ACE 0.8.3 is a release candidate over the published 0.8.2 developer preview. This page distinguishes the public contract from implemented surfaces that remain experimental.
The supported self-hosted path is:
install ace-core → configure SurrealDB and an optional provider → start ACE → authenticate
→ connect → map → watch → receive a cited brief → activate → inspect resources → stop cleanly
The 0.8.3 candidate identities are:
- Python distribution:
ace-core - Python import:
ace - CLI command:
ace - thin MCP command:
ace-mcp-client - version:
0.8.3
The thin MCP surface contains exactly eleven tools:
| Tool | Purpose |
|---|---|
ace_start |
Establish product and session context |
ace_load |
Load relevant accumulated intelligence |
ace_capture |
Persist an observation or correction |
ace_task |
Submit complex orchestration with a durable receipt |
ace_status |
Retrieve task or system status |
ace_capture_idea |
Preserve an emerging idea |
ace_search |
Search accumulated intelligence |
ace_briefing |
Retrieve a return briefing |
ace_impact |
Inspect likely code impact |
ace_history |
Inspect file or symbol history |
ace_related |
Find related code and knowledge |
The CLI, thin MCP adapter, public Core/Intelligence/Application packages, authorized Intelligence resource plane, inert Domain Pack boundary, persistence migrations, reference extension mechanism, and documented provider routes are the compatibility focus for the current developer preview. Atrium is a supported optional repository-delivered preview over the same resource plane. Changes to these surfaces receive migration notes when needed.
The supported CLI also includes ace landscape, a versioned, authenticated, strictly read-only
Living Product Graph snapshot. It exposes stable object identity, canonical and non-operational
assertion states, evidence, provenance, uncertainty, history, decisions, corrections, and outcomes
without adding an MCP tool or any write, execution, extension, or model-inference authority. Its
read contract freezes ordering, bounds, absence, failure, redaction, and
0.3.x compatibility behavior.
Roadmap outcome E1 is passed for the exact ace-core 0.3.0 release. The E1 release evidence binds the implementation to the published package matrix, deployment inventory, independent Claude Fable 5 AI security review, and release-owner countersignature. The security record is intentionally described as an independent AI review, not a human penetration test, professional audit, or certification.
ACE 0.3.0 supports one canonical governed-cognition contract for reusable recipes, instruments,
frameworks, tools, perspectives, and procedural knowledge. Stable cognition identity, immutable
revision identity, scoped active heads, proposals, human review receipts, bounded selection/use
receipts, and effectiveness observations remain distinct. The supported lifecycle is
teach → propose → inspect → approve → use → measure → revise or retire; proposal or use alone
cannot approve, activate, or promote cognition.
Core recipes and current trusted in-process extensions enter the same typed catalog. Current Core adapts the v0.2.0 reference registration contract, while v0.2.0 Core refuses the current reference extension before any partial registration. Unknown, future, malformed, conflicting, unapproved, expired, superseded, unavailable, or scope-incompatible cognition fails closed. The naked kernel, zero-extension configuration, mixed wheel/source-distribution packages, and independently packaged consumer remain supported conformance cases.
The authenticated cognition HTTP boundary owns proposal, semantic diff, review, activation, rollback, expiry, disablement, retirement, discovery, and use inspection. Deprecated skill, framework, and self-optimizer facades project into this boundary without manufacturing approval provenance. The existing thin MCP surface remains exactly eleven tools and gains only additive cognition receipt projections. E1 adds no writable model authority and supports no untrusted in-process extension execution, distributed approval, or exactly-once external side-effect claim.
See the governed-cognition operations runbook, threat model, and architecture/migration packet.
The product-scoped State Engine v1 contracts are supported for the measured single-node topology: one ACE API/worker deployment, one SurrealDB/SurrealKV database, and bounded synchronous adapter clients. The supported Core boundary includes material-derived evidence identities; item/batch and operational receipts; deterministic candidate/evidence packs; reviewed as-of belief projections; reviewed transition revisions; labeled action/no-action simulations and reconciliation; I3 reasoning-use receipts; and human-authority promotion, correction, and supersession lineage.
The supported extension-first journey is executable from a source checkout as documented in the product-builder guide and README. It performs, in order:
- build, clean-install, and discover the independent product extension;
- schema-zero and supported-predecessor upgrade verification;
- bounded adapter ingestion, exact replay, counts, lineage, and scope checks;
- grounded evidence query and frozen as-of projection with five explicit epistemic meanings;
- transition hypothesis challenge/review with mechanism, uncertainty, evidence, and causal limits;
- action/no-action and named-alternative/no-action consequence comparisons;
- structured decision, task-time I3 use, and explicit eligible promotion receipts;
- incomplete and matched later-outcome reconciliation without rollout mutation;
- real database/API/worker restart and fresh-client material use;
- correction plus append-only supersession and post-correction restart; and
- interruption recovery, degraded/failure cases, and exact eleven-tool revalidation.
The adapter supplies bounded, untrusted proposals and content digests. Authenticated Core context owns product scope, stable identity, validation, transactions, replay, and receipts. Source or model content cannot select product, task, tool, review, mutation, causal, or promotion authority. Source claims stay in the grounded evidence plane unless an independently eligible conclusion completes the explicit promotion lifecycle; simulations remain separate from observations and beliefs.
The reference evidence-query and promotion-review task actions are real production-router
integration but remain on the explicitly experimental extension-invocation HTTP surface. They are
not a new stable CLI/MCP contract. The public thin MCP adapter remains exactly eleven tools, and the
legacy in-process engine MCP surface is not the supported public MCP boundary.
The frozen reference envelope is 200 records per item, 200 items per manifest, 200 candidate
records, 50 returned candidates, 20 evidence-pack records, 8 runtime evidence records, 8 rollout
branches/steps/transitions, and 20 promotion retrieval results. The measured initial corpus was
200,000 claims and 236,000 semantic records under a 2 GiB store budget on an 11-core, 18 GB M3 Pro;
the retained post-sustained store held 220,000 claims and 256,000 semantic records. K1-K3 passed
means only the published bounded single-node contracts, scale packet, and the frozen extension-first
journey recorded in the
K1-K3 product evidence.
Schema zero and the public 0.1.4 predecessor upgrade to v168 are supported. Current-head partial application resumes through the ordinary installer; arbitrary interruption inside historical pre-v142 migrations requires restoring a pre-migration backup. The operations runbook covers health, backup/restore, interrupted replay, archival/reactivation, and stop conditions. Distributed ordering, multi-writer consistency, multi-region recovery, real-world causal accuracy or calibration, autonomous learning, general world-model intelligence, and general real-world L1 beneficial impact are not supported claims.
ACE 0.3.1 adds a builder-facing ace state workflow and authenticated Product State HTTP boundary.
The supported claim is bounded to the contract, topology, and failure behavior below.
The release supports:
GET /product-state/capabilitiesfor deterministic adapter/action/version discovery;POST /product-state/ingestionsfor extension mapping under authenticated Core product scope;ace state capabilities,ingest,invoke,correct, andinspectas one outcome-led journey;- additive, allowlisted Living Product Graph metadata for ingestion, belief, transition, task-evidence, material-use, rollout, reconciliation, promotion, decision, deliberation, and correction receipts; and
- schema-zero v171 and v168→v171 compatibility in the frozen extension-first acceptance.
The ingestion and inspection contracts are supported surfaces. The underlying grounded- state adapter and extension task-action registration remain experimental extension hooks: installed extensions are explicitly trusted in-process, action authority stays bounded by their manifest, and failures remain visible. The CLI adds no MCP tool and does not make capture equivalent to promotion.
The Productized State guide and release evidence record the exact journey, failure behavior, receipt identities, and artifact boundary.
ACE 0.5.0 supports one ACE host, one durable store, and explicitly trusted in-process Action adapters. Approved reasoning proceeds through an exact Decision, human review of the effect-free plan, Core authorization and admission, a terminal result, separate post-effect verification, and explicit promotion. Cancellation, attempt identity, replay, limits, restart recovery, and resource outcomes remain durable and inspectable. Neither a successful effect nor verification promotes itself.
The supported claim is not distributed execution, untrusted adapter isolation, globally exactly- once effects, compensation, or remote orchestration. The separate World Intelligence adapter proves public-contract portability for this bounded topology. The exact tag, packages, restart journey, hashes, and non-claims are in the 0.5.0 release evidence.
ACE 0.6.0 supports bounded exact measured-impact evaluation. It links an intelligence artifact or immutable cognition revision to product-owned criteria, matched conditions, material-use attribution, Decision, reviewed Action, observed Outcome, uncertainty, and a useful, harmful, or unproven receipt. Its append-only proposal may name promotion, rejection, rollback, or retirement but is non-effective, non-selectable, and requires separate human/Core authority. Exact replay survives a fresh real-store process without reclassification.
The public World Intelligence 0.10.0 journey composes those contracts over recorded official data.
It reproduces seven World-owned matched evaluations as useful, preserves every proposal as
non-effective and non-selectable, and records a separately authorized reject / no_action
Decision without changing governed state. Core's deterministic coverage also proves harmful and
unproven classifications plus fail-closed attribution, condition, availability, leakage,
duplicate, replay, interruption, and authority controls. The exact tag, public packages, hashes,
checks, and limitations are in the
0.6.0 release closeout.
The kickoff, disposition, observed-result, convergence, audit, #49 hardening, and merged-source
packets remain historical construction records for this public closeout. Issue #49 marks F1 and F5
complete against merged Core and re-dates unresolved, unwaived F3 to 2026-11-05 under documented
containment.
The supported claim does not include proposal application, self-certification, autonomous optimization, causal effect, population performance, general model quality, human benefit, live network freshness, or general SI4 completion. Product measures, thresholds, source policy, and domain nouns remain outside Core and Intelligence. Market's public reproduction is the next independent falsifier, not a condition retroactively added to the released World-based boundary.
The broader HTTP and engine MCP APIs, Atrium, worker automation, MAKE/SHIP execution arms, foresight, calibration, proactive intelligence, continuous-learning paths, and advanced extension hooks are implemented parts of ACE. They are not stable 0.3.x contracts: their APIs, supported end-to-end journeys, and compatibility guarantees can change. This maturity label limits the public promise; it does not reduce those systems to roadmap concepts or peripheral demos.
The authenticated extension-invocation-v1 HTTP envelope and
extension-invocation-receipt-v1 projection are experimental. They add an extension-owned
preparation/outcome bridge over Core's durable task lifecycle, including linked attempt-level
resume after restart. This is real HTTP execution authority, not a supported CLI/MCP surface, and
does not broaden the supported governed-cognition boundary. See the
experimental extension-invocation contract.
Long-running public tasks use persisted receipts and expose pending, running, completed, failed, cancelled, and degraded outcomes. In the supported 0.5.0 single-host topology, durable attempt identity, cancellation, exact replay, limits, and restart recovery are passed. This does not claim distributed task claiming, multi-host coordination, or general exactly-once external effects. The experimental extension-invocation surface separately supports linked retries and negotiated process-local cancellation; neither is a distributed guarantee.
The supported I1 nested decision/correction receipt contract, explicit incomplete provenance, privacy boundary, lifecycle history, and replay evidence are documented in Decision and correction receipts. I1 is passed without widening the eleven-tool surface or adding execution authority.
The supported I2 deliberation-receipt-v1 projection is exposed through existing task/status,
opt-in CLI, thin-client, and read-only Living Product Graph paths. It records bounded observable
shape selection, execution-identity contributor artifacts, artifact-grounded conflicts, synthesis
dispositions, and honest partial/degraded coverage without exposing hidden reasoning or adding
execution authority. I2 is passed, but attribution does not establish correctness, causality, or
benefit. See I2 closeout evidence.
The supported I3 intelligence-use-receipt-v1 projection is exposed through the existing
ace_status task result and read-only Living Product Graph. It distinguishes retrieved, injected,
reflected, and decision-material evidence; limits comparison to the six structured I1 fields; and
degrades missing, mismatched, failed, or partial controls without reconstruction. I3 is passed, but
material influence is not beneficial impact and does not imply L1 success. See
I3 closeout evidence.
The L1 ace.foresight.impact-evaluation/v1 evaluator computes bounded,
cluster-aware later-outcome comparisons without accepting caller-supplied quality labels. Its first
checksum-frozen public-data probe did not establish benefit. A tamper-evident prospective protocol
was frozen, but a pre-outcome collection-start audit invalidated v1 because it cannot be executed
without adding unfrozen cohort, route, outcome, estimator, attribution, and leakage choices.
The fully frozen agent-only v5 successor later completed 144 decisions over 36 eligible clusters
but failed naïve/base-rate and matched model-only. V6 preserved a favorable statistical result but
failed closed on one I3 field-level lineage record. Its independently seeded v7 correction
replicate froze that already-required validation before collection, completed 192 decisions over
48 eligible clusters, and passed persistence, naïve/base-rate, and matched model-only under the
all-controls interval rule. L1 is therefore passed for the frozen executable-workload claim only;
human, customer, external-product, provider, and general real-world benefit remain unsupported. See the
L1 evidence gate.
Atrium is a supported repository-delivered developer-preview workspace. It consumes the same authorized Intelligence resource plane as machine clients and is covered by repository tests and production-build acceptance. It remains optional, is not included in the Python wheel or sdist, and is never an authoritative state, policy, or effect path.
A capability moves into the preview contract only after it has a documented user journey, failure behavior, compatibility boundary, and reproducible tests. Product ideas and planned work belong in the public roadmap, not in this support inventory.