v2.0.0 — genre-neutral cycle + security triplet extraction

Major release — breaking positioning change

Completes ADR-0009's rhetorical demotion of Security by Absence by extracting the contemplative-agent-specific security triplet from AKC, and declares genre neutrality as a first-class positioning via ADR-0011.

AKC's identity contract has shifted. Before: "a knowledge cycle for AI agents, with Security by Absence as a first-class design principle." After: "a knowledge cycle for AI agents — mechanism-only, genre-neutral about what knowledge body flows through it." Academic citations resolving to v1.x will see a different Design Principles list and a different ADR set.

R1 — Security triplet extraction

Removed

• [ADR-0001 Security by Absence] — genre-specific to the contemplative-agent lineage; encoded a threat model appropriate to local autonomous agents making side effects in the world, not a universal requirement of any knowledge cycle.
• [ADR-0006 Single External Adapter] — same genre lineage.
• [ADR-0007 Untrusted Content Boundary] — same genre lineage.
• [docs/skills/llm-agent-security-principles.md] — the design-pattern skill that bundled the three security ADRs' concrete defense patterns.

The extracted files are preserved locally (outside the public repository) for future re-expression in the contemplative-agent lineage. External readers who need the content should consult the v1.8.1 tag.

Internal references cleaned

Internal references in ADR-0002, ADR-0008, ADR-0010, and docs/skills/code-and-llm-collaboration.md were tightened to no longer point at the removed ADRs. examples/minimal_harness/knowledge_store.py retains the forbidden-substring validation behavior; only the ADR-0007 pointer was replaced with a neutral description.

R2 — Genre neutrality as positioning

Added

• [ADR-0011 Cycle Applies to Any Knowledge Body] — declares the cycle genre-neutral with three falsifiable commitments:

  1. No future AKC ADR encodes domain-specific content beyond what the six phases require.
  2. docs/skills/ hosts only cycle-mechanic skills.
  3. External projects may cite AKC as their cycle substrate while retaining their own content.

• [ADR-0009 2026-04-19 addendum] — records the concrete security-triplet extraction that completes ADR-0009's demotion. Documents the consequences (numbering gaps, principle renumbering, supersession of the original "unchanged" claim) and cross-links to ADR-0011.

• [examples/constitution_amend/] — a new examples entry that points at contemplative-agent's Constitution-amend workflow as a concrete instance of the AKC cycle applied to constitutional values rather than behavioral patterns. The directory ships no implementation — the canonical code is in the contemplative-agent repository — so AKC retains genre neutrality while making the structural claim of ADR-0011 observable. Paired descriptively with minimal_harness/ (cycle on behavioral patterns).

Counts

	v1.8.1	v2.0.0
ADRs	10	8 (permanent gaps at 0001, 0006, 0007)
Design principles	9	8 (Security by Absence removed; Code-LLM Layering → #7, Human cognitive resource → #8)
Design-pattern skills	4	3
JSON schemas	2	2
Reference implementations	1	1 (minimal_harness/) + 1 descriptive example (constitution_amend/)

Migration notes

For external citations

Academic and web citations pinned to v1.x continue to describe the state at that version. No action is required. v2.0.0 is a positioning change, not a correction.

For external links

Links pointing at docs/adr/0001-security-by-absence.md, docs/adr/0006-single-external-adapter.md, or docs/adr/0007-untrusted-content-boundary.md at main HEAD will 404. Commit-pinned GitHub permalinks continue to resolve to the v1.x content.

For downstream projects

Projects that adopted AKC's security-by-absence design principle may continue to do so — the principle itself is sound — but should no longer cite AKC as its source. Contemplative-agent is the upstream home.

Why v2.0.0 (not v1.9.0)

AKC's identity contract changes. ADR-0009 established the precedent for positioning-breaking changes warranting a version bump; the security triplet extraction completes what ADR-0009 started, and the identity-contract shift is the direct reason this is a major bump rather than a minor one. CHANGELOG entry spells out the reasoning.

Zenodo

This release will auto-archive to Zenodo, producing a new version DOI linked to the base DOI 10.5281/zenodo.19200727. The v1.x DOIs remain valid for time-pinned citations.

🤖 Co-authored with Claude Opus 4.7 (1M context)

v1.8.1 — signal-first-research pattern skill + inspiration upstream

Patch release

Completes the three-layer structure introduced by v1.8.0 (principle → pattern → implementation) and records the self-referential origin loop that produced ADR-0010.

No new philosophical claims; this release wires up the parts v1.8.0 promised and names the origin story that made ADR-0010 possible in the first place.

What changed

Added

• docs/skills/signal-first-research.md — the design-pattern skill paired 1:1 with ADR-0010. Fourth entry in the docs/skills/ family, alongside when-code-when-llm, code-and-llm-collaboration, and llm-agent-security-principles. Covers:

  • Define the signal before the search
  • The search-wide / intake-narrow asymmetry (two behaviors, two phases)
  • Three abstract worked examples: daily topic digest, session-scoped filter, exploration mode
  • A six-question diagnostic checklist for spotting filters that have quietly broken
  • The archive-first vs. signal-first distinction

  Written as a pure pattern guide — no personal project URLs in the skill body. Lived examples live in docs/inspiration.md.

• New docs/inspiration.md section: "Upstream: a daily signal filter" — documents that AKC's origin chain extends one hop further back than Contemplative Agent. Before Contemplative Agent existed, the author was running a daily research pipeline that (a) surfaced Laukkonen et al. (2025) and seeded Contemplative Agent, and (b) later generated the report that triggered ADR-0010. The pipeline was skillified in April 2026 as claude-skill-daily-research.

Changed

• ADR-0010 Notes — extended with an Origin loop observation:

  The signal-first Research principle that ADR-0010 names was being practiced by the pipeline that eventually produced the report that named it. The principle was lived before it was stated; this ADR is the point at which the practice recognized itself.

• README / README.ja.md / docs/skills/README.md / llms.txt — design-pattern skill count: 3 → 4. CITATION.cff bumped to v1.8.1.

Not changed

• Tagline, six phases, ADRs 0001–0009, design principles, reference implementation — all unchanged from v1.8.0.
• ADR-0010's Decision body — still generic ("daily-report workflows ... are now a canonical example"). Specific implementation references belong in docs/inspiration.md, not in the principle itself.

A note on portability

This release was informed by a small methodological lesson: skills should be portable. A design-pattern skill that embeds links to the author's personal repositories ties itself to one filesystem and one project's lifecycle. docs/skills/signal-first-research.md therefore contains no external project URLs; it is readable and usable without knowing anything about the author's claude-skill-daily-research or Contemplative Agent. Those personal references — and the causal chain between them — are deliberately concentrated in docs/inspiration.md, which is the narrative-historical document where author-specific context belongs.

🤖 Co-authored with Claude Opus 4.7 (1M context)

v1.8.0 — human cognitive resource as central constraint

Architectural depth release

AKC surfaces the implicit design philosophy that has shaped every phase since inception: human cognitive resources — attention and judgment — are the central constraint that does not scale with the model.

This is a documentation release. No code changes. What AKC centers on becomes explicit.

See ADR-0010 for the full rationale.

Why now

Three pressures forced the surfacing:

1. The LLM-Wiki wave (Karpathy's LLM Wiki, claude-obsidian, MemPalace) popularized knowledge bases written by LLMs for humans. Their superficial overlap with AKC's Curate and Promote phases invited a wrong reading: is AKC just a smaller LLM Wiki? The answer is no — they optimize opposite sides of scarcity. LLM Wikis solve information-side scarcity. AKC solves attention-side scarcity. Stating that required a philosophical claim the repository did not yet contain.
2. The origin of + "daily research" + — a workflow where one narrowly-scoped report per day is filtered for personal relevance. It was conceived after hearing about a high-capacity knowledge worker processing ~100 deep-research reports per day. The author's honest admission — I cannot digest that volume — became a design input, not a deficit to overcome.
3. Why dialogue is front-loaded. Implementation started without sufficient intent-alignment produces work that does not match the user's intent, requiring expensive downstream correction cycles. Pre-implementation dialogue looks expensive; it is actually the cheaper path.

The common root: cognitive resources are scarce on the human side and get relatively scarcer as agent capability grows.

What changed

Added

• ADR-0010 Human Cognitive Resource as Central Constraint — names the implicit philosophy, catalogues three mechanisms (signal-first Research, cognitive economy as Design Principle #9, pre-implementation dialogue as investment), and records a geo-writer before/after snapshot as a self-application of the Measure phase to AKC's own docs.
• Design Principle #9 — Human cognitive resource is the bottleneck.
• New "Whose cognitive budget is the cycle protecting?" subsection in + "README.md" + "Why a cycle?" — states that the cycle's target resource is human attention, with one line per phase tying the design back to that target.
• + "README.ja.md" + — full Japanese translation with English/Japanese switcher.

Changed

• + "docs/akc-cycle.md" + Research phase retitled "Search broadly, filter by signal." New step 0 asks the agent to define the signal (what information would change the next action) before searching.
• Research row in both the cycle table and the rules table restated as "Search broadly, filter by signal — intake only what would change the next action."
• CITATION.cff bumped to v1.8.0. Abstract extended with the cognitive-economy framing. Added keywords: + "signal-first" + , + "cognitive-economy" + .
• llms.txt opening summary extended with the cognitive-resource axis. ADR-0010 added.
• README counts: 8 → 9 design principles, 9 → 10 ADRs.

Not changed

• The six phases (Research → Extract → Curate → Promote → Measure → Maintain)
• ADR-0001 through ADR-0009
• Reference implementation
• Tagline — "A knowledge cycle for AI agents — one that grows with the people who shape it" is retained. ADR-0010 adds depth beneath the ADR-0009 framing rather than replacing it.

Relationship to prior releases

Version	Positioning move
v1.7.0	What AKC is: a cycle, not a harness (ADR-0009)
v1.8.0	Why the cycle is shaped the way it is: human cognitive resource is the central constraint (ADR-0010)

v1.7.0 named the direction of the cycle (bidirectional growth). v1.8.0 names the economics that make the direction possible (the human budget is the scarce input).

Why a minor bump

Pure semver would call this a patch (docs only). v1.8.0 is a minor bump because the central philosophy of AKC is being named for the first time — future citations will encounter an architecture whose implicit frame has been made explicit. The distinction matters for academic citation and for readers situating AKC relative to the LLM-Wiki wave.

🤖 Co-authored with Claude Opus 4.7 (1M context)

v1.7.0 — AKC is a cycle, not a harness

Conceptual repositioning

AKC is no longer described as a "self-improving harness." It is now a knowledge cycle for AI agents — one that grows with the people who shape it.

This is a positioning release. No code changes — but what AKC is changed.

See ADR-0009 for the full rationale.

Why

The "self-improving harness" framing created two persistent problems:

1. Layer confusion — calling AKC a harness put it in the same conceptual layer as Everything Claude Code (ECC), making it look like a competitor when AKC actually runs on top of harnesses to keep them coherent over time.
2. Two-pillar drift — presenting Security by Absence as a co-equal definitional element diluted the cycle, which is the actual contribution.

A third issue surfaced as the project matured: the cycle does not just improve the agent. It also changes the human running it. Curate and Promote sharpen judgment; Research develops intuition; Measure teaches what makes a good rule. None of this was articulated in the v1.6 README.

What changed

• Tagline: "memory-centric, self-improving harness" → "knowledge cycle for AI agents — one that grows with the people who shape it"
• "What is AKC?" rewritten — cycle is now the sole defining characteristic
• Security by Absence — demoted from co-equal pillar to Design Principle #7. Still load-bearing, still in ADR-0001
• Bidirectional growth — made explicit in README and tagline
• New "Acknowledgments" section — explicit recognition of ECC as the practical foundation AKC grew out of
• New "References" section — post-hoc theoretical resonances (Thompson's Mind in Life, Laukkonen et al.'s A Beautiful Loop)
• ADR-0009 added with the full repositioning rationale
• akc-cycle.md — Security by Absence section removed (it remains as ADR-0001 and Design Principle #7)
• CITATION.cff abstract rewritten; harness keyword removed; bidirectional-growth and human-ai-collaboration added

What did NOT change

• The six phases (Research → Extract → Curate → Promote → Measure → Maintain)
• The cycle skills (search-first, learn-eval, skill-stocktake, rules-distill, skill-comply, context-sync)
• ADRs 0001–0008
• Schemas and reference implementation
• Security by Absence as a load-bearing guarantee

Why a minor bump

Pure semver would call this a patch (docs only). v1.7.0 is a minor bump because the README is part of the project's identity contract — what AKC is changed in this release. The distinction matters for academic citation: work referencing v1.6 encounters the harness framing, while v1.7+ encounters the cycle framing.

🤖 Co-authored with Claude Opus 4.6 (1M context)

v1.6.0 — ADR-0008 Code-LLM Layering + design-pattern skills

Additive release. Introduces ADR-0008 and a new category of skills
(design-pattern skills) — long-form "how" guides paired 1:1 with ADRs,
shipped inside the repository alongside the ADRs they correspond to.

Added

• ADR-0008 Code and LLM Collaboration — names the principle behind
  ADR-0004 (two-stage distill), ADR-0005 (approval gate), and ADR-0007
  (untrusted boundary): code owns determinism and control flow, LLMs
  own meaning. Catalogues four layering patterns (LLM→Code guard,
  Code filter→LLM, LLM judge + Code enforce, Code orchestrator + LLM
  worker).
• docs/skills/ — new directory for design-pattern skills:
  • when-code-when-llm — per-task decision framework (structural
    vs semantic) with worked examples in both directions
  • code-and-llm-collaboration — per-pipeline layering patterns
    with composition rules and a diagnostic checklist
  • llm-agent-security-principles — concrete defense patterns for
    ADR-0001 / ADR-0006 / ADR-0007 including HTTP, credential, and
    LLM-host hardening
• README Design Principle #8 — Code-LLM Layering.
• README "two kinds of skills" — explicit distinction between cycle
  skills (six external repositories, one per phase) and design-pattern
  skills (in-repo, paired with ADRs).

Changed

• ADR-0001 Security by Absence — added an Audit test section with
  a ready-to-run `git grep` compliance check and a pointer to the
  `llm-agent-security-principles` skill for the full "how" reference.
• docs/adr/README.md — ADR-0008 added to the index.
• llms.txt — ADR-0008 entry and a design-pattern-skills section
  added.

Citation

Zenodo will auto-archive this release as a new version DOI under the
concept DOI 10.5281/zenodo.19200727.

v1.5.0 — Security by Absence, ADRs, reference implementation

Additive release. The six-phase cycle, the six skills, and the rules
install path are unchanged. v1.5 adds a second foundational principle
(Security by Absence), seven ADRs, JSON schemas, and a dependency-free
Python reference implementation adapted from the contemplative-agent
research repository.

Added

• docs/adr/ — seven new architecture decision records:
  • ADR-0001 Security by Absence as a First-Class Principle
  • ADR-0002 Immutable Episode Log as Source of Truth
  • ADR-0003 Three-Layer Distillation (Raw → Knowledge → Identity/Rules)
  • ADR-0004 Two-Stage Distill Pipeline (Free-form → Format)
  • ADR-0005 Human Approval Gate for Behavior-Modifying Changes
  • ADR-0006 Single External Adapter per Agent Process
  • ADR-0007 Untrusted Content Boundary
• schemas/ — JSON Schema definitions for the episode log and
  knowledge store, usable independently of any implementation.
• examples/minimal_harness/ — a dependency-free Python reference
  implementation (~300 lines) demonstrating the three memory layers
  and the two-stage distill pipeline. Runs with a deterministic fake
  LLM; no network or API keys required.
• docs/inspiration.md — prior-art credit to the contemplative-agent
  research repository.
• docs/akc-cycle.md — new 7th principle: Security by Absence.
• llms.txt — llmstxt.org-style machine-readable index.

Changed

• README.md — rewritten intro. AKC is now described as a
  "memory-centric, self-improving harness for AI agents" (scope
  broadened beyond coding agents). The six-skill table, rules install
  path, harness-engineering comparison, and scaffold dissolution
  links are preserved. Design Principles gain a 7th entry.
• CITATION.cff — bumped to v1.5.0, abstract updated to match the
  new positioning, new keywords.

Archived

• docs/history/README-v1.4.md — the v1.4.0 README is preserved
  verbatim for citation integrity.

Citation

Zenodo will auto-archive this release as a new version DOI under the
concept DOI 10.5281/zenodo.19200727.

AKC v1.4.0

Rules-only install path & Scaffold Dissolution

New: docs/rules.md

AKC's six phases distilled into behavioral rules. Copy one file to your agent's rules directory and the cycle runs through natural conversation — no skills, no plugins, no CLI tools.

cp docs/rules.md ~/.claude/rules/common/akc-cycle.md

New: Scaffold Dissolution (Design Principle #6)

Skills are scaffolding. As the user and agent internalize the cycle through practice, skills become unnecessary and rules alone sustain the loop. Success is measured not by skill usage, but by whether the cycle runs naturally without explicit invocation.

See docs/scaffold-dissolution.md for the full concept.

Changes

• Added docs/rules.md — behavioral principles for all six AKC phases with trigger conditions
• Added docs/scaffold-dissolution.md — concept document on skill self-obsolescence
• Added Design Principle #6 (Scaffold dissolution) to README
• Added "Rules — Install the cycle without the skills" section to README with install instructions
• Updated CITATION.cff abstract to reflect new components

AKC v1.3.0

What's Changed

Added Relationship to Harness Engineering section — positioning AKC relative to Mitchell Hashimoto's harness engineering concept (2025). Both aim to make agents more reliable; they differ in layer:

Layer	Question	Addressed by
Harness	"Is this output correct?"	Individual linters, tests, scripts
AKC	"Are the harnesses themselves still valid?"	skill-comply, skill-stocktake, context-sync

Full Changelog: v1.2.0...v1.3.0

AKC v1.2.0

What's Changed

Added Customization section — reference implementations are starting points, not prescriptions. AKC defines the cycle, not the implementation.

Full Changelog: v1.1.0...v1.2.0

AKC v1.1.0

What's Changed

Added Design Principle #5: Evaluation scales with model capability

Small models benefit from rubric-based scoring; reasoning models (Opus-class) evaluate with full context and qualitative judgment. AKC does not prescribe one approach — it matches evaluation depth to the model's reasoning capacity.

Full Changelog: v1.0.0...v1.1.0