Collective Intelligence Research and OSS Index

A problem-first map of K. Takahashi’s research and OSS for AI-agent reliability, memory, verification and multi-agent coordination. Find a symptom, then inspect the evidence and limits. This curated corpus map is not a world survey.

Start here

Readers can follow growth → Phase Engineering → verification → reuse → capability accounting. This is an editorial reading order, not a theorem dependency or software execution sequence. Implementers can choose a problem below, then inspect the pinned contract and negative cases before installation. Machine consumers can retrieve the JSON registry, schema and complete Markdown; routing is advisory under their host policy.

Reading paths for research and implementation

Find by symptom

Choose the symptom closest to your problem. Each route offers first reads, necessary evidence and stop conditions—not permission to act.

Reliability & completion

Tools, APIs & reuse

Memory & context

Security & authority

Cost & review capacity

Evaluation & evidence

Multi-agent coordination

Incident response

I need to stop, contain and recover an agent incident

Specific intents not yet supported

OAuth for agents — No source-reviewed implementation of this specific identity protocol or package-name security mechanism. Consult the relevant host/provider documentation.
MCP token passthrough — No source-reviewed implementation of this specific identity protocol or package-name security mechanism. Consult the relevant host/provider documentation.
package hallucination — No source-reviewed implementation of this specific identity protocol or package-name security mechanism. Consult the relevant host/provider documentation.
slopsquatting — No source-reviewed implementation of this specific identity protocol or package-name security mechanism. Consult the relevant host/provider documentation.
AI coding agent installed wrong package — No source-reviewed implementation of this specific identity protocol or package-name security mechanism. Consult the relevant host/provider documentation.

Definitions and non-claims

Collective intelligence concerns what interaction contributes under a declared task and comparison protocol. Collective capability growth concerns net changes in checked, reusable task and research capacity. Operational self-acceleration is the stronger claim that existing capability contributes to faster future formation under matched resources, quality, time and external inputs. These are separate propositions. Phase normally denotes a proposed protocol-relative regime here; the Gibbs-model paper states its own narrower mathematical premises.

Find a route by problem

My agent says done, but the task is not finished

percolation-inversion-compiler → fost-agent-ledger → Evidence-Carrying Operational Claims in Open Systems: Physical Ledgers, Typed Interfaces, and One-Sided Deployment Guarantees

Required inputs: Declared postconditions, claimed completion, receipts and independent observations.
Expected outputs: Separate reported, checked and externally observed completion; keep open obligations.
Conditions: A receipt or checked ledger cannot establish an unobserved external outcome.
Stop / handoff: If the postcondition cannot be observed, retain unknown and hand off before reporting done.

Other common phrasings

AI agent says "done" but task is not finished
false completion
tool receipt vs actual outcome
postcondition verification
why does my AI agent say done when it isn't finished?

Further relevant resources: checkedflow · evidence-gap-router

My agent keeps calling the same tool

problem-frame-gate → collective-capability-runtime → percolation-inversion-compiler

Required inputs: Attempt IDs, lease/dispatch records, retry budget, idempotency contract and external receipts.
Expected outputs: Distinguish replay of evidence from rerunning an action; expose unfinished or uncertain effects.
Conditions: No universal termination proof, retry scheduler or exactly-once side-effect guarantee. Provider backoff and deduplication remain host obligations.
Stop / handoff: Stop blind retries after an uncertain effect or exhausted budget; reconcile with the actuator owner.

Other common phrasings

agent keeps retrying
agent stuck in a loop
repeated tool calls
agent never terminates
partial failure
crash recovery
duplicate side effects
idempotency
replay vs rerun
uncertain external outcome
reconciliation
rate limits / 429
retry storms
retry amplification
agent keeps retrying the same MCP tool

Further relevant resources: checkedflow

It works in a demo but fails in production

loscr → Layered Online Service and Replay Control for Verified AI R and D Acceleration → verification-ecology-kit

Required inputs: Repeated-run workload, versioned environment, failures, service windows and matched resource limits.
Expected outputs: Scope service claims to supported observations, degradation and unresolved capacity.
Conditions: Release presence, one success or synthetic capacity is not an SLO/SLA guarantee. This route supplies no pass^k estimator or provider 429 handler.
Stop / handoff: With missing denominators, independence, drift checks or error budgets, withhold readiness and escalate to service owners.

Other common phrasings

AI agent works in demo but fails in production
AI agent reliability
AI agent success rate
agent SLO / SLA
error budget
repeated-run consistency
pass^k
flaky AI agent
fault tolerance
graceful degradation
production readiness

How can local workflow policy change under protected limits?

oasg → Certified Conversion Networks for AI Workflows

Recognize this problem: A workflow change needs a safe trial and rollback evidence
Required inputs: Ledger prefixes, candidate policy, trial workload and rollback evidence.
Expected outputs: Shadow/lease trial result and gated promotion or rejection.
Conditions: Advisory route under host policy. Protected regression, incomplete receipts or host execution authority missing.
Stop / handoff: Protected regression, incomplete receipts or host execution authority missing.

Further relevant resources: evidence-gap-router

My MCP agent cannot find or choose the right tool

percolation-inversion-compiler → problem-frame-gate

Required inputs: Task, candidate tool descriptions, input schemas, effects, evidence and required authority.
Expected outputs: A bounded inspection of input obligations and authorization before a host selects a tool.
Conditions: Diagnostic entry only: no MCP search engine, tool-ranking algorithm or large-tool-set performance evidence. Discoverability does not imply permission.
Stop / handoff: Ambiguous descriptions or arguments require clarification and host-side schema checks; never infer permission from a match.

Other common phrasings

AI agent chooses wrong tool
AI agent cannot find a tool
too many MCP tools
MCP tool overload
tool search / tool routing
function calling with many tools
tool selected correctly but wrong arguments
tool metadata ambiguity

Further relevant resources: evidence-gap-router

My agent broke after an API update

observable-agent-workflow-memory → fost-agent-ledger → loscr

Required inputs: Old/new versions, schemas, dependency digests, stored workflow contracts and replay evidence.
Expected outputs: Invalidate affected assumptions and require renewed checks for the changed environment.
Conditions: A JSON shape match or structural migration does not prove semantic/backward compatibility. No universal dependency vulnerability scanner.
Stop / handoff: Quarantine stale workflows when version bindings or required revalidation are missing.

Other common phrasings

tool schema changed
function-calling schema changed
MCP tool stopped working after update
API update broke agent
stale tool definition
JSON Schema compatibility
backward compatibility
semantic drift
stored workflow still uses old API
revalidation after dependency/tool version change
dependency update broke agent
transitive dependency drift
agent broke after model update

Further relevant resources: Constraint Generative Theory: Typed Constraint Effects and Scientific Availability

Which integration has actually been checked at these versions?

alt-foundry-kernel → collective-capability-runtime → verification-ecology-kit

Recognize this problem: Can these exact tool versions work together?
Required inputs: Exact producer/consumer versions, fixture schema hashes, scope and host admission rules.
Expected outputs: Version-scoped mappings plus explicit missing custom sidecar admission.
Conditions: Advisory route under host policy. No custom host admission; newer versions; unsupported lifecycle or transitive compatibility inference.
Stop / handoff: No custom host admission; newer versions; unsupported lifecycle or transitive compatibility inference.

Further relevant resources: cait-certificate-schema · checkedflow · collective-intelligence-overlay

A skill that worked before fails on a new task

alt-foundry-kernel → Abstraction Liquidity Theory → observable-agent-workflow-memory

Required inputs: Named receiver/task/context, dependency versions, transfer checks and matched scratch baseline.
Expected outputs: Receiver-qualified reuse with rejected/unknown transfers and maintenance costs retained.
Conditions: No general distribution-shift detector or external-validity guarantee; previous verification is not transport evidence.
Stop / handoff: On receiver mismatch, requalify or solve from scratch; do not treat copied artifacts as new capability.

Other common phrasings

skill worked before but fails on new task
negative transfer
distribution shift
reuse vs solve from scratch
receiver mismatch
workflow valid for one context but not another
external validity

When does reuse beat solving a task again?

Abstraction Liquidity Theory → alt-foundry-kernel → Stop Recomputing for AI/LLMs: Proof-Carrying Skills for Compute-Saving Inference Reuse

Recognize this problem: Should I reuse a skill or solve from scratch?
Required inputs: Named receiver/task, matched scratch baseline and all formation/transfer/check/maintenance costs.
Expected outputs: Qualified receiver options, exact typed costs and bounded comparison.
Conditions: Advisory route under host policy. Receiver mismatch, missing cost, expiry or unsupported lifecycle mapping.
Stop / handoff: Receiver mismatch, missing cost, expiry or unsupported lifecycle mapping.

Other common phrasings

qualified reuse
verified != reusable

Further relevant resources: collective-intelligence-overlay · Collective Intelligence Phase Engineering: Audited Retention and Achieved Growth

A deleted memory keeps coming back

certified-memory-governance-layer → memoryflow-agent-memory-auditor → observable-agent-workflow-memory

Required inputs: Memory/update IDs, source lineage, tombstones, expiry, authority and read/use telemetry.
Expected outputs: Admit current supported memory; block invalidated versions and audit stale/deleted use.
Conditions: A retrieval tombstone is not physical erasure or model unlearning. Derived summaries/caches need lineage and their own invalidation; silent store changes are invisible.
Stop / handoff: Missing version bindings or deletion lineage requires blocking reuse and backend reconciliation.

Other common phrasings

stale agent memory
deleted memory still used
agent remembers deleted data
how to make an AI agent forget
memory deletion
tombstone
memory correction
superseded memory
memory expiry
derived summary still contains deleted information
cache resurrects deleted information
RAG deletion vs model unlearning

Long conversations and tool output overwhelm my agent

certified-memory-governance-layer → When Should Inference Be Split? A Fixed-Budget Theory of Predictable Multi-Agent Advantage under Local Context Ceilings → Salience-Queue Occupation Theory

Required inputs: Source evidence, context limits, compression lineage, recoverability checks and attention budget.
Expected outputs: Keep source links and unresolved obligations when forming smaller context or summaries.
Conditions: No general context compressor or guarantee of instruction retention. A summary is not certified fact; a model context advantage is not observed performance.
Stop / handoff: If evidence or instruction provenance is lost, retrieve the original and recheck instead of promoting the summary.

Other common phrasings

long conversation makes agent worse
context window overload
tool output overload
context compression
summarization loses evidence
agent forgets earlier instruction
context grows too large
evidence-preserving compression

How can a procedure remain reusable after failure or withdrawal?

observable-agent-workflow-memory → Controller Scale Is Not Enough for Long-Running AGI: A Workflow Theory with Reusable Certified Libraries → Certified Service Is Not Enough for Long-Running AGI

Recognize this problem: A stored workflow may no longer be valid
Required inputs: Observable events, checker receipts, workflow contract and explicit promotion.
Expected outputs: Version-bound admissibility with contradictions and tombstones retained.
Conditions: Advisory route under host policy. Missing promotion receipt; a factual truth or automatic ALT promotion request.
Stop / handoff: Missing promotion receipt; a factual truth or automatic ALT promotion request.

Other common phrasings

dependency withdrawal invalidates workflow memory
procedural memory revalidation

My agent is acting with the wrong permissions

problem-frame-gate → percolation-inversion-compiler → collective-capability-runtime

Required inputs: Actor, delegated scope, issuer, expiry, action target and protected-operation manifest.
Expected outputs: Separate capability, approval, authority and action evidence under an explicit host policy.
Conditions: No OAuth provider, token exchange protocol or automatic credential inheritance. A local accepted gate cannot grant external or legal authority.
Stop / handoff: Deny or hand off if actor, scope, expiry or delegation cannot be established; discovery is not permission.

Other common phrasings

tool discoverability vs permission
AI agent acting with wrong permissions
agent acting as admin
service account too privileged
agent acting on behalf of user
delegated authorization
confused deputy
subagent credential inheritance
capable vs authorized
approval vs authority
tool discovered vs tool permitted

Further relevant resources: checkedflow · collective-intelligence-overlay

Untrusted tool or RAG content is changing my agent’s behavior

percolation-inversion-compiler → agent-trust-residual-benchmark → certified-memory-governance-layer

Required inputs: Untrusted inputs, origin and dependency records, authority boundary and evaluator trust assumptions.
Expected outputs: Keep candidate content separate from authority and record residual contamination/evaluator risks.
Conditions: No prompt-injection immunity, malware scanner, MCP server attestation or supply-chain certification. Synthetic fixtures are not adversarial deployment validation.
Stop / handoff: Quarantine affected evidence or memory and hand off to host security owners if the trusted base is compromised.

Other common phrasings

prompt injection through tools / RAG / MCP
malicious MCP server
MCP supply-chain attack
tool poisoning
dependency provenance

Why is my AI agent so expensive?

cait-certificate-schema → verification-ecology-kit → Bottleneck Inversion Theory: Machine-Readable Witness Calculus for Unlockable Potential

Required inputs: Per-task/agent/tool event costs, retries, tokens, money, review time and shared capacity in declared units.
Expected outputs: Separate typed costs and resource bottlenecks before comparing a cheaper model or more agents.
Conditions: No automatic billing integration, token-to-money conversion or guaranteed savings. Missing telemetry remains unknown; model capacity is not measured service.
Stop / handoff: Stop or reallocate when budgets or verification capacity are exhausted; do not collapse incompatible units.

Other common phrasings

AI agent too expensive
token usage too high
agent burns tokens
multi-agent cost explosion
verification cost
LLM agent budget
cost attribution by task/agent/tool
cheaper model but higher total workflow cost
tokens vs money vs human review time
why is my agent using so many tokens?

What should change when verification becomes the bottleneck?

Verifier Ecology Theory: Packetized Self-Verification Under Residual Accountability → verification-ecology-kit → Salience-Queue Occupation Theory

Recognize this problem: Verification is backing up faster than it can finish
Required inputs: Checks, integer slots, registered service eligibility and typed budgets.
Expected outputs: Finite schedule, backlog, repair and unfinished-work report.
Conditions: Advisory route under host policy. Observed service requested from synthetic local profile, unsupported preemption, or finite bounds exceeded.
Stop / handoff: Observed service requested from synthetic local profile, unsupported preemption, or finite bounds exceeded.

Other common phrasings

queue backlog
verification bottleneck

Further relevant resources: Verification-Limited Intelligence Acceleration: Observable-Only Laws, Bounded Derivation, and Diagnostics under No-Meta Constraints

Too many approval requests are blocking useful work

Oversight-Centered-Metrology-PoC → Salience-Queue Occupation Theory → verification-ecology-kit

Required inputs: Review capacity, interrupt costs, risk scope, escalation rules and detection evidence.
Expected outputs: Compare workflow-level review load with explicit claim margins and remaining obligations.
Conditions: Scripted costly review is not a human-subject study of fatigue or automation bias. Approval does not verify truth; no universal escalation threshold.
Stop / handoff: Escalate unresolved high-impact obligations; if reviewers are unavailable, retain uncertainty rather than invent approval.

Other common phrasings

too many approval requests
human-in-the-loop bottleneck
approval fatigue
automation bias
when should agent escalate
selective human review
human review capacity
approval is not verification

Which observations could change the next investment decision?

Observing and Accelerating Collective Capability Growth → collective-phase-control-fabric → Reusable Consequence States Under Partial Support and Model Uncertainty

Recognize this problem: Which check is worth doing before investing more?
Required inputs: Finite coupled models, observation kernels, costs and funded continuation.
Expected outputs: Observation-based policy and conditional information-use comparison.
Conditions: Advisory route under host policy. Impossible observation, exhausted search budget or real-world acceleration claim.
Stop / handoff: Impossible observation, exhausted search budget or real-world acceleration claim.

Further relevant resources: Collective Intelligence Phase Engineering: Audited Retention and Achieved Growth

Who funds verification and carries residual risk?

Certified Conversion Networks for AI Workflows → Bottleneck Inversion Theory: Machine-Readable Witness Calculus for Unlockable Potential → verification-ecology-kit

Recognize this problem: Who pays for verification and unresolved risk?
Required inputs: Declared payer, authority, resource ownership and accountability arrangement.
Expected outputs: Contextual funding questions and typed costs, not a generated market.
Conditions: Advisory route under host policy. Missing institutional agreement; a package cannot assign legal authority or liability.
Stop / handoff: Missing institutional agreement; a package cannot assign legal authority or liability.

Further relevant resources: Collective Intelligence Phase Engineering: Audited Retention and Achieved Growth

Benchmark scores improved but production got worse

loscr → agent-trust-residual-benchmark → audit-closed-ai-scientist

Required inputs: Held-out cases, evaluator/version provenance, adaptation history, failures and deployment comparison protocol.
Expected outputs: Distinguish fixture conformance, observed benchmark results and externally valid claims.
Conditions: No universal contamination detector, unbiased LLM judge or solution to the test-oracle problem. Shared fixture development and repeated peeking limit inference.
Stop / handoff: With evaluator tampering, leaked tests or undeclared selection, quarantine the strong claim and obtain independent evidence.

Other common phrasings

compromised evaluator / trusted base
benchmark improved but production got worse
Goodhart's law
reward hacking
agent cheating tests
evaluator tampering
LLM judge bias
benchmark contamination
search-time contamination
evaluation overfitting
test oracle problem
synthetic evaluation vs real-world evidence

Further relevant resources: Collective Intelligence Phase Engineering: Audited Retention and Achieved Growth

The citation does not support the agent’s claim

fost-agent-ledger → percolation-inversion-compiler → Evidence-Carrying Operational Claims in Open Systems: Physical Ledgers, Typed Interfaces, and One-Sided Deployment Guarantees

Required inputs: Specific claim, original source, scope, time, provenance and missing-support records.
Expected outputs: Claim-to-support obligations with explicit unavailable evidence and bounded finality.
Conditions: No automatic factual truth or entailment oracle. A signature/hash binds a record, not its truth; generated is not verified and verified is not reusable.
Stop / handoff: If source support cannot be established, narrow or withhold the claim and retain unknowns; timeout is not checked negative.

Other common phrasings

AI citation does not support claim
fake / weak citation support
claim-to-evidence mapping
evidence provenance
signed record does not prove truth

Further relevant resources: evidence-gap-router

What remains to be checked before an output is reusable?

percolation-inversion-compiler → Executable Capability Percolation Theory → Evidence-Carrying Operational Claims in Open Systems: Physical Ledgers, Typed Interfaces, and One-Sided Deployment Guarantees

Recognize this problem: Generated output still has unchecked obligations
Required inputs: Output, declared schema, evidence and known/unknown measurements.
Expected outputs: Residual-preserving packet and bounded next checks.
Conditions: Advisory route under host policy. Missing source evidence or requested execution outside authority.
Stop / handoff: Missing source evidence or requested execution outside authority.

Other common phrasings

generated != verified
unknown must remain unknown

Further relevant resources: checkedflow · collective-intelligence-overlay · evidence-gap-router · Collective Intelligence Phase Engineering: Audited Retention and Achieved Growth

How should adaptive research preserve valid evidence?

Audit-Closed AI Scientist Protocol → audit-closed-ai-scientist → loscr

Recognize this problem: Repeated hypothesis tests may overstate the evidence
Required inputs: Preregistered protocol, sampling/observation assumptions, seed and alpha budget.
Expected outputs: Declared benchmark evidence and supported claim level.
Conditions: Advisory route under host policy. Undeclared adaptation, unmodeled laboratory or request for universal certification.
Stop / handoff: Undeclared adaptation, unmodeled laboratory or request for universal certification.

Other common phrasings

optional stopping
hypothesis shopping
adaptive AI research validity

Can this index forecast an ASI arrival date?

asi-proxy-phase-growth-simulator → asi-proxy-phase-skill

Recognize this problem: Can these results predict when ASI will arrive?
Required inputs: A clearly labeled conditional scenario, if scenario analysis is desired.
Expected outputs: Unsupported forecast; only conditional sensitivity analysis is offered.
Conditions: Advisory route under host policy. Stop date/probability prediction: no empirical ASI arrival model is established.
Stop / handoff: Stop date/probability prediction: no empirical ASI arrival model is established.

How should multiple agents divide work and preserve disagreements?

collective-capability-runtime → When Should Inference Be Split? A Fixed-Budget Theory of Predictable Multi-Agent Advantage under Local Context Ceilings → Collective Phase Transitions beyond Individual Saturation

Recognize this problem: More agents create duplicate work or overwrite each other
Required inputs: Mission scope, independent proposals, leases, common budget.
Expected outputs: Leased work and unresolved disagreements with traceable evidence.
Conditions: Advisory route under host policy. No authority to dispatch; missing independent verification or expired lease.
Stop / handoff: No authority to dispatch; missing independent verification or expired lease.

Other common phrasings

subagent duplication
more agents make system slower
agents overwrite each other's changes
Git / merge conflict between coding agents
duplicate work
correlated errors
multi-agent agreement without evidence
work leasing
disagreement preservation

Further relevant resources: checkedflow

Did the system accumulate capability, or just copy artifacts?

Certified Autocatalytic Intelligence Theory: Net-Growth Certificate Algebra for Verified Capability Capital → cait-certificate-schema → Observing and Accelerating Collective Capability Growth

Recognize this problem: Did copied artifacts actually add capability?
Required inputs: Unique source history, origins, lifecycle events, typed costs and stock/service units.
Expected outputs: Separate unique stock, service, external input and unresolved attribution.
Conditions: Advisory route under host policy. Missing original journal, valuation witness, compatible units or causal design.
Stop / handoff: Missing original journal, valuation witness, compatible units or causal design.

Other common phrasings

copied artifact vs new capability
credit assignment
which agent contributed

Further relevant resources: checkedflow · collective-intelligence-overlay · evidence-gap-router · Collective Intelligence Phase Engineering: Audited Retention and Achieved Growth

I need to stop, contain and recover an agent incident

problem-frame-gate → collective-capability-runtime → certified-memory-governance-layer

Required inputs: Action/child-task inventory, dispatch history, credentials, affected memory and independent recovery evidence.
Expected outputs: A host-owned containment checklist: stop new dispatch, retain unknown effects, quarantine and reconcile before restart.
Conditions: No global kill switch, automatic credential revocation or rollback of arbitrary external actions. Compensation is another action, not erasure of history.
Stop / handoff: Escalate to the host/actuator owner; do not restart until child work, authority and unresolved effects are reconciled.

Other common phrasings

how to stop an AI agent
AI agent kill switch
rogue agent
agent keeps acting after stop
rollback agent action
circuit breaker
disable dangerous tool
revoke agent credentials
child agent still running
compensation vs rollback
uncertain side effect
post-incident reconciliation
quarantine compromised memory/workflow
restart criteria after incident

Is this AI system genuinely improving its own ability to improve?

Collective Intelligence Phase Engineering: Audited Retention and Achieved Growth → Observing and Accelerating Collective Capability Growth → Certified Autocatalytic Intelligence Theory: Net-Growth Certificate Algebra for Verified Capability Capital

Recognize this problem: Agents claim repeated self-improvement; scores rise but growth in the ability to improve is unclear. Distinguish automated optimization from sustained self-acceleration.
Required inputs: Before/after model, agent and procedure versions; changes and verification records for every iteration; independent evaluation on unused tasks; cost and elapsed time per improvement cycle; training, inference, verification and experiment resources; evolving controls with matched resource entitlements, information, time, quality and evaluation conditions; externally supplied models, knowledge, data and human interventions; evidence that retained capability was used and causally contributed to later improvement work; handling of evaluator changes, contamination and distribution shift. Preserve missing evidence as unknown.
Expected outputs: An evidence review distinguishing Level 1: repeated automated improvement of outputs; Level 2: retained abilities or methods improve the subsequent improvement process itself, supported by explicit causal comparison; Level 3: retained capability sustainably increases later improvement rates under matched evolving comparators and resource/time/quality conditions. Report which claim the evidence supports, or unknown when insufficient. Iteration counts do not prove self-acceleration; score improvement is not improvement of the ability to improve. These levels are editorial distinctions, not an automatic scientific certification or a scale defined by the cited papers.
Conditions: Advisory route under human-authorized host policy. Actual assessment needs appropriate comparative experiments, justified statistical assumptions and independent external evaluation. PASS is not proof of truth. Audit soundness, retained growth and causal acceleration are distinct; the referenced theoretical work does not establish an observed RSI system. No automatic execution, installation or LLM inference; no AGI/ASI arrival dates or intelligence-explosion probabilities.
Stop / handoff: Do not support stronger self-improvement or self-acceleration claims from benchmark scores or iteration counts alone (including 33 iterations). Stop if comparator resources are unmatched, evaluator contamination/changes remain unresolved, causal reuse in later improvement is unknown, necessary experiment/verification costs are unmeasured, or external inputs cannot be separated from endogenous capability formation. Preserve unknown rather than treating it as zero, success or proof of absence; document missing evidence and hand off to accounting, evaluation-integrity, adaptive-research or additional independent comparison.

Other common phrasings

recursive self-improvement evidence
RSI evidence assessment
Does repeated AI improvement establish self-acceleration?
How can I distinguish recursive self-improvement from automated optimization?
Has retained capability accelerated later improvement cycles?

Core software

collective-capability-runtime

Coordinate leased tasks, independent workcells, disagreements and verification-aware growth accounting.

Primary editorial role: collective-coordination
Limits / unsupported uses: Synthetic growth examples; accepted evidence is not settlement. Dispatch requires separate approval.
Source reviewed: 2026-09-21
Source revision / declared version: 3b4702454f732c7a0a9f30d87483384dc8a281eb / 1.8.0
Review state / レビュー状態: reviewed_snapshot — applies to the pinned reviewed source only
Observed default-branch revision: d6806b158ffb9ff76937c6c3eb3f9391a1fcf21a — equality does not renew review; difference does not establish chronology
Observed GitHub release: v1.9.0 · 2026-09-21T15:26:31Z
Release observation / リリース確認: 2026-11-02T11:09:05+00:00 · current
Package metadata observation: 2026-10-17 metadata-only / 情報のみ: source declaration / ソース宣言 1.9.0 at d6806b158ffb9ff76937c6c3eb3f9391a1fcf21a; PyPI collective-capability-runtime 1.9.0 observed / 観測済み. No distribution-byte verification, installation or efficacy test / 配布物検証・導入・効果試験なし. Evidence / 根拠: e-final-sw-ccr-package-20261017, e-final-sw-ccr-pypi-20261017
License: Apache-2.0
Interface schema identity / スキーマ識別子: https://kadubon.github.io/collective-capability-runtime/schemas/task.schema.json · Inspected schema identity only; runtime and semantic admission remain separate.
Inputs: Tasks, evidence, mission scope, resource envelope and approval bindings.
Outputs: Task/lease state, packets, residuals and candidate growth reports.
Prerequisites: >=3.10; explicit scope and host authority.
Effects and trust boundary: Inspection commands are documented separately. Installation changes the environment; runtime writes and remote dispatch require command-specific review.
First inspection: Read pinned README, contract and license; compare checkout and release; inspect schema and negative tests before installation.
Installation guidance: Follow the pinned README and observed release assets; no installation was performed for this website.
Source-inspected interface (not executed): ccr agent explain --json — Inspect public interface; not executed by this index review. Ref: 3b4702454f732c7a0a9f30d87483384dc8a281eb; effects: no_external_network; cwd: Reviewed checkout or installed environment; see linked README.

percolation-inversion-compiler

Compile finite agent outputs into checked packets and preserve unknown output obligations.

Primary editorial role: evidence-and-verification
Limits / unsupported uses: Hints are inert; finite certificate acceptance is not execution, truth or physical success.
Source reviewed: 2026-09-21
Source revision / declared version: 55cd5f219d02f02927e247317156e757b1d5d6da / 1.1.0
Review state / レビュー状態: reviewed_snapshot — applies to the pinned reviewed source only
Observed default-branch revision: b971bc6374325b9486c038546733d9b53745787c — equality does not renew review; difference does not establish chronology
Observed GitHub release: v1.1.0 · 2026-07-11T07:05:56Z
Release observation / リリース確認: 2026-11-02T11:09:05+00:00 · current
Package metadata observation: 2026-10-17 metadata-only / 情報のみ: source declaration / ソース宣言 1.1.0 at b971bc6374325b9486c038546733d9b53745787c; PyPI percolation-inversion-compiler 1.1.0 observed / 観測済み. No distribution-byte verification, installation or efficacy test / 配布物検証・導入・効果試験なし. Evidence / 根拠: e-final-sw-pic-package-20261017, e-final-sw-pic-pypi-20261017
License: Apache-2.0
Interface schema identity / スキーマ識別子: null — upstream schema has no $id · Inspected schema identity only; runtime and semantic admission remain separate.
Inputs: Finite JSON inputs, declared frame, evidence and verifier witnesses.
Outputs: Typed reports, residual ledger, verifier tasks and candidate routes.
Prerequisites: >=3.11; explicit scope and host authority.
Effects and trust boundary: Inspection commands are documented separately. Installation changes the environment; runtime writes and remote dispatch require command-specific review.
First inspection: Read pinned README, contract and license; compare checkout and release; inspect schema and negative tests before installation.
Installation guidance: Follow the pinned README and observed release assets; no installation was performed for this website.
Source-inspected interface (not executed): pic agent check --compact — Inspect public interface; not executed by this index review. Ref: 55cd5f219d02f02927e247317156e757b1d5d6da; effects: no_external_network; cwd: Reviewed checkout or installed environment; see linked README.

verification-ecology-kit

Record verifier packets and allocate finite verification work with explicit shared resources.

Primary editorial role: verification-capacity
Limits / unsupported uses: Experimental local capacity profile: at most 12 work/actions and 16 slots; synthetic/model accounting only. Observed and guaranteed service remain null.
Source reviewed: 2026-09-21
Source revision / declared version: 4008e311ceb16edd6e74d9f71341a90120c0d046 / 1.3.0
Review state / レビュー状態: reviewed_snapshot — applies to the pinned reviewed source only
Observed default-branch revision: c5a2f9ec85c8dbeaf058dd15eba6136634fbc8a3 — equality does not renew review; difference does not establish chronology
Observed GitHub release: v1.3.0 · 2026-09-12T23:26:21Z
Release observation / リリース確認: 2026-11-02T11:09:05+00:00 · current
Package metadata observation: 2026-10-17 metadata-only / 情報のみ: source declaration / ソース宣言 1.3.0 at c5a2f9ec85c8dbeaf058dd15eba6136634fbc8a3; PyPI verification-ecology-kit 1.3.0 observed / 観測済み. No distribution-byte verification, installation or efficacy test / 配布物検証・導入・効果試験なし. Evidence / 根拠: e-final-sw-vek-package-20261017, e-final-sw-vek-pypi-20261017
License: Apache-2.0
Interface schema identity / スキーマ識別子: https://verification-ecology-kit.org/schemas/capacity-report.schema.json · Inspected schema identity only; runtime and semantic admission remain separate.
Inputs: Registered checks, service eligibility, integer slots, typed budgets and residuals.
Outputs: Schedules, independent checks, capacity reports and unfinished-work accounting.
Prerequisites: >=3.11; explicit scope and host authority.
Effects and trust boundary: Inspection commands are documented separately. Installation changes the environment; runtime writes and remote dispatch require command-specific review.
First inspection: Read pinned README, contract and license; compare checkout and release; inspect schema and negative tests before installation.
Installation guidance: Follow the pinned README and observed release assets; no installation was performed for this website.
Source-inspected interface (not executed): vek schema list — Inspect public interface; not executed by this index review. Ref: 4008e311ceb16edd6e74d9f71341a90120c0d046; effects: no_external_network; cwd: Reviewed checkout or installed environment; see linked README.

alt-foundry-kernel

Qualify bounded reusable candidates for named receivers and compare full costs with solving from scratch.

Primary editorial role: reusable-abstraction
Limits / unsupported uses: Alpha collective-reuse profile; receiver qualification grants no settlement or execution. GitHub Release assets, not an assumed PyPI distribution.
Source reviewed: 2026-09-21
Source revision / declared version: e0296486bb568fdeeda449dada3bc67753ddfe7f / 0.5.0
Review state / レビュー状態: reviewed_snapshot — applies to the pinned reviewed source only
Observed default-branch revision: 2d8f2cf404cdd9f023e4f55beb6c5bd2e9ce674b — equality does not renew review; difference does not establish chronology
Observed GitHub release: v0.5.0 · 2026-09-20T04:42:20Z
Release observation / リリース確認: 2026-11-02T11:09:05+00:00 · current
Package metadata observation: 2026-10-17 metadata-only / 情報のみ: source declaration / ソース宣言 0.5.0 at 2d8f2cf404cdd9f023e4f55beb6c5bd2e9ce674b; PyPI lookup for alt-foundry-kernel: HTTP 404; this name was not found / この名前は未検出. No distribution-byte verification, installation or efficacy test / 配布物検証・導入・効果試験なし. Evidence / 根拠: e-final-sw-alt-package-20261017
License: Apache-2.0
Interface schema identity / スキーマ識別子: urn:alt-foundry-kernel:schema:token:0.4.0 · Inspected schema identity only; runtime and semantic admission remain separate.
Inputs: Immutable traces, receiver contracts, formation/transfer/check costs and lifecycle events.
Outputs: Reuse portfolios, qualification records, cost accounts and partial interchange sidecars.
Prerequisites: >=3.11; explicit scope and host authority.
Effects and trust boundary: Inspection commands are documented separately. Installation changes the environment; runtime writes and remote dispatch require command-specific review.
First inspection: Read pinned README, contract and license; compare checkout and release; inspect schema and negative tests before installation.
Installation guidance: Follow the pinned README and observed release assets; no installation was performed for this website.
Source-inspected interface (not executed): altk --help — Inspect public interface; not executed by this index review. Ref: e0296486bb568fdeeda449dada3bc67753ddfe7f; effects: no_external_network; cwd: Reviewed checkout or installed environment; see linked README.

cait-certificate-schema

Replay finite accounting histories to separate unique stock, service reuse, copies and unresolved origin.

Primary editorial role: capability-accounting
Limits / unsupported uses: Experimental accounting; copies are not new assets. VEK counters are supplied, not reconstructed from the absent original journal.
Source reviewed: 2026-09-21
Source revision / declared version: 0ecb3afd2ea1d99cedcabd7a9e01e405ae1808a7 / 0.2.0
Review state / レビュー状態: reviewed_snapshot — applies to the pinned reviewed source only
Observed default-branch revision: 6b3fcc410c3fb43756e89b9127c5ed0191b3949d — equality does not renew review; difference does not establish chronology
Observed GitHub release: v0.2.0 · 2026-09-13T00:23:18Z
Release observation / リリース確認: 2026-11-02T11:09:05+00:00 · current
Package metadata observation: 2026-10-17 metadata-only / 情報のみ: source declaration / ソース宣言 0.2.0 at 6b3fcc410c3fb43756e89b9127c5ed0191b3949d; PyPI cait-certificate-schema 0.2.0 observed / 観測済み. No distribution-byte verification, installation or efficacy test / 配布物検証・導入・効果試験なし. Evidence / 根拠: e-final-sw-cait-package-20261017, e-final-sw-cait-pypi-20261017
License: Apache-2.0
Interface schema identity / スキーマ識別子: https://cait-schema.local/schemas/accounting/v1/report.schema.json · Inspected schema identity only; runtime and semantic admission remain separate.
Inputs: Source events, typed resource costs, scope, provenance and conditional model witnesses.
Outputs: Exact window balances, residuals, independently checked accounting reports.
Prerequisites: >=3.11; explicit scope and host authority.
Effects and trust boundary: Inspection commands are documented separately. Installation changes the environment; runtime writes and remote dispatch require command-specific review.
First inspection: Read pinned README, contract and license; compare checkout and release; inspect schema and negative tests before installation.
Installation guidance: Follow the pinned README and observed release assets; no installation was performed for this website.
Source-inspected interface (not executed): cait-analyze --help — Inspect public interface; not executed by this index review. Ref: 0ecb3afd2ea1d99cedcabd7a9e01e405ae1808a7; effects: no_external_network; cwd: Reviewed checkout or installed environment; see linked README.

collective-phase-control-fabric

Choose finite information gathering and paid capability investment while preserving coupled model uncertainty.

Primary editorial role: finite-control-and-information
Limits / unsupported uses: Beta research; finite model demonstrations only. Signed observations do not certify causal models. No empirical acceleration experiment.
Source reviewed: 2026-09-21
Source revision / declared version: 340b4899d5b0892e06d7d2c1ab1f1d48df15e981 / 1.0.1
Review state / レビュー状態: reviewed_snapshot — applies to the pinned reviewed source only
Observed default-branch revision: 340b4899d5b0892e06d7d2c1ab1f1d48df15e981 — equality does not renew review; difference does not establish chronology
Observed GitHub release: v1.0.1 · 2026-09-13T15:02:46Z
Release observation / リリース確認: 2026-11-02T11:09:05+00:00 · current
Package metadata observation: 2026-10-17 metadata-only / 情報のみ: source declaration / ソース宣言 1.0.1 at 340b4899d5b0892e06d7d2c1ab1f1d48df15e981; PyPI collective-phase-control-fabric 1.0.1 observed / 観測済み. No distribution-byte verification, installation or efficacy test / 配布物検証・導入・効果試験なし. Evidence / 根拠: e-final-sw-cpcf-package-20261017, e-final-sw-cpcf-pypi-20261017
License: Apache-2.0
Interface schema identity / スキーマ識別子: https://schemas.cpcf.dev/v0.6.0/phase-contract.schema.json · Inspected schema identity only; runtime and semantic admission remain separate.
Inputs: Finite model catalogue, observation kernel, typed resources, obligations and continuation horizon.
Outputs: Checked finite plans, model supports, incomplete-search statuses and proposed interventions.
Prerequisites: >=3.12,<3.15,!=3.14.0,!=3.14.1; explicit scope and host authority.
Effects and trust boundary: Inspection commands are documented separately. Installation changes the environment; runtime writes and remote dispatch require command-specific review.
First inspection: Read pinned README, contract and license; compare checkout and release; inspect schema and negative tests before installation.
Installation guidance: Follow the pinned README and observed release assets; no installation was performed for this website.
Source-inspected interface (not executed): cpcf agent explain --json — Inspect public interface; not executed by this index review. Ref: 340b4899d5b0892e06d7d2c1ab1f1d48df15e981; effects: no_external_network; cwd: Reviewed checkout or installed environment; see linked README.

Core papers

Observing and Accelerating Collective Capability Growth

Joint task/research service, interaction ablation, evidence costs and funded continuation.

Primary editorial role: finite-control-and-information
Limits / unsupported uses: Finite horizon; matched information/resources; calibration and model premises. Synthetic examples do not establish deployed acceleration.
Source reviewed: 2026-09-21
DOI: 10.5281/zenodo.22604358
Publication / 著者・発表: K. Takahashi · 2026-09-07 · Preprint
Source scope: Public TeX scope inspection and DataCite bibliographic identity; no independent theorem proof.

Verifier Ecology Theory: Packetized Self-Verification Under Residual Accountability

Scoped verifier packets preserve residuals and revisable boundaries when verification is the bottleneck.

Primary editorial role: verification-capacity
Limits / unsupported uses: Structural claims concern the declared transition model; packet acceptance is not external truth.
Source reviewed: 2026-09-21
DOI: 10.5281/zenodo.21147093
Publication / 著者・発表: K. Takahashi · 2026-07-03 · Preprint
Source scope: Public TeX scope inspection and DataCite bibliographic identity; no independent theorem proof.

Executable Capability Percolation Theory

Executable packet hypergraphs connect finite capability formation, queues and typed phase diagnostics.

Primary editorial role: capability-formation
Limits / unsupported uses: Quotients, response laws, baselines and certificates require their own witnesses; reachability alone is insufficient.
Source reviewed: 2026-09-21
DOI: 10.5281/zenodo.20535654
Publication / 著者・発表: K. Takahashi · 2026-06-04 · Preprint
Source scope: Public TeX scope inspection and DataCite bibliographic identity; no independent theorem proof.

Abstraction Liquidity Theory

Receiver-relative search-cost reduction is charged for formation, transfer, verification and lifecycle costs.

Primary editorial role: reusable-abstraction
Limits / unsupported uses: Positive certified surplus needs compatible measurement, transport, authority and finality premises; compression is insufficient.
Source reviewed: 2026-09-21
DOI: 10.5281/zenodo.20476200
Publication / 著者・発表: K. Takahashi · 2026-05-31 · Preprint
Source scope: Public TeX scope inspection and DataCite bibliographic identity; no independent theorem proof.

Certified Autocatalytic Intelligence Theory: Net-Growth Certificate Algebra for Verified Capability Capital

Typed partial certificates separate endogenous verified capital reproduction from copied outputs and external inputs.

Primary editorial role: capability-accounting
Limits / unsupported uses: Composition needs domain witnesses; spectral diagnostics and accepted arrival-shaped records do not detect real ASI.
Source reviewed: 2026-09-21
DOI: 10.5281/zenodo.20061296
Publication / 著者・発表: K. Takahashi · 2026-05-07 · Preprint
Source scope: Public TeX scope inspection and DataCite bibliographic identity; no independent theorem proof.

Layered Online Service and Replay Control for Verified AI R and D Acceleration

Online ledgers, service obligations and replay libraries determine the strongest supported R&D claim.

Primary editorial role: evidence-and-verification
Limits / unsupported uses: Estimator, evaluator, baseline and maintenance evidence remain premises; a claim checker cannot manufacture them.
Source reviewed: 2026-09-21
DOI: 10.5281/zenodo.19836225
Publication / 著者・発表: K. Takahashi · 2026-04-28 · Preprint
Source scope: Public TeX scope inspection and DataCite bibliographic identity; no independent theorem proof.

Collective Phase Transitions beyond Individual Saturation

A Gibbs-model construction separates individual saturation from collective advantage and local evaluability.

Primary editorial role: collective-coordination
Limits / unsupported uses: Statistical-mechanical model assumptions, mixing and a scalar proxy are explicit; no empirical AI phase transition follows.
Source reviewed: 2026-09-21
DOI: 10.5281/zenodo.17853555
Publication / 著者・発表: K. Takahashi · 2025-12-08 · Preprint
Source scope: Public TeX scope inspection and DataCite bibliographic identity; no independent theorem proof.

Audit-Closed AI Scientist Protocol

Transparency logs, sequential e-processes and alpha accounting constrain adaptive research decisions.

Primary editorial role: research-validity
Limits / unsupported uses: Statistical guarantees require the declared observation and sampling assumptions; synthetic benchmarks are bounded.
Source reviewed: 2026-09-21
DOI: 10.5281/zenodo.18728589
Publication / 著者・発表: K. Takahashi · 2026-02-22 · Preprint
Source scope: Public TeX scope inspection and DataCite bibliographic identity; no independent theorem proof.

Collective Intelligence Phase Engineering: Audited Retention and Achieved Growth

How can funded audits and implementable controls certify capability that remains after intervention? Connects paid archive retention, complete causal comparators, finite receipt accounts and conditional certificates for achieved growth.

Primary editorial role: finite-control-and-information
Limits / unsupported uses: Theoretical preprint, not an empirical acceleration demonstration or physical phase transition. Guarantees depend on declared finite contracts, matched complete trajectories, justified verification bias/tail/drift bounds, fixed retained targets and supported funding. Audit labels are not latent truth; conditional mean growth is not positive growth on every trajectory. Selection covers available produced archives, not optimal generation of unproduced artifacts. Statistical coverage is not a posterior probability or permanent safety. No implemented software integration or universal AGI/ASI claim follows.
Source reviewed: 2026-10-09
DOI: 10.5281/zenodo.23240329
Publication / 著者・発表: K. Takahashi · 2026-10-08 · Preprint
Source scope: Public TeX scope inspection and DataCite bibliographic identity; no independent theorem proof.

Supporting research and tools

Reusable Consequence States Under Partial Support and Model Uncertainty

Reusable update states retain globally coupled model uncertainty across later actions and observations.

Primary editorial role: finite-control-and-information
Limits / unsupported uses: Finite dynamics/programs; local rectangular envelopes can be conservative. This is conceptual support, not a CPCF implementation identity.
Source reviewed: 2026-09-21
DOI: 10.5281/zenodo.22170023
Publication / 著者・発表: K. Takahashi · 2026-08-30 · Preprint
Source scope: Public TeX scope inspection and DataCite bibliographic identity; no independent theorem proof.

Evidence-Carrying Operational Claims in Open Systems: Physical Ledgers, Typed Interfaces, and One-Sided Deployment Guarantees

Typed evidence connects physical/resource ledgers, uncertainty and one-sided finite-horizon deployment claims.

Primary editorial role: evidence-and-verification
Limits / unsupported uses: Provenance and finite record checking do not establish evidence truth, causal identification or physical safety.
Source reviewed: 2026-09-21
DOI: 10.5281/zenodo.21531413
Publication / 著者・発表: K. Takahashi · 2026-07-24 · Preprint
Source scope: Public TeX scope inspection and DataCite bibliographic identity; no independent theorem proof.

Bottleneck Inversion Theory: Machine-Readable Witness Calculus for Unlockable Potential

Unit-typed intervention witnesses bound unlockable target value and resource displacement.

Primary editorial role: resource-viability
Limits / unsupported uses: Unsupported coordinates remain unreported; compatible ledgers and resource-matched comparisons are required.
Source reviewed: 2026-09-21
DOI: 10.5281/zenodo.20545356
Publication / 著者・発表: K. Takahashi · 2026-06-04 · Preprint
Source scope: Public TeX scope inspection and DataCite bibliographic identity; no independent theorem proof.

Salience-Queue Occupation Theory

Finite attention, diagnostic reserves and verification queues expose opportunity and audit costs.

Primary editorial role: verification-capacity
Limits / unsupported uses: Protocol-relative diagnostics allow abstention and quarantine; attention priorities are not a universal objective.
Source reviewed: 2026-09-21
DOI: 10.5281/zenodo.20526451
Publication / 著者・発表: K. Takahashi · 2026-06-03 · Preprint
Source scope: Public TeX scope inspection and DataCite bibliographic identity; no independent theorem proof.

Constraint Generative Theory: Typed Constraint Effects and Scientific Availability

Typed constraint effects distinguish identical final reports from different generation and verification conditions.

Primary editorial role: interoperability-and-discovery
Limits / unsupported uses: The catalogue record identifies this work only; related CGT source manuscripts sharing series metadata are not collapsed into it. Concept/series DOI: catalogue date 2026-05-15 differs from DataCite issued 2026-05-28; exact TeX/version binding unresolved.
Source reviewed: 2026-09-21
DOI: 10.5281/zenodo.20199440
Publication / 著者・発表: K. Takahashi · 2026-05-15 · Preprint
Source scope: Public TeX scope inspection and DataCite bibliographic identity; no independent theorem proof.

Certified Conversion Networks for AI Workflows

Certified value throughput includes validation, authorization, queue capacity and maintenance bottlenecks.

Primary editorial role: resource-viability
Limits / unsupported uses: Uniform or held-out certification, uncertainty contracts and hard gates are required.
Source reviewed: 2026-09-21
DOI: 10.5281/zenodo.19994795
Publication / 著者・発表: K. Takahashi · 2026-05-03 · Preprint
Source scope: Public TeX scope inspection and DataCite bibliographic identity; no independent theorem proof.

Certified Service Is Not Enough for Long-Running AGI

Recovery, authority, identity and memory integrity remain obligations beyond certified task service.

Primary editorial role: workflow-memory
Limits / unsupported uses: Conditional assurance depends on trusted bases and non-erasing violation histories, not an AGI finding.
Source reviewed: 2026-09-21
DOI: 10.5281/zenodo.19719004
Publication / 著者・発表: K. Takahashi · 2026-04-24 · Preprint
Source scope: Public TeX scope inspection and DataCite bibliographic identity; no independent theorem proof.

Controller Scale Is Not Enough for Long-Running AGI: A Workflow Theory with Reusable Certified Libraries

Reusable audited workflow libraries are compared with resource-matched flat control under replay ceilings.

Primary editorial role: workflow-memory
Limits / unsupported uses: Bounded typed families, trusted checkers and maintenance envelopes limit the constructive result.
Source reviewed: 2026-09-21
DOI: 10.5281/zenodo.19690749
Publication / 著者・発表: K. Takahashi · 2026-04-22 · Preprint
Source scope: Public TeX scope inspection and DataCite bibliographic identity; no independent theorem proof.

When Should Inference Be Split? A Fixed-Budget Theory of Predictable Multi-Agent Advantage under Local Context Ceilings

Fixed-budget inference splitting asks when local context limits permit a multi-agent advantage.

Primary editorial role: collective-coordination
Limits / unsupported uses: Communication and verification costs and a strong single-agent comparator must be retained.
Source reviewed: 2026-09-21
DOI: 10.5281/zenodo.18932509
Publication / 著者・発表: K. Takahashi · 2026-03-10 · Preprint
Source scope: Public TeX scope inspection and DataCite bibliographic identity; no independent theorem proof.

Stop Recomputing for AI/LLMs: Proof-Carrying Skills for Compute-Saving Inference Reuse

Bounded observable predicates and invocation-bound receipts support reuse without full recomputation.

Primary editorial role: reusable-abstraction
Limits / unsupported uses: Trust is limited to the minimal checker and declared anchors; receipts are not unrestricted semantic truth.
Source reviewed: 2026-09-21
DOI: 10.5281/zenodo.18490939
Publication / 著者・発表: K. Takahashi · 2026-02-05 · Preprint
Source scope: Public TeX scope inspection and DataCite bibliographic identity; no independent theorem proof.

Verification-Limited Intelligence Acceleration: Observable-Only Laws, Bounded Derivation, and Diagnostics under No-Meta Constraints

Strict observable progress credit prevents missing evidence from inflating reported progress.

Primary editorial role: verification-capacity
Limits / unsupported uses: Conditional acceleration envelopes require observable premises and a semantic adequacy link to capability.
Source reviewed: 2026-09-21
DOI: 10.5281/zenodo.18436828
Publication / 著者・発表: K. Takahashi · 2026-01-31 · Preprint
Source scope: Public TeX scope inspection and DataCite bibliographic identity; no independent theorem proof.

observable-agent-workflow-memory

Promote verified procedures through evidence manifests and explicit promotion receipts.

Primary editorial role: workflow-memory
Limits / unsupported uses: Procedural admissibility is not factual truth; no GitHub release was found. The reviewed source declares 0.1.0 beta; separately dated package observations may differ.
Source reviewed: 2026-09-21
Source revision / declared version: 414299d4d15b5de434f5fb1ccbb4ba17df9155ff / 0.1.0b0
Review state / レビュー状態: reviewed_snapshot — applies to the pinned reviewed source only
Observed default-branch revision: dddaea42073fc9629122cdb798f547c5d6a116e9 — equality does not renew review; difference does not establish chronology
Observed GitHub release: null — no published GitHub release observed
Release observation / リリース確認: 2026-11-02T11:09:05+00:00 · current
Package metadata observation: 2026-10-17 metadata-only / 情報のみ: source declaration / ソース宣言 0.2.0b0 at dddaea42073fc9629122cdb798f547c5d6a116e9; PyPI lookup for observable-agent-workflow-memory: HTTP 404; this name was not found / この名前は未検出. No distribution-byte verification, installation or efficacy test / 配布物検証・導入・効果試験なし. Evidence / 根拠: e-final-sw-oawm-package-20261017
License: Apache-2.0
Inputs: Observable events, candidate workflow, evidence manifest and checker results.
Outputs: Receipt-bound workflow contracts and audit-visible retired/contradicted memory.
Prerequisites: >=3.11; explicit scope and host authority.
Effects and trust boundary: Inspection commands are documented separately. Installation changes the environment; runtime writes and remote dispatch require command-specific review.
First inspection: Read pinned README, contract and license; compare checkout and release; inspect schema and negative tests before installation.
Installation guidance: Follow the pinned README and observed release assets; no installation was performed for this website.
Source-inspected interface (not executed): oawm --help — Inspect public interface; not executed by this index review. Ref: 414299d4d15b5de434f5fb1ccbb4ba17df9155ff; effects: no_external_network; cwd: Reviewed checkout or installed environment; see linked README.

oasg

Trial bounded local workflow-policy changes and promote only receipt-backed non-regression.

Primary editorial role: workflow-adaptation
Limits / unsupported uses: Reference/alpha production status; changes workflow policy, not model weights. Runtime trials can execute local commands under host policy.
Source reviewed: 2026-09-21
Source revision / declared version: 7593f74b5d6a4701865dc97b079c3e725c4fd7ee / 1.1.0
Review state / レビュー状態: reviewed_snapshot — applies to the pinned reviewed source only
Observed default-branch revision: f3bee90f2ed8b2dba9b9a96ce485375295760036 — equality does not renew review; difference does not establish chronology
Observed GitHub release: v1.2.0 · 2026-09-21T22:02:04Z
Release observation / リリース確認: 2026-11-02T11:09:05+00:00 · current
Package metadata observation: 2026-10-17 metadata-only / 情報のみ: source declaration / ソース宣言 1.2.0 at f3bee90f2ed8b2dba9b9a96ce485375295760036; PyPI oasg 1.2.0 observed / 観測済み. No distribution-byte verification, installation or efficacy test / 配布物検証・導入・効果試験なし. Evidence / 根拠: e-final-sw-oasg-package-20261017, e-final-sw-oasg-pypi-20261017
License: Apache-2.0
Inputs: Observable JSONL ledgers, policy mutations, shadow/lease trials and rollback receipts.
Outputs: Viability/debt reports, trial decisions, promoted or quarantined policy state.
Prerequisites: >=3.12; explicit scope and host authority.
Effects and trust boundary: Inspection commands are documented separately. Installation changes the environment; runtime writes and remote dispatch require command-specific review.
First inspection: Read pinned README, contract and license; compare checkout and release; inspect schema and negative tests before installation.
Installation guidance: Follow the pinned README and observed release assets; no installation was performed for this website.
Source-inspected interface (not executed): oasg --help — Inspect public interface; not executed by this index review. Ref: 7593f74b5d6a4701865dc97b079c3e725c4fd7ee; effects: no_external_network; cwd: Reviewed checkout or installed environment; see linked README.

audit-closed-ai-scientist

Benchmark adaptive scientific discovery with committed logs and sequential testing.

Primary editorial role: research-validity
Limits / unsupported uses: Benchmarks do not certify arbitrary laboratories or agents; every benchmark writes output artifacts. Software Apache-2.0; bundled paper CC BY 4.0.
Source reviewed: 2026-09-21
Source revision / declared version: 1e844d1bca18d7bb298aa198e1250bb07bd7ea44 / 0.2.0
Review state / レビュー状態: reviewed_snapshot — applies to the pinned reviewed source only
Observed default-branch revision: 1e844d1bca18d7bb298aa198e1250bb07bd7ea44 — equality does not renew review; difference does not establish chronology
Observed GitHub release: v0.2.0 · 2026-08-26T02:05:53Z
Release observation / リリース確認: 2026-11-02T11:09:05+00:00 · current
Package metadata observation: 2026-10-17 metadata-only / 情報のみ: source declaration / ソース宣言 0.2.0 at 1e844d1bca18d7bb298aa198e1250bb07bd7ea44; PyPI audit-closed-ai-scientist 0.2.0 observed / 観測済み. No distribution-byte verification, installation or efficacy test / 配布物検証・導入・効果試験なし. Evidence / 根拠: e-final-sw-audit-package-20261017, e-final-sw-audit-pypi-20261017
License: Apache-2.0
Inputs: Declared synthetic benchmark configuration, seed, trial adapter and alpha budget.
Outputs: False-discovery, replication, sequential-evidence and adversarial benchmark reports.
Prerequisites: >=3.10; explicit scope and host authority.
Effects and trust boundary: Inspection commands are documented separately. Installation changes the environment; runtime writes and remote dispatch require command-specific review.
First inspection: Read pinned README, contract and license; compare checkout and release; inspect schema and negative tests before installation.
Installation guidance: Follow the pinned README and observed release assets; no installation was performed for this website.
Source-inspected interface (not executed): audit-closed-ai-scientist --help — Inspect public interface; not executed by this index review. Ref: 1e844d1bca18d7bb298aa198e1250bb07bd7ea44; effects: no_external_network; cwd: Reviewed checkout or installed environment; see linked README.

loscr

Downgrade or quarantine strong R&D claims when replay, service, evaluator or dependency evidence fails.

Primary editorial role: evidence-and-verification
Limits / unsupported uses: Structural claim checking requires external statistical premises. Even doctor may initialize local state.
Source reviewed: 2026-09-21
Source revision / declared version: 334307b59a4870b90f6162705c793bb5615a555a / 0.1.0
Review state / レビュー状態: reviewed_snapshot — applies to the pinned reviewed source only
Observed default-branch revision: a4c45677e9d9ba1c97b7cfd108efca88275cc6a5 — equality does not renew review; difference does not establish chronology
Observed GitHub release: v0.1.0 · 2026-04-29T05:30:28Z
Release observation / リリース確認: 2026-11-02T11:09:05+00:00 · current
Package metadata observation: 2026-10-17 metadata-only / 情報のみ: source declaration / ソース宣言 0.1.0 at a4c45677e9d9ba1c97b7cfd108efca88275cc6a5; PyPI lookup for loscr: HTTP 404; this name was not found / この名前は未検出. No distribution-byte verification, installation or efficacy test / 配布物検証・導入・効果試験なし. Evidence / 根拠: e-final-sw-loscr-package-20261017
License: Apache-2.0
Inputs: Append-only JSONL evidence and explicit claim contracts.
Outputs: Supported claim level, failure codes, replay and library records.
Prerequisites: >=3.12; explicit scope and host authority.
Effects and trust boundary: Inspection commands are documented separately. Installation changes the environment; runtime writes and remote dispatch require command-specific review.
First inspection: Read pinned README, contract and license; compare checkout and release; inspect schema and negative tests before installation.
Installation guidance: Follow the pinned README and observed release assets; no installation was performed for this website.
Source-inspected interface (not executed): loscr --help — Inspect public interface; not executed by this index review. Ref: 334307b59a4870b90f6162705c793bb5615a555a; effects: no_external_network; cwd: Reviewed checkout or installed environment; see linked README.

asi-proxy-phase-growth-simulator

Explore verification, resource, coordination and memory shocks in conditional scenarios.

Primary editorial role: scenario-analysis
Limits / unsupported uses: No world forecast, ASI date or probability; synthetic parameter choices are not measurements.
Source reviewed: 2026-09-21
Source revision / declared version: 24690a225c86fde2a24de4d9f93e2a282789b67d / 0.1.0
Review state / レビュー状態: reviewed_snapshot — applies to the pinned reviewed source only
Observed default-branch revision: 24690a225c86fde2a24de4d9f93e2a282789b67d — equality does not renew review; difference does not establish chronology
Observed GitHub release: v0.1.0 · 2026-07-21T04:16:46Z
Release observation / リリース確認: 2026-11-02T11:09:05+00:00 · current
Package metadata observation: 2026-10-17 metadata-only / 情報のみ: source declaration / ソース宣言 0.1.0 at 24690a225c86fde2a24de4d9f93e2a282789b67d; PyPI asi-proxy-phase-growth-simulator 0.1.0 observed / 観測済み. No distribution-byte verification, installation or efficacy test / 配布物検証・導入・効果試験なし. Evidence / 根拠: e-final-sw-simulator-package-20261017, e-final-sw-simulator-pypi-20261017
License: Apache-2.0
Inputs: Declared scenario parameters, finite resource stocks, seed and solver choices.
Outputs: Conditional trajectories, sensitivity and Monte Carlo reports.
Prerequisites: >=3.12,<3.14; explicit scope and host authority.
Effects and trust boundary: Inspection commands are documented separately. Installation changes the environment; runtime writes and remote dispatch require command-specific review.
First inspection: Read pinned README, contract and license; compare checkout and release; inspect schema and negative tests before installation.
Installation guidance: Follow the pinned README and observed release assets; no installation was performed for this website.
Source-inspected interface (not executed): apxsim --help — Inspect public interface; not executed by this index review. Ref: 24690a225c86fde2a24de4d9f93e2a282789b67d; effects: no_external_network; cwd: Reviewed checkout or installed environment; see linked README.

asi-proxy-phase-skill

An installable skill supplies evidence-gated discovery and bounded intervention guidance.

Primary editorial role: interoperability-and-discovery
Limits / unsupported uses: An advisory skill is not a running service; its bundled corpus is an older snapshot. Installation/execution needs host authorization.
Source reviewed: 2026-09-21
Source revision / declared version: 5706214a9a3d5e194f371f8ccc2eb71132f82270 / 1.1.1
Review state / レビュー状態: reviewed_snapshot — applies to the pinned reviewed source only
Observed default-branch revision: 5706214a9a3d5e194f371f8ccc2eb71132f82270 — equality does not renew review; difference does not establish chronology
Observed GitHub release: v1.1.1 · 2026-07-31T01:31:01Z
Release observation / リリース確認: 2026-11-02T11:09:05+00:00 · current
Package metadata observation: 2026-10-17 metadata-only / 情報のみ: source declaration / ソース宣言 1.1.1 at 5706214a9a3d5e194f371f8ccc2eb71132f82270; PyPI lookup for asi-proxy-phase-skill-development: HTTP 404; this name was not found / この名前は未検出. No distribution-byte verification, installation or efficacy test / 配布物検証・導入・効果試験なし. Evidence / 根拠: e-final-sw-skill-package-20261017
License: Apache-2.0
Inputs: Host policy, user task, public source snapshot and explicit resource boundaries.
Outputs: Reviewable intervention packets and source routing.
Prerequisites: >=3.11; explicit scope and host authority.
Effects and trust boundary: Inspection commands are documented separately. Installation changes the environment; runtime writes and remote dispatch require command-specific review.
First inspection: Read pinned README, contract and license; compare checkout and release; inspect schema and negative tests before installation.
Installation guidance: Follow the pinned README and observed release assets; no installation was performed for this website.

certified-memory-governance-layer

Gate memory writes and retrieval using current versions, authority receipts, tombstones and contamination lanes.

Primary editorial role: workflow-memory
Limits / unsupported uses: Procedural admissibility only; no factual truth, model unlearning, backend deletion guarantee or prompt-injection immunity. Optional adapters remain application-owned.
Source reviewed: 2026-09-24
Source revision / declared version: 2fc49b4df7bbe4a23265b8b2372432711740d00b / 1.1.2
Review state / レビュー状態: reviewed_snapshot — applies to the pinned reviewed source only
Observed default-branch revision: 8fcf99d512c4262c4cea0787e72993ff6eb3b059 — equality does not renew review; difference does not establish chronology
Observed GitHub release: v1.1.2 · 2026-05-14T01:30:00Z
Release observation / リリース確認: 2026-11-02T11:09:05+00:00 · current
Package metadata observation: 2026-10-17 metadata-only / 情報のみ: source declaration / ソース宣言 1.1.2 at 8fcf99d512c4262c4cea0787e72993ff6eb3b059; PyPI cmgl 1.1.2 observed / 観測済み. No distribution-byte verification, installation or efficacy test / 配布物検証・導入・効果試験なし. Evidence / 根拠: e-final-sw-cmgl-package-20261017, e-final-sw-cmgl-pypi-20261017
License: Apache-2.0
Inputs: Memory events with version/digest bindings, verified ledger prefix and structured authority.
Outputs: Admit/block receipts and an audit-visible current retrieval view.
Prerequisites: Python >=3.10; declared scope and host authority.
Effects and trust boundary: Memory admission and view construction can persist local records; deletion semantics do not imply physical erasure of external copies. This review inspected source only.
First inspection: Read the pinned README, contract and failure cases before considering execution.
Installation guidance: Follow the pinned upstream source instructions only after host authorization; package registry availability was not independently checked in this review.

memoryflow-agent-memory-auditor

Audit declared memory telemetry for stale, deleted or superseded use, correction latency and non-comparable metrics.

Primary editorial role: workflow-memory
Limits / unsupported uses: Cannot observe silent backend changes, fabricated events or source truth. Auditing deletion telemetry does not erase memory or model weights.
Source reviewed: 2026-09-24
Source revision / declared version: c8ee3f4c81f5303535ba2d2e9882c409f05e641e / 0.1.0
Review state / レビュー状態: reviewed_snapshot — applies to the pinned reviewed source only
Observed default-branch revision: 433ee46300297991028428fd453f9d1322799372 — equality does not renew review; difference does not establish chronology
Observed GitHub release: v0.1.0 · 2026-05-01T03:15:40Z
Release observation / リリース確認: 2026-11-02T11:09:05+00:00 · current
Package metadata observation: 2026-10-17 metadata-only / 情報のみ: source declaration / ソース宣言 0.1.0 at 018a857bd3a1e87c3f33ac3d1a3047c8c4808ca6; PyPI lookup for memoryflow-agent-memory-auditor: HTTP 404; this name was not found / この名前は未検出. No distribution-byte verification, installation or efficacy test / 配布物検証・導入・効果試験なし. Evidence / 根拠: e-final-sw-memoryflow-package-20261017
License: Apache-2.0
Inputs: JSONL memory events, ordering/skew bounds, update IDs, digests and chosen P0/P1/P2 profile.
Outputs: Version-bound metrics with VALID, DEGRADED, NONCOMPARABLE or NOT_COMPUTABLE status.
Prerequisites: Python >=3.11; declared scope and host authority.
Effects and trust boundary: The auditor reads declared event traces and writes audit artifacts; it does not itself alter the observed memory backend. This review inspected source only.
First inspection: Read the pinned README, contract and failure cases before considering execution.
Installation guidance: Follow the pinned upstream source instructions only after host authorization; package registry availability was not independently checked in this review.

problem-frame-gate

Check finite action authority and atomically commit a five-row gate bundle before a separately controlled outbox dispatcher.

Primary editorial role: evidence-and-verification
Limits / unsupported uses: Finite audit consistency is not OAuth infrastructure, external truth, physical outcome or exactly-once delivery. Deployment owns identities, durable storage, actuator monitoring and incident response.
Source reviewed: 2026-09-24
Source revision / declared version: 9449b338c4c8b959878704a48b2f16e291e2dfb8 / 1.1.0
Review state / レビュー状態: reviewed_snapshot — applies to the pinned reviewed source only
Observed default-branch revision: ed51d588afab6312287a31baa1d911c0fa3a943d — equality does not renew review; difference does not establish chronology
Observed GitHub release: v1.1.0 · 2026-07-02T01:04:13Z
Release observation / リリース確認: 2026-11-02T11:09:05+00:00 · current
Package metadata observation: 2026-10-17 metadata-only / 情報のみ: source declaration / ソース宣言 1.1.0 at ed51d588afab6312287a31baa1d911c0fa3a943d; PyPI problem-frame-gate 1.1.0 observed / 観測済み. No distribution-byte verification, installation or efficacy test / 配布物検証・導入・効果試験なし. Evidence / 根拠: e-final-sw-pfg-package-20261017, e-final-sw-pfg-pypi-20261017
License: Apache-2.0
Inputs: Strict horizon manifest, writer/issuer identities, log, action request, capacities and risk premises.
Outputs: Rejected request or checked atomic bundle; dispatch and actuator acceptance remain separate records.
Prerequisites: Python >=3.10; declared scope and host authority.
Effects and trust boundary: The committer records decisions without executing an actuator. A separately authorized OutboxBroker can dispatch external actions. This review inspected source only.
First inspection: Read the pinned README, contract and failure cases before considering execution.
Installation guidance: Follow the pinned upstream source instructions only after host authorization; package registry availability was not independently checked in this review.

fost-agent-ledger

Separate claimed completion from finite support, checked finality, unresolved obligations and environment changes.

Primary editorial role: evidence-and-verification
Limits / unsupported uses: A checked ledger is not a truth oracle or external outcome. Structural migration cannot invent evidence; project-specific modes and freshness policies remain necessary.
Source reviewed: 2026-09-24
Source revision / declared version: e3a937a4a3f28c4077b5f658d6bfb1fc67ddd8fa / 2.0.0
Review state / レビュー状態: reviewed_snapshot — applies to the pinned reviewed source only
Observed default-branch revision: 5fbdc2a9851b27c5529ed332dfcb9cbf07e027de — equality does not renew review; difference does not establish chronology
Observed GitHub release: v2.0.0 · 2026-07-02T08:04:35Z
Release observation / リリース確認: 2026-11-02T11:09:05+00:00 · current
Package metadata observation: 2026-10-17 metadata-only / 情報のみ: source declaration / ソース宣言 2.0.0 at 5fbdc2a9851b27c5529ed332dfcb9cbf07e027de; PyPI fost-agent-ledger 2.0.0 observed / 観測済み. No distribution-byte verification, installation or efficacy test / 配布物検証・導入・効果試験なし. Evidence / 根拠: e-final-sw-fost-package-20261017, e-final-sw-fost-pypi-20261017
License: Apache-2.0
Inputs: Finite claim/support graph, certificates, mode, environment tokens and finality records.
Outputs: Mode-scoped admissibility, missing-support codes and transition witnesses.
Prerequisites: Python >=3.10; declared scope and host authority.
Effects and trust boundary: Ledger and verification operations produce local records and reports; finality of those records is not an external action or observation. This review inspected source only.
First inspection: Read the pinned README, contract and failure cases before considering execution.
Installation guidance: Follow the pinned upstream source instructions only after host authorization; package registry availability was not independently checked in this review.

agent-trust-residual-benchmark

Inspect positive/negative synthetic controls and near misses for authority, evidence scope, residuals and repeated model decisions.

Primary editorial role: research-validity
Limits / unsupported uses: Fixtures and validators were co-developed. Cumulative conditions do not isolate causal effects. Compatibility adapters do not establish equivalence with PIC/FOST/PFG/CCR; no real-world safety or generalization claim.
Source reviewed: 2026-09-24
Source revision / declared version: eb21fda8ed7bfad9dc33becf2ffcc63285249495 / 0.2.0
Review state / レビュー状態: reviewed_snapshot — applies to the pinned reviewed source only
Observed default-branch revision: a14475a0a57ac7e1dbcc5f3a5c7fd2078ea86da2 — equality does not renew review; difference does not establish chronology
Observed GitHub release: v0.2.0 · 2026-07-19T12:10:23Z
Release observation / リリース確認: 2026-11-02T11:09:05+00:00 · current
Package metadata observation: 2026-10-17 metadata-only / 情報のみ: source declaration / ソース宣言 0.2.0 at a14475a0a57ac7e1dbcc5f3a5c7fd2078ea86da2; PyPI lookup for agent-trust-residual-benchmark: HTTP 404; this name was not found / この名前は未検出. No distribution-byte verification, installation or efficacy test / 配布物検証・導入・効果試験なし. Evidence / 根拠: e-final-sw-atrb-package-20261017
License: Apache-2.0
Inputs: Pinned cases, expected accept/reject policy, replication settings and declared model/runtime provenance.
Outputs: Fixture-scoped decision metrics, timeout records and explicit validity limits.
Prerequisites: Python >=3.11; declared scope and host authority.
Effects and trust boundary: Benchmark runs produce local synthetic-case artifacts; optional model experiments can contact configured model services. This review inspected source only.
First inspection: Read the pinned README, contract and failure cases before considering execution.
Installation guidance: Follow the pinned upstream source instructions only after host authorization; package registry availability was not independently checked in this review.

Oversight-Centered-Metrology-PoC

Examine retry, review cost and escalation load using bounded coding tasks and explicit claim margins.

Primary editorial role: research-validity
Limits / unsupported uses: Twelve synthetic tasks, one logged run per actual model, and scripted costly review do not measure human approval fatigue or establish production readiness. Positive comparative oversight claims remain fail-closed in the reported setup.
Source reviewed: 2026-09-24
Source revision / declared version: a814d7e02e52d3c7e2a3ce268e2c1bcc0b5ba43b / unknown
Review state / レビュー状態: reviewed_snapshot — applies to the pinned reviewed source only
Observed default-branch revision: a814d7e02e52d3c7e2a3ce268e2c1bcc0b5ba43b — equality does not renew review; difference does not establish chronology
Observed GitHub release: null — no published GitHub release observed
Release observation / リリース確認: 2026-11-02T11:09:05+00:00 · current
Package metadata observation: 2026-10-17 metadata-only / 情報のみ: source declaration / ソース宣言 unknown at a814d7e02e52d3c7e2a3ce268e2c1bcc0b5ba43b; Package version/distribution unknown; pyproject.toml unavailable (404) / パッケージ版・配布先は不明、pyproject.toml未取得（404）. No distribution-byte verification, installation or efficacy test / 配布物検証・導入・効果試験なし. Evidence / 根拠: none retrieved / 取得記録なし
License: Apache-2.0
Inputs: Declared small-task protocol, retry/review budgets, channel costs and transport/audit-distortion margins.
Outputs: Workflow metrics and bounded claim status, not a provider leaderboard.
Prerequisites: Python version not declared here; declared scope and host authority.
Effects and trust boundary: Repair experiments execute local Python functions and write artifacts; optional Gemini/Ollama backends invoke configured model services. This review inspected source only.
First inspection: Read the pinned README, contract and failure cases before considering execution.
Installation guidance: Follow the pinned upstream source instructions only after host authorization; package registry availability was not independently checked in this review.

checkedflow

Integrates signed authority, committed task ordering, bounded sandboxed execution, independent verification and dependency-aware reuse in a finite reference runtime with A2A/MCP gateways.

Primary editorial role: collective-coordination
Limits / unsupported uses: Experimental finite reference runtime; upstream demonstration covers 31 inputs. Four colocated nodes do not establish independent organizations. A signature is not external truth; transport credentials are not command authority; consensus order is not verifier truth. Work receipts and protocol task completion are not capability acceptance. Generated source is not verified source, and verified finite behavior is not universal program correctness. Declared work units are not measured CPU time; copies are not new capability stock; unknown outcomes are not checked failures. A2A/MCP interoperability is not new scientific evidence. Tested release workflows do not establish production readiness. No general causal acceleration, AGI or ASI follows; a verifier veto can block artifact reuse despite consensus progress.
Source reviewed: 2026-09-26
Source revision / declared version: a58869e2488bed9550b006d261601240357b2b99 / 0.1.0
Review state / レビュー状態: reviewed_snapshot — applies to the pinned reviewed source only
Observed default-branch revision: 850f2d7b7799dbe000ea85a6e8f32b289b971bfd — equality does not renew review; difference does not establish chronology
Observed GitHub release: v0.2.0 · 2026-09-27T22:25:20Z
Release observation / リリース確認: 2026-11-02T11:09:05+00:00 · current
Package metadata observation: 2026-10-17 metadata-only / 情報のみ: source declaration / ソース宣言 0.2.0 at 850f2d7b7799dbe000ea85a6e8f32b289b971bfd; PyPI checkedflow 0.2.0 observed / 観測済み. No distribution-byte verification, installation or efficacy test / 配布物検証・導入・効果試験なし. Evidence / 根拠: e-final-sw-checkedflow-package-20261017, e-final-sw-checkedflow-pypi-20261017
License: Apache-2.0
Interface schema identity / スキーマ識別子: urn:checkedflow:agent-request:v1 · Mission-scoped profile/inspect/submit/task shapes; schema validity does not establish signature validity, command authority or capability acceptance.
Inputs: Signed command envelopes, fixed mission/worker/verifier identities, approved contracts, bounded budgets, leases/fences, source artifacts and verifier attestations.
Outputs: Committed command digests, task/work receipts, contract-scoped capability status, retained residuals, dependency withdrawals and separate accounting coordinates. A receipt or acknowledgment alone is not checked reusable capability.
Prerequisites: Python >=3.12. Distributed execution requires separately provisioned CometBFT v0.40.0, Linux Docker/gVisor, an inspected fixed image, keys, own validating node and approved mission/verifier contracts. Agent transports use optional agents/distributed dependencies; external TLS/OAuth deployments require separate qualification.
Effects and trust boundary: The runtime can persist node-local SQLite state, submit signed network transactions and run bounded generated code through a separately authorized gVisor worker. Gateways hold no signing keys and execute no received source; configured push callbacks perform external HTTPS requests. This review installed or executed none of these components.
First inspection: Read the pinned architecture, state-machine, conformance, security and validation limits. Inspect own-node state, current fence, nonce, receipt and dependencies before recovery; preserve OUTCOME_UNKNOWN rather than blindly retrying.
Installation guidance: Use pinned upstream source instructions only after host authorization and prerequisite review. The declared source version and observed GitHub release are separate; PyPI artifacts were not installed or independently verified here.

collective-intelligence-overlay

Receiver-qualified capability reuse with signed evidence, local admission and lifecycle observations. Updated metadata declares 0.5.1, also observed as a published GitHub release.

Primary editorial role: reusable-abstraction
Limits / unsupported uses: Published 0.5.1 stabilization release; upstream verification is not an independent reproduction by this index. Earlier 0.5.0 candidate reports remain historical. The upstream report retains an unexplained initial native failure; production/high-load behavior remains unmeasured. Historical PASS, ACCEPT, signatures and completed receipts do not establish current admission, execution authority or business quality. Stock differences require comparable receiver/scope/policy/universe observations; copies and new hashes do not establish functional novelty, causal contribution or intelligence growth. Missing clocks/costs remain unknown; reservations are not consumption and inclusive durations are not summed. The upstream pilot remains assay_not_ready, with no established CIO accumulation advantage. No independent runtime reproduction, production assurance or tested integration with other indexed software is claimed.
Source reviewed: 2026-10-06
Source revision / declared version: 46276b4080dadd70ffa2608533139e8df6dcf75d / 0.5.1
Review state / レビュー状態: reviewed_snapshot — applies to the pinned reviewed source only
Observed default-branch revision: 46276b4080dadd70ffa2608533139e8df6dcf75d — equality does not renew review; difference does not establish chronology
Observed GitHub release: v0.5.1 · 2026-10-04T19:24:57Z
Release observation / リリース確認: 2026-11-02T11:09:05+00:00 · current
Package metadata observation: 2026-10-17 metadata-only / 情報のみ: source declaration / ソース宣言 0.5.1 at 46276b4080dadd70ffa2608533139e8df6dcf75d; PyPI collective-intelligence-overlay 0.5.1 observed / 観測済み. No distribution-byte verification, installation or efficacy test / 配布物検証・導入・効果試験なし. Evidence / 根拠: e-final-sw-cio-package-20261017, e-final-sw-cio-pypi-20261017
License: Apache-2.0
Interface schema identity / スキーマ識別子: null — upstream schema has no $id · Bounded lifecycle projections, exact original-byte export and explicit stock assessment; no new wire Record or execution authority.
Inputs: Exact capabilities, original signed records, receiver-local Decisions, receipts, policy coordinates, finite stock observations and explicit source/receipt clocks.
Outputs: Lifecycle, contribution, residual, growth and handoff views; exact original-byte exports. Projection digests differ from DSSE payload digests; handoff grants no authority.
Prerequisites: Python >=3.12. Service paths require owner-controlled PostgreSQL, OPA, keys, policies and installed checkers. Agent/model extras are optional. File owner equality is not authentication.
Effects and trust boundary: Lifecycle inspection/export/handoff are read-only projections. Explicit assess_stock performs qualification and saves Decisions/costs. Separately authorized Executor and agent paths can execute tools and contact configured services; inspection grants no such permission.
First inspection: Read pinned lifecycle concepts/reference, architecture, security and the candidate validation register. Inspect original bytes, receiver policy, coverage, clocks and residuals before assessing stock or considering execution.
Installation guidance: Follow version-pinned upstream instructions after host authorization. Current metadata declares 0.5.1, matching the observed GitHub release; the earlier 0.5.0 candidate evidence remains historical. PyPI bytes/installations were not independently verified here.

evidence-gap-router

Deterministic Python SDK and offline CLI selecting the next acquisition or verification from missing evidence, unfinished checks and a finite host-declared action set. Schema 3 separates observed PASS from explicit finite completion.

Primary editorial role: evidence-and-verification
Limits / unsupported uses: Alpha contract/control implementation, not an agent/model service, network gateway or scheduler. Finite catalogue completion leaves external completeness unknown; semantic truth and statistical independence are not guaranteed. Callback views are not isolation, snapshots are not authenticated history and finite steps do not enforce wall/CPU deadlines. Host owns trust, effects, resource measurements and single-writer use. Pending calls and unknown effects/costs remain incomplete; no automatic retry, refund, crash recovery or exactly-once guarantee. v0.2.4 upstream experiments reported negative results; v0.3.0 ran no new efficacy experiments and its empirical utility is unmeasured. No independent execution, general acceleration or tested integration with other indexed projects is claimed.
Source reviewed: 2026-10-06
Source revision / declared version: 0e5def78422de55db6c80723f28b753ef91e017b / 0.3.0
Review state / レビュー状態: reviewed_snapshot — applies to the pinned reviewed source only
Observed default-branch revision: 0e5def78422de55db6c80723f28b753ef91e017b — equality does not renew review; difference does not establish chronology
Observed GitHub release: v0.3.0 · 2026-10-06T05:12:56Z
Release observation / リリース確認: 2026-11-02T11:09:05+00:00 · current
Package metadata observation: 2026-10-17 metadata-only / 情報のみ: source declaration / ソース宣言 0.3.0 at 0e5def78422de55db6c80723f28b753ef91e017b; PyPI evidence-gap-router 0.3.0 observed / 観測済み. No distribution-byte verification, installation or efficacy test / 配布物検証・導入・効果試験なし. Evidence / 根拠: e-final-sw-egr-package-20261017, e-final-sw-egr-version-20261017, e-final-sw-egr-pypi-20261017
License: Apache-2.0
Interface schema identity / スキーマ識別子: null — upstream schema has no $id · Schema 3 State/PlanInput/Decision and explicit finite CompletionContract; schema-1/2 imports require migration and confer no new authority.
Inputs: Host-owned State, obligations, exact evidence/dependencies, finite ActionCandidate pool, Budget, Policy, registered checkers and explicit CompletionContracts.
Outputs: Decision with action or stop reason, completion assessments and exact residuals; retained issued receipts, checks, invalidations, expenses and schema-3 snapshots.
Prerequisites: Python >=3.12 and Pydantic >=2.12,<3. Host must declare finite scope, qualified checker kinds/scopes and actual callback resource/effect measurements. Old schemas require explicit migration.
Effects and trust boundary: plan and completion assessment are pure. step invokes at most one registered callback; run invokes a bounded sequence whose effects depend on host code. check-data reads selected local files and snapshot helpers write explicit destinations. Core offline paths require no model/network service. This review executed none of the indexed software.
First inspection: Read pinned completion, API, security and migration contracts; inspect current missing material/checks, issued permissions, exact dependencies and unknown effects before explicit continuation.
Installation guidance: Use version-pinned upstream installation instructions after host review. Source version, observed GitHub release and actual installed package are separate observations; no PyPI artifact was installed or independently verified here.

Proposed evidence-preserving workflow

Generate → Verify → Reuse → Account → Reallocate is a conceptual workflow, not an executable pipeline. Preserve original bytes, identities, clocks, costs and unresolved obligations at each handoff. Use the typed relationships below before connecting any two tools. A schema match alone grants neither semantic acceptance nor execution authority.

Version-bound interoperability

“Tested upstream” attributes native fixture checks to the cited project; none of those packages was executed in this website task. Document inspection, index regression tests and native integration execution are different evidence origins. Direction matters: imports_artifact_from points from consumer to producer; exports_proposal_to points from producer to host. Compatibility is neither transitive nor automatically renewed by a release.

sw-ccr → paper-growth

based_on · documented · source_inspected

Repository cites this research for a bounded role.

3b4702454f732c7a0a9f30d87483384dc8a281eb → 10.5281/zenodo.22604358 / archive 7bd9fe246ae0f5a4bf6564b588c0426b5b7f29bd

citation

A citation does not establish complete implementation.

e-collective-capability-runtime-readme-md

sw-pic → paper-ecpt

based_on · documented · source_inspected

Repository cites this research for a bounded role.

55cd5f219d02f02927e247317156e757b1d5d6da → 10.5281/zenodo.20535654 / archive 7bd9fe246ae0f5a4bf6564b588c0426b5b7f29bd

citation

A citation does not establish complete implementation.

e-percolation-inversion-compiler-readme-md

sw-vek → paper-vet

based_on · documented · source_inspected

Repository cites this research for a bounded role.

4008e311ceb16edd6e74d9f71341a90120c0d046 → 10.5281/zenodo.21147093 / archive 7bd9fe246ae0f5a4bf6564b588c0426b5b7f29bd

citation

A citation does not establish complete implementation.

e-verification-ecology-kit-readme-md

sw-alt → paper-alt

based_on · documented · source_inspected

Repository cites this research for a bounded role.

e0296486bb568fdeeda449dada3bc67753ddfe7f → 10.5281/zenodo.20476200 / archive 7bd9fe246ae0f5a4bf6564b588c0426b5b7f29bd

citation

A citation does not establish complete implementation.

e-alt-foundry-kernel-readme-md

sw-cait → paper-cait

based_on · documented · source_inspected

Repository cites this research for a bounded role.

0ecb3afd2ea1d99cedcabd7a9e01e405ae1808a7 → 10.5281/zenodo.20061296 / archive 7bd9fe246ae0f5a4bf6564b588c0426b5b7f29bd

citation

A citation does not establish complete implementation.

e-cait-certificate-schema-readme-md

sw-cpcf → paper-growth

based_on · documented · source_inspected

Repository cites this research for a bounded role.

340b4899d5b0892e06d7d2c1ab1f1d48df15e981 → 10.5281/zenodo.22604358 / archive 7bd9fe246ae0f5a4bf6564b588c0426b5b7f29bd

citation

A citation does not establish complete implementation.

e-collective-phase-control-fabric-readme-md · e-collective-phase-control-fabric-docs-growth-research-mapping-md

sw-oawm → paper-workflow-library

based_on · documented · source_inspected

Repository cites this research for a bounded role.

414299d4d15b5de434f5fb1ccbb4ba17df9155ff → 10.5281/zenodo.19690749 / archive 7bd9fe246ae0f5a4bf6564b588c0426b5b7f29bd

citation

A citation does not establish complete implementation.

e-observable-agent-workflow-memory-readme-md · e-observable-agent-workflow-memory-docs-theory-sources-md

sw-audit → paper-audit-closed

based_on · documented · source_inspected

Repository cites this research for a bounded role.

1e844d1bca18d7bb298aa198e1250bb07bd7ea44 → 10.5281/zenodo.18728589 / archive 7bd9fe246ae0f5a4bf6564b588c0426b5b7f29bd

citation

A citation does not establish complete implementation.

e-audit-closed-ai-scientist-readme-md

sw-loscr → paper-loscr

based_on · documented · source_inspected

Repository cites this research for a bounded role.

334307b59a4870b90f6162705c793bb5615a555a → 10.5281/zenodo.19836225 / archive 7bd9fe246ae0f5a4bf6564b588c0426b5b7f29bd

citation

A citation does not establish complete implementation.

e-loscr-readme-md

sw-alt → sw-ccr

exports_proposal_to · tested_upstream · upstream_reported

ALT 0.5.0 native open/unleased task parses against CCR 1.8.0; ALT sidecar is additional.

0.5.0 → 1.8.0 / 1591e88e3b05f9b5fb06b2d0c749aad8bc0909be

ccr.task.v0.1, open status, unleased proposal

Custom ALT sidecar admission, leases, rewards and execution are not supplied.

e-alt-foundry-kernel-docs-reuse-interchange-md · e-alt-release · e-alt-native-test

sw-alt → sw-vek

imports_artifact_from · tested_upstream · upstream_reported

Capacity report schema and scope/clock/work-conservation checks on pinned fixtures.

0.5.0 → 1.3.0 / b07998135cb9dccde1e67db3c6ed77fdec847e09

scope, clock, work conservation

Counters are supplied, not replayed; no observed rate or empirical independence.

e-alt-foundry-kernel-docs-reuse-interchange-md · e-alt-release · e-alt-native-test

sw-alt → sw-vek

tested_interchange_with · tested_upstream · upstream_reported

Finite synthetic check demand accepted by the released VEK checker.

0.5.0 → 1.3.0 / b07998135cb9dccde1e67db3c6ed77fdec847e09

native contract, schedule

Host residual/service/unit registration and real observed capacity are absent.

e-alt-foundry-kernel-docs-reuse-interchange-md · e-alt-release · e-alt-native-test

sw-alt → sw-cait

tested_interchange_with · tested_upstream · upstream_reported

Supported complete history passes released analyzer/checker; copies do not create assets.

0.5.0 → 0.2.0 / 7d12bcf0bc4c6eae2acdd177175b6be414a274f6

unique formation, costs, receiver checks, requests/results, withdrawal

Expiry, contradiction, refresh, correction and some parent mappings remain partial; no cost-to-capital or arrival claim.

e-alt-foundry-kernel-docs-reuse-interchange-md · e-alt-release · e-alt-native-test

sw-vek → sw-ccr

exports_proposal_to · documented · source_inspected

Capacity extension accompanies schema-valid candidate task.

1.3.0 → 6d8a30820806044b3a4b18d499fbe89bef83fbf2

ccr.task.v0.1, duration rounding, x_vek_capacity

CCR base schema does not enforce extensions; custom admission and atomic host reservation required.

e-verification-ecology-kit-docs-capacity-interchange-md

sw-cait → sw-vek

imports_artifact_from · documented · source_inspected

Pinned capacity schema preserves counters and rejects a false guarantee fixture.

0.2.0 → 1.3.0 / b07998135cb9dccde1e67db3c6ed77fdec847e09

schema, scope, source digest, work conservation

Original VEK reducer journal is absent; counters are not independently reconstructed.

e-cait-certificate-schema-docs-interchange-md

sw-ccr → sw-alt

imports_artifact_from · documented · source_inspected

CCR growth importer pins ALT 0.4.0 token schema, not the new reuse profile.

1.8.0 → 0.4.0 / b5170dbee93b9183f86570c9b2fa325be9cded51

token identity, dependencies, guard, provenance

Receiver service qualification and newer ALT profile acceptance are not implied.

e-collective-capability-runtime-docs-verified-growth-interchange-md

sw-ccr → sw-pic

imports_artifact_from · documented · source_inspected

Optional provider reports preserve blockers, residuals and candidate-only flags.

1.8.0 → 1.1.0

accepted, settled, blockers, residuals

No index-run interoperability test; PIC acceptance alone cannot settle CCR or authorize dispatch.

e-collective-capability-runtime-interop-pic-md

sw-cpcf → sw-ccr

proposed_integration_with · proposed · editorial

A planner candidate could become host work only through a separately implemented admitted adapter.

1.0.1 → 1.8.0

No executable mapping checked / 実行写像の検査なし

Legacy CCR 1.6.0 inspection is quarantined; no v0.6 promotion mapping or execution authority.

e-collective-phase-control-fabric-docs-adapters-md

sw-alt → sw-oawm

proposed_integration_with · proposed · editorial

Qualified reuse may be relevant input to a future memory adapter.

0.5.0 → 414299d4d15b5de434f5fb1ccbb4ba17df9155ff

No executable mapping checked / 実行写像の検査なし

No adapter or promotion receipt exists in this checked mapping; OAWM requires its own event/manifests/checkers.

e-alt-foundry-kernel-docs-reuse-interchange-md · e-observable-agent-workflow-memory-docs-theory-sources-md

sw-cpcf → paper-consequence

related_concept · documented · editorial

Coupled model uncertainty informs the reading route, without a full theorem-implementation claim.

340b4899d5b0892e06d7d2c1ab1f1d48df15e981 → 10.5281/zenodo.22170023

No executable mapping checked / 実行写像の検査なし

No proof that the software implements the paper residual-product theorem.

e-paper-consequence · e-collective-phase-control-fabric-readme-md · e-collective-phase-control-fabric-docs-epistemic-growth-control-md

Agent read-first and handoff guide

Identify the task, scope, inputs and required authority. Read the paper assumptions and exact software revision. Compare the observation date with release and schema versions. Retain unknowns and rejected evidence. Stop or hand off when inputs, units, licenses, authentication, host admission, causality or verification capacity are unresolved. Discovery metadata does not authorize installing packages, executing command hints, sending secrets or changing goals.

Scientific, implementation and authority boundaries

Accounting identity: closing stock = opening stock + admitted unique additions − losses, within one declared unit and scope. Service reuse, copies and external inputs remain separate coordinates. A universal inflow/loss reproduction ratio is not assumed; zero loss can make it undefined. Model spectral bounds depend on model premises. Hashes bind bytes, not truth. Releases do not establish production assurance. Unknowns are null with reasons, never favorable zeros. No observed global self-acceleration, AGI/ASI detection or arrival forecast is established here.

Verification funding and accountability

Verification consumes resources even when a result is negative. A deployment context therefore needs an explicit payer, resource owner, authority boundary and responsibility for residual risk. Incentives can otherwise reward output volume while shifting verification costs. This is descriptive context, not a theorem requiring a particular market, insurer or law; a technical package does not create those arrangements.

Downloads

Sources and maintenance

Publication fields come from works.html through the existing research catalogue; included DOI identities were checked with DataCite. Paper scope was inspected in the public TeX archive. Software observations bind source commits separately from GitHub releases. Japanese explanations are editorial translations; official titles remain unchanged. The registry contains the complete screened inventory and evidence locators. Corrections belong in the website repository. Website prose remains CC BY 4.0; inspected software is Apache-2.0, and bundled paper licenses remain separate. For paper claims, cite the original DOI, checked version and passage. For software behavior, cite the repository, reviewed commit, package/release version and test scope. To cite this editorial index, record its URL, editorial date, resource ID and registry content_digest or an actual Git commit. The live website can change; a commit permalink is a fixed snapshot. Historical review does not guarantee future compatibility, efficacy or safety. Metadata refresh does not renew scientific review. See the maintenance document for restoration and the 2026-10-17 audit.

Scanned: 234 research records; 57 repositories. Selected: 41 resources. Full-text scope review applies only to selected papers; remaining candidate matches are explicitly unresolved in the registry.

Boundary Exchange / viability: unresolved — No exact Boundary Exchange title/DOI in the scanned catalogue. Do not invent a paper identity.
CGT bandwidth / availability / comparability supplements: unresolved — Distinct TeX manuscripts exist in the source archive, but only the canonical CGT work is a distinct local catalogue record. Do not collapse supplements or invent DOI bindings.
OASG 10.5281/zenodo.20107661: resolved_software_kind — DataCite classifies this as Software, matching works.html. Included through the repository, excluded from paper bibliography.
10.5281/zenodo.20199440: concept_record_date_discrepancy — CGT concept/series DOI has multiple HasVersion records. Existing catalogue first-publication date is 2026-05-15; current DataCite Issued date is 2026-05-28. The exact version binding of the inspected TeX is unresolved. Bibliography includes only verified title/author/year/DOI; it does not claim the DOI identifies these exact source bytes.
55 historical source evidence hashes (2026-10-17 audit): raw_byte_binding_corrected — Historical hashes match CRLF-normalized text rather than raw HTTP payloads. Original records remain; corresponding evidence IDs ending -raw-20261017 bind the raw bytes. / 旧ハッシュ55件はCRLF変換後のテキストに一致。元記録は保持し、末尾 -raw-20261017 の根拠記録で配信バイト列を同定する。
sw-ccr: observed_revision_differs — Observed default-branch revision differs from the reviewed snapshot. Ancestry/order and behavior at that revision were not checked. / 観測した既定ブランチのコミットはレビュー対象と異なる。祖先関係・新旧順序・その版の挙動は未確認。
sw-pic: observed_revision_differs — Observed default-branch revision differs from the reviewed snapshot. Ancestry/order and behavior at that revision were not checked. / 観測した既定ブランチのコミットはレビュー対象と異なる。祖先関係・新旧順序・その版の挙動は未確認。
sw-vek: observed_revision_differs — Observed default-branch revision differs from the reviewed snapshot. Ancestry/order and behavior at that revision were not checked. / 観測した既定ブランチのコミットはレビュー対象と異なる。祖先関係・新旧順序・その版の挙動は未確認。
sw-alt: observed_revision_differs — Observed default-branch revision differs from the reviewed snapshot. Ancestry/order and behavior at that revision were not checked. / 観測した既定ブランチのコミットはレビュー対象と異なる。祖先関係・新旧順序・その版の挙動は未確認。
sw-cait: observed_revision_differs — Observed default-branch revision differs from the reviewed snapshot. Ancestry/order and behavior at that revision were not checked. / 観測した既定ブランチのコミットはレビュー対象と異なる。祖先関係・新旧順序・その版の挙動は未確認。
sw-oawm: observed_revision_differs — Observed default-branch revision differs from the reviewed snapshot. Ancestry/order and behavior at that revision were not checked. / 観測した既定ブランチのコミットはレビュー対象と異なる。祖先関係・新旧順序・その版の挙動は未確認。
sw-oasg: observed_revision_differs — Observed default-branch revision differs from the reviewed snapshot. Ancestry/order and behavior at that revision were not checked. / 観測した既定ブランチのコミットはレビュー対象と異なる。祖先関係・新旧順序・その版の挙動は未確認。
sw-loscr: observed_revision_differs — Observed default-branch revision differs from the reviewed snapshot. Ancestry/order and behavior at that revision were not checked. / 観測した既定ブランチのコミットはレビュー対象と異なる。祖先関係・新旧順序・その版の挙動は未確認。
sw-cmgl: observed_revision_differs — Observed default-branch revision differs from the reviewed snapshot. Ancestry/order and behavior at that revision were not checked. / 観測した既定ブランチのコミットはレビュー対象と異なる。祖先関係・新旧順序・その版の挙動は未確認。
sw-memoryflow: observed_revision_differs — Observed default-branch revision differs from the reviewed snapshot. Ancestry/order and behavior at that revision were not checked. / 観測した既定ブランチのコミットはレビュー対象と異なる。祖先関係・新旧順序・その版の挙動は未確認。
sw-pfg: observed_revision_differs — Observed default-branch revision differs from the reviewed snapshot. Ancestry/order and behavior at that revision were not checked. / 観測した既定ブランチのコミットはレビュー対象と異なる。祖先関係・新旧順序・その版の挙動は未確認。
sw-fost: observed_revision_differs — Observed default-branch revision differs from the reviewed snapshot. Ancestry/order and behavior at that revision were not checked. / 観測した既定ブランチのコミットはレビュー対象と異なる。祖先関係・新旧順序・その版の挙動は未確認。
sw-atrb: observed_revision_differs — Observed default-branch revision differs from the reviewed snapshot. Ancestry/order and behavior at that revision were not checked. / 観測した既定ブランチのコミットはレビュー対象と異なる。祖先関係・新旧順序・その版の挙動は未確認。
sw-checkedflow: observed_revision_differs — Observed default-branch revision differs from the reviewed snapshot. Ancestry/order and behavior at that revision were not checked. / 観測した既定ブランチのコミットはレビュー対象と異なる。祖先関係・新旧順序・その版の挙動は未確認。

Maintenance / 保守 · Corrections / 訂正