Datasets:
Formats:
json
Size:
< 1K
Tags:
ai-agent
agent-knowledge-cycle
knowledge-cycle
self-improvement
cognitive-economy
signal-first
License:
Upload graph.jsonld with huggingface_hub
Browse files- graph.jsonld +4 -4
graph.jsonld
CHANGED
|
@@ -645,7 +645,7 @@
|
|
| 645 |
"description": "Skill scaffolding Scaffold Dissolution's model-generation trigger (a snapshot, mutable; ADR-0019). Since 2026-09-25 (v2) it is the single entry for generation-change audits and runs two checks in one pass: a runtime cross-check of self-authored assets against the system prompt and tool descriptions actually loaded at inference time (conflict / redundancy), and a dated-pattern scan delegated to the model vendor's own prompt-audit procedure bundled with Claude Code — the pattern table is read at run time, not copied, so the audit's own catalog of outdated patterns has dissolved downward into the vendor. It judges each finding with an intent / evidence / freshness / expiry frame, applies high- and medium-confidence line fixes under explicit consent, and hands whole-asset verdicts to the per-asset-class stocktake skills. The instrument that produces generation review, the fourth retirement evidence class (ADR-0023).",
|
| 646 |
"url": "https://github.com/shimo4228/generation-audit",
|
| 647 |
"extends": "https://doi.org/10.5281/zenodo.19200726",
|
| 648 |
-
"implements": "https://shimo4228.github.io/shimo4228/vocab#concept/scaffold-dissolution",
|
| 649 |
"groundedIn": "https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/adr/0023-generation-review-as-a-fourth-evidence-class.md"
|
| 650 |
},
|
| 651 |
{
|
|
@@ -773,7 +773,7 @@
|
|
| 773 |
"@id": "https://github.com/shimo4228/claude-harness",
|
| 774 |
"@type": ["EcosystemRepo", "SoftwareSourceCode"],
|
| 775 |
"name": "claude-harness",
|
| 776 |
-
"description": "Bundled distribution of the author's Claude Code harness — the running system the AKC cycle operates on and co-evolves with. As of 2026-09 it ships an origin-tracked library
|
| 777 |
"url": "https://github.com/shimo4228/claude-harness",
|
| 778 |
"extends": "https://doi.org/10.5281/zenodo.19200726",
|
| 779 |
"implements": [
|
|
@@ -815,7 +815,7 @@
|
|
| 815 |
"name": "llm-as-judge",
|
| 816 |
"description": "Agent Skill for designing LLM-as-judge evaluators: binary checks as evidence, one named holistic verdict, no score aggregation. It grounds the LLM side of the code-LLM judge pattern (ADR-0008) — the bounded, named options the judge chooses among — and states that the enforce side (code taking the deterministic action) is outside its scope. Also bundled in claude-harness. Relation fact only.",
|
| 817 |
"url": "https://github.com/shimo4228/llm-as-judge",
|
| 818 |
-
"implements": "https://shimo4228.github.io/shimo4228/vocab#akc/pattern/judge"
|
| 819 |
},
|
| 820 |
{
|
| 821 |
"@id": "https://github.com/shimo4228/akc-mcp",
|
|
@@ -1274,7 +1274,7 @@
|
|
| 1274 |
"@id": "https://github.com/shimo4228/akc-cycle",
|
| 1275 |
"@type": "TechArticle",
|
| 1276 |
"name": "akc-cycle (single-file cycle rules)",
|
| 1277 |
-
"description": "The install entry point for the AKC cycle: a single rules file copy-installable to any agent's rules directory (installs all six phases without the skill repos), plus
|
| 1278 |
"groundedIn": "https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/adr/0009-akc-is-a-cycle-not-a-harness.md"
|
| 1279 |
},
|
| 1280 |
|
|
|
|
| 645 |
"description": "Skill scaffolding Scaffold Dissolution's model-generation trigger (a snapshot, mutable; ADR-0019). Since 2026-09-25 (v2) it is the single entry for generation-change audits and runs two checks in one pass: a runtime cross-check of self-authored assets against the system prompt and tool descriptions actually loaded at inference time (conflict / redundancy), and a dated-pattern scan delegated to the model vendor's own prompt-audit procedure bundled with Claude Code — the pattern table is read at run time, not copied, so the audit's own catalog of outdated patterns has dissolved downward into the vendor. It judges each finding with an intent / evidence / freshness / expiry frame, applies high- and medium-confidence line fixes under explicit consent, and hands whole-asset verdicts to the per-asset-class stocktake skills. The instrument that produces generation review, the fourth retirement evidence class (ADR-0023).",
|
| 646 |
"url": "https://github.com/shimo4228/generation-audit",
|
| 647 |
"extends": "https://doi.org/10.5281/zenodo.19200726",
|
| 648 |
+
"implements": ["https://shimo4228.github.io/shimo4228/vocab#concept/scaffold-dissolution", "https://shimo4228.github.io/shimo4228/vocab#akc-phase/curate"],
|
| 649 |
"groundedIn": "https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/adr/0023-generation-review-as-a-fourth-evidence-class.md"
|
| 650 |
},
|
| 651 |
{
|
|
|
|
| 773 |
"@id": "https://github.com/shimo4228/claude-harness",
|
| 774 |
"@type": ["EcosystemRepo", "SoftwareSourceCode"],
|
| 775 |
"name": "claude-harness",
|
| 776 |
+
"description": "Bundled distribution of the author's Claude Code harness — the running system the AKC cycle operates on and co-evolves with. As of 2026-09 it ships an origin-tracked skill library (the AKC cycle-phase skills among them, each also published as its own standalone repo and bundled in the akc-cycle plugin), a worldview rules layer (always-loaded rules such as LLM-first artifact readability and knowledge staleness, whose AKC-side judgments are ADR-0025 and ADR-0026), deterministic hooks and machine gates, and a public rfcs/ task ledger operated by the judge/build/human three-role triage loop (ADR-0024), whose build sessions run as cloud sessions by default since 2026-09-24. Since 2026-09-15 its always-loaded boundary rule (rules/common/boundary.md) is the running instance of the human approval gate (ADR-0005): one list of what is handed to the human (publication, billing, outbound messages, and unattended changes to rules, hooks, permissions, scheduled tasks, the verify gate, ADRs, and skills), the risks taken without asking, and the conditions for stopping. The author-calibrated-eval skill (2026-09-25) grounds intent alignment in an evaluation loop for generated reading material: the author's blind reading is the target for what is worth reading, an LLM judge only screens fidelity to the source material, and code checks form. The jev-judgment-design skill (2026-09-23) grounds the code-LLM judge pattern (ADR-0008): closed judgments move to a typed judgment model that answers with probabilities, and code decides acceptance. Three older skills ground concepts the same way: harness-boundary asks, before a mechanism enters the harness, whether the next model generation will make it unnecessary and can return Make temporary or Delete — Scaffold Dissolution applied at design time rather than at retirement; review-to-lint moves the machine-decidable items of an LLM reviewer's checklist into a deterministic script and leaves the reviewer only the semantic checks — code-LLM layering applied to review; and adr-writer makes every decision record carry a Review-when section of expiry conditions and routes decisions below the ADR bar into commit bodies that carry the same Context / Decision / Review-when triple — expiry-conditioned knowledge in its running form. Its own ADR series records the harness-side lineage of these practices.",
|
| 777 |
"url": "https://github.com/shimo4228/claude-harness",
|
| 778 |
"extends": "https://doi.org/10.5281/zenodo.19200726",
|
| 779 |
"implements": [
|
|
|
|
| 815 |
"name": "llm-as-judge",
|
| 816 |
"description": "Agent Skill for designing LLM-as-judge evaluators: binary checks as evidence, one named holistic verdict, no score aggregation. It grounds the LLM side of the code-LLM judge pattern (ADR-0008) — the bounded, named options the judge chooses among — and states that the enforce side (code taking the deterministic action) is outside its scope. Also bundled in claude-harness. Relation fact only.",
|
| 817 |
"url": "https://github.com/shimo4228/llm-as-judge",
|
| 818 |
+
"implements": ["https://shimo4228.github.io/shimo4228/vocab#akc/pattern/judge", "https://shimo4228.github.io/shimo4228/vocab#akc-phase/measure"]
|
| 819 |
},
|
| 820 |
{
|
| 821 |
"@id": "https://github.com/shimo4228/akc-mcp",
|
|
|
|
| 1274 |
"@id": "https://github.com/shimo4228/akc-cycle",
|
| 1275 |
"@type": "TechArticle",
|
| 1276 |
"name": "akc-cycle (single-file cycle rules)",
|
| 1277 |
+
"description": "The install entry point for the AKC cycle: a single rules file copy-installable to any agent's rules directory (installs all six phases without the skill repos), plus a Claude Code plugin carrying the skills of all six phases, grouped by phase (the companion skills formerly listed apart are placed in the phase where they are used, among them skill-creator under Extract, verify-bootstrap under Maintain and measurement-discipline under Measure; a mutable snapshot per ADR-0019). Published as a standalone repository that is also its own plugin marketplace; the AKC repository points to it instead of carrying a divergent copy. Canonical sources are the author's live harness, mirrored one-way. Operational form of scaffold dissolution — the cycle floor becomes a rules-layer concern, with the skill layer as an optional plugin on top.",
|
| 1278 |
"groundedIn": "https://github.com/shimo4228/agent-knowledge-cycle/blob/main/docs/adr/0009-akc-is-a-cycle-not-a-harness.md"
|
| 1279 |
},
|
| 1280 |
|