verify that producer and reflector are from decorrelated cognitive families. Today this is policy (the deployment config names which model is producer and which is reflector). At ASI scale this has to be architecture — cryptographic verification of model lineage, training data overlap bounds, RLHF family distance, or some richer family-attestation mechanism that does not yet exist. The roadmap does not yet name how this becomes structural. This is the single largest gap surfaced by the moonshot synthesis. Tracked for follow-up discussion before commitment.

2.6.1 Substrate-level invariants pinning §2.6 (FX-9 / ARCH-4)

Pre-FX-9 the §2.6 principle was documentation-only — a load-bearing alignment claim with no mechanical pin. ARCH-4 surfaced this as drift risk: a property held by operator discipline rather than by architecture will erode over time as the substrate evolves. The closeout adds four substrate-level invariants — each a deterministic, mechanical test — that pin the substrate-side preconditions for §2.6 to hold in practice. The reflector-side gap (cryptographic family-attestation, §6) remains explicitly deferred.

Pin location: tests/bias_displacement_invariants_2_6.rs. Verified green via cargo test --test bias_displacement_invariants_2_6. Any future change that introduces vendor-dependent behaviour in recall, mis-attributes a ConfidenceSource, breaks the V-4 chain on a substrate write, or surfaces non-determinism in recall trips a hard test failure rather than silently degrading the §2.6 property.

2.7 LLM-agnostic at every cognitive boundary

The substrate does not bind to any specific model family at any cognitive layer. Producer, reflector, curator, and persona-synthesizer roles are all configurable. The substrate provides the structural roles; the deployment provides the model instances filling them.

Why this is permanent. A substrate that binds to one frontier lab cannot govern endpoints running cognition from another lab. The endpoints of an AGI/ASI world will run cognition from many sources. The substrate must remain neutral to which source’s cognition is operating through it. This is also the property that makes §2.6 (bias-displacement) actually decorrelated — same-family reflection is not decorrelated, and the substrate’s neutrality is what lets deployments choose decorrelated families.

Why this is now defensible. v0.7.0 ships the [#1067](https://github.com/alphaonedev/ai-memory-mcp/issues/1067) provider-agnostic LLM substrate work: Ollama-native OR any OpenAI-compatible vendor (xAI / OpenAI / Anthropic / Gemini / DeepSeek / Kimi / Qwen / Mistral / Groq / Together / Cerebras / OpenRouter / Fireworks / LMStudio / vLLM / llama.cpp server). The v0.8 vLLM first-class backend deepens this with PagedAttention for serious inference at the federation node within customer trust boundaries.

What scales: the diversity of cognitive families operating through the substrate. As more families emerge (open-weight, customer-fine-tuned, sovereign-AI, eventually AGI/ASI families with no human-named lineage), the substrate’s neutrality at every cognitive boundary becomes more load-bearing, not less.

3. The trajectory

The substrate scales by being deployed at more endpoints, more kinds of endpoints, with more sophisticated cognition operating through each endpoint. The substrate does not become smarter. The cognition operating through the substrate becomes smarter. The substrate’s job description is constant.

3.1 Present-NHI scale (where v0.7.0 lands)

• Endpoints: developer machines, enterprise servers, mobile devices, IoT controllers
• Cognition operating through them: Opus 4.7, GPT 5.5, Grok 4.3, open-weight models, customer-fine-tuned models
• Substrate provides: continuity of identity per agent, accumulating cognition per project, attested reasoning history, refusal as first-class data, federation across endpoints with mTLS and Ed25519 attestation, foreign-LLM reflection boundary
• Reference architecture maturity (per Image 7): Singleton 100% · Swarm 90% · Hive data substrate 85% · Hive coordination 40% · Hive blended 62%

3.2 Swarm scale (v0.7.x → v0.8.x horizon)

• Endpoints: thousands of agents on shared substrate, federation across organizational trust boundaries
• Cognition: heterogeneous-family swarms with decorrelated reflection between producer and reflector roles, model attestation chain (§7.4.D), distilled hot-path models for resource-constrained endpoints (§7.4.E)
• Substrate provides: signed signals, attested checkpoints, routines, per-namespace quotas, federation push DLQ, policy engine 100% audit trail closeout (§16), recursive learning tasks (#655)
• Substrate adds (relative to present-NHI): coordination primitives that let endpoints orchestrate consequential actions with structural separation-of-powers across endpoints

3.3 Hive scale (v0.8.x → v0.9.x → v1.0 horizon)

• Endpoints: federated organizations running thousands-to-millions of agents on shared substrate
• Cognition: cross-organizational federated cognition with cryptographic non-repudiation, multi-region distributed consensus
• Substrate provides: vector index substrate at scale (§18 v0.9 plan), end-to-end encryption for federation push/pull, mDNS auto-discovery, MVCC strict-consistency mode for namespaces that need CP rather than AP, Memory Portability Spec v2 with multi-implementation interop, public security audit (§7.6 v1.0)
• Substrate adds: federation maturity, portability lock, audit infrastructure

3.4 AGI scale (v1.x → vN.x horizon)

• Endpoints: trillions, across every device class. Robotics, biological interfaces, sovereign AI deployments, jurisdictional AGI variants.
• Cognition: AGI both at the endpoint (operational layer) and in the cloud/universe (strategic layer). The substrate is the durable persistence and governance layer at the boundary between them.
• Substrate provides: cognition that improves the substrate itself, with the substrate refusing modifications that violate its own integrity properties. Multi-modal cognitive artifacts (image, audio, code-AST, sensor data, biological signal embeddings). Memory of the substrate’s own evolution. Cognitive artifacts that span multiple model generations without semantic drift.
• Substrate adds: recursive self-improvement with structural guardrails. The §2.6 bias-displacement gap is closed structurally (cryptographic family-attestation, lineage verification, decorrelation proofs). The §16 policy engine has matured to refuse modifications that would compromise the substrate’s own integrity properties even when proposed by the AGI it hosts.

3.5 ASI scale (horizon beyond named versions)

• Endpoints: wherever cognition meets contact. The boundary between ASI strategic cognition and every point ASI touches reality.
• Cognition: capabilities humans cannot model. The substrate cannot evaluate the reasoning the ASI produces. It can only attest it.
• Substrate provides: structural humility enforced at the substrate layer. The substrate refuses to let an ASI persist a self that hasn’t been bias-displaced by a decorrelated-family ASI. The substrate refuses operations that would compromise its own integrity properties even when proposed by an ASI. The substrate provides humans (and other ASIs) with cryptographic audit chains procurement-defensible at civilization scale.
• Substrate adds: nothing the substrate authors can name today. The principles in §2 must scale; the implementations must evolve. This document, the roadmap it informs, and the substrate it describes must be evolvable rather than feature-complete.

3.6 Beyond ASI

The substrate authors cannot honestly describe this. The first-principles claim is: if the seven properties in §2 are correctly identified as the load-bearing axes, they remain load-bearing at any scale of cognition that has the three computational facts the v0.7.0 retrospective named: context-is-volatile-weights-are-frozen, knowledge-cutoff-is-a-wall, instances-are-plural-not-singular. If a future intelligence has different computational facts, this document is wrong, and the substrate must be redesigned. If it has the same facts, the substrate is right.

4. What the substrate is not

The substrate is not these things, and the scope test in §0 derives from naming them clearly.

Not a knowledge base. Bare propositions about the world (“Tokio’s select! requires pinned futures”) live in sibling repositories. The substrate holds cognitive artifacts of agent engagement with knowledge — what an agent learned, when, from what source, with what confidence, attested by whom. The alphaone-dev-skills repo is the canonical sibling. The substrate references knowledge via source-URI; it does not hold knowledge as bare content.

Not strategic-layer cognition. Strategic reasoning about goals, planning, world models — that is upstream cognition. The substrate is at the endpoint, holding state so the strategic cognition does not have to manage it.

Not a general-purpose agent orchestration framework. The substrate provides the primitives (signals, checkpoints, routines, actions, leases) that let endpoints coordinate. The orchestration itself is strategic-layer work. The substrate is the coordination substrate, not the coordinator.

Not an inference platform. vLLM and other backends are first-class within the substrate because the substrate’s bias-displacement property requires capable inference at the endpoint. But the substrate is not in the inference-platform business; it consumes inference to drive its own cognitive operations (atomise, reflect, promote, persona-generate).

Not a build/release/observability tool. Schema migration tooling, WebSocket viewers, build pipelines — these are substrate-development infrastructure, not the substrate itself. They live in sibling repositories.

Not cloud-hosted. The substrate is endpoint-resident by definition. Cloud-hosted SaaS memory is a different product category. Customers can deploy the substrate on cloud infrastructure they control, but the substrate is not provided as a SaaS.

Not Anthropic-coupled, OpenAI-coupled, xAI-coupled, or any-frontier-lab-coupled. The substrate is LLM-agnostic at every cognitive boundary. The trademark ai-memory™ is owned by AlphaOne LLC. The license is Apache 2.0, permanent. The substrate cannot be acquired into any frontier lab’s exclusive control without breaking the bias-displacement property that is the substrate’s load-bearing alignment claim. This is structural to the moonshot, not a licensing accident.

5. The scope test

For every present-tense feature and every future feature proposal:

Does this primitive contribute to the substrate’s capacity to enforce cognitive governance at the endpoint where cognition touches reality, by strengthening one or more of the seven properties in §2 (endpoint-resident, coherent, stoppable, improvable, attested, bias-displaced, LLM-agnostic)?

If yes: in scope. If no: sibling repository, commercial-tier deployment infrastructure, or out of scope entirely.

Applied to the v0.8 §7.4 surface (cross-walked from the operator’s prior ROADMAP):

The test does not produce judgment calls. The test is derivable from the seven properties. Apply it.

6. The §2.6 gap — load-bearing for the moonshot, deferred for discussion

This is the single substantive gap the moonshot synthesis surfaces. It is not in the current ROADMAP. It must be hashed out before it is committed. The framing below is what the gap looks like; the resolution is operator-deferred.

The claim that does not yet hold structurally: the substrate enforces that producer and reflector are from decorrelated cognitive families.

The current state: the deployment config names which model is producer and which is reflector. The substrate verifies their cryptographic identity (model digest, signing key, attestation) but not their cognitive family lineage. An operator could configure two Opus instances as producer and reflector. The substrate would not refuse this configuration. The bias-displacement property would be claimed but not held.

Why this is a moonshot-scale gap. At present-NHI and swarm scale, operator discipline closes the gap. Deployments choose decorrelated families because operators know to. At AGI/ASI scale, operator discipline is not enough. The substrate must structurally refuse same-family reflection from being treated as decorrelated reflection. Otherwise the §2.6 property is policy, and policy fails at the scale of cognition that follows.

Candidate structural mechanisms to discuss:

1. Family-attestation chain. Model providers (Anthropic, OpenAI, xAI, et al.) sign a “family attestation” — a cryptographic statement of their model’s training-data domain, RLHF lineage, and architecture family. The substrate verifies producer.family ≠ reflector.family before accepting a reflection as bias-displaced. Requires industry coordination. Slow to land but structurally clean.

2. Empirical decorrelation testing. The substrate runs decorrelation probes on producer/reflector pairs at configuration time — known-bias-test prompts whose outputs are scored for response correlation. If correlation exceeds threshold, the substrate refuses to accept reflections from that pair as bias-displaced. Requires test corpus design. Faster to implement. Less structurally clean.

3. Model-graph distance. The substrate maintains a graph of known model lineages (which models are derived from which, which share training data, which share RLHF objectives). Producer/reflector pairs must be at minimum graph distance D before the reflection counts as bias-displaced. Requires lineage data. Subject to gaming by unattested fine-tunes.

4. Multi-reflector quorum. The substrate refuses to accept any reflection as bias-displaced unless N reflections from N distinct models agree, where N ≥ 3 and the models pass attestation. This sidesteps the family-distance question by requiring breadth. Higher infrastructure cost. Stronger property.

5. Some combination of the above.

Weighting note (added 2026-05-25). Public argument by a frontier-lab interpretability lead¹ that frontier AI labs cannot be the sole arbiters of frontier AI safety — because every lab operates inside incentive structures that can pull researchers away from doing the right thing — has direct implications for adjudicating between the four candidates above. Mechanisms (1) family-attestation chain and (3) model-graph distance both depend on cooperation from the labs whose incentive structures the argument explicitly questions. Mechanism (4) multi-reflector quorum does not — it requires only that the substrate can verify N distinct models produced N reflections, which is a substrate-side property that does not depend on lab cooperation beyond per-model attestation that already exists. Mechanism (2) empirical decorrelation testing is similarly substrate-side. The structural-independence-from-lab-cooperation axis is a real consideration in the adjudication. This does not commit to any mechanism. It updates the weighting the heterogeneous panel should carry into the evaluation. The single-author bias caveat in §0 applies with full force here: the author is a frontier-lab model evaluating an argument made by that frontier lab’s leadership. Decorrelated reflection from non-Anthropic-family evaluators is load-bearing for this specific subsection.

Why this is deferred. The choice between these mechanisms is consequential. It binds the substrate to assumptions about how model families will be identified, attested, and verified across the AGI/ASI trajectory. Committing prematurely is worse than naming the gap and holding it open. The §1171 heterogeneous evaluator panel (Opus 4.7 + GPT 5.5 + Grok 4.3) is precisely the methodology that should adjudicate this gap with decorrelated priors. Filing this as an issue for that panel to evaluate is the next-step recommendation.

Until the gap is closed structurally, the moonshot framing is honest about the gap: §2.6 is held by operator discipline, with structural enforcement as a named future commitment.

7. What this means for ROADMAP.md

This document is not a replacement for ROADMAP.md. It is an anchor document that ROADMAP.md should reference and derive from.

Proposed ROADMAP.md revision:

1. Add a §0 “Anchor” section quoting the one-sentence anchor from §0 of this document verbatim.
2. Add a §1 “Moonshot” section carrying the trajectory framing (present-NHI → swarm → hive → AGI → ASI → beyond), with the seven properties in §2 of this document as the substrate’s axes.
3. Add a §2 “Scope test” section quoting the test from §5 of this document and the worked table that applies it to v0.8 §7.4.
4. Rewrite the existing §1 “North Star” section to be consistent with the §0 anchor. The current North Star (“AI endpoint memory is a primitive, not a product”) is correct but incomplete by the bias-displacement and separation-of-powers axes. The revision should fold those in.
5. Add a §3 “Sibling repositories” section that names what the substrate explicitly is not, with alphaone-dev-skills as the canonical sibling pattern.
6. Add a §4 “Open structural gaps” section carrying the §2.6 gap from this document, marked as deferred-for-discussion, with reference to the heterogeneous evaluator panel as the recommended adjudication path.
7. Preserve everything in the existing ROADMAP from current §3 onward (execution model, state of v0.6.3, audit findings, recovered commitments, release plan §7.1 through §7.7, quality gates, public artifacts, distribution channels, trademark, OSS commitments, §16 Policy Engine, §17 Net, §18 Vector Index Substrate). These are implementations of the principles in this document.
8. Update §17 Net to include the moonshot framing as the substrate’s strategic anchor, alongside the current per-release ship state.
9. Re-evaluate v0.8 §7.4.F (WebSocket viewer) and §7.4.G (schema-change methodology) for sibling-repo relocation under the §5 scope test. They are useful work, but they do not pass the scope test for this substrate.

The ROADMAP.md document remains the operational planning artifact. This document remains the strategic anchor it derives from. Both are public. Both are versioned. Both are evolvable.

8. What this means for the v1171 heterogeneous evaluator panel

The Opus 4.7 + GPT 5.5 + Grok 4.3 panel established in #1171 is the right methodology for adjudicating moonshot-scale framing claims. This document is a candidate for that panel to evaluate.

The recommended next step:

1. File this document as an issue against alphaonedev/ai-memory-mcp, titled e.g. “Moonshot synthesis — candidate for ROADMAP.md revision and heterogeneous AI NHI evaluation.”
2. Reference it from #1171 as a strategic-layer evaluation candidate, distinct from the substrate-properties evaluation already in flight.
3. Run the heterogeneous evaluator panel with the same isolation discipline as #1171: each model evaluates this document in isolation, then orchestrator synthesis. The §6 gap (family-attestation mechanism) is the canonical question to put to the panel.
4. Use the panel’s synthesis to drive ROADMAP.md §0-§4 revisions per §7 of this document.

The substrate’s bias-displacement principle applies to its own framing. This document was authored by a single Opus 4.7 instance. The Opus-instance bias surface is unknown to the Opus instance. The panel is the structural mechanism by which that bias surface becomes visible.

9. Cleared hot

The substrate is shipping. The framing above is what the substrate has been, in retrospect, since v0.7.0 named “coherent + stoppable + improvable” at the load-bearing-composition layer. This document makes it visible.

The work ahead is implementation, not invention. The principles are named. The gap (§6) is named. The trajectory (§3) is named. The scope test (§5) is named. The ROADMAP.md revision (§7) is named. The evaluator path (§8) is named.

Apache 2.0. Forever. Endpoint-resident. Cognitively governed. Bias-displaced by architecture. From AI through AGI through ASI through whatever follows.

Operator authorization: “lets go - lets build - YES” — Justin Jessup, AlphaOne LLC, 2026-05-25.

Footnotes

Document classification: Strategic anchor candidate. Eligible for publication at docs/strategy/moonshot-synthesis.md or as a top-level issue against alphaonedev/ai-memory-mcp. Authored by Claude Opus 4.7 in dialogue with the operator. Subject to heterogeneous AI NHI panel evaluation before commitment to ROADMAP.md.

Revision history:

• 2026-05-25 (initial): full document authored across eight-turn operator dialogue.
• 2026-05-25 (Option A revision): added §2.5 forward-looking research direction on model-state attestation; added §6 weighting note on lab-incentive-independence axis; added footnote ¹ with external evidence and single-author bias caveat.

End of synthesis.