Rename to affect-kernel, ground every constant, and evaluate the scorer

.github/workflows/ci.yml

@@ -48,13 +48,15 @@ jobs:
           cache-dependency-path: python/requirements-dev.lock
       - run: python -m pip install --require-hashes -r python/requirements-dev.lock
       - run: python -m pip install --no-deps --no-build-isolation -e ./python
-      - run: python -m pytest python/tests scripts/tests --cov=anjo_core --cov-branch --cov-report=term-missing
-      - run: python -m ruff check python scripts examples
-      - run: python -m ruff format --check python scripts examples
+      - run: python -m pytest python/tests scripts/tests --cov=affect_kernel --cov-branch --cov-report=term-missing
+      - run: python -m ruff check python scripts examples bench
+      - run: python -m ruff format --check python scripts examples bench
       - run: python -m mypy --config-file python/pyproject.toml python/src
       # Examples are executable documentation and assert their own invariants.
       - run: python examples/python-headless/main.py
       - run: python examples/game-npc/main.py
+      # The benchmark report is generated; fail if the committed numbers drifted.
+      - run: python bench/run.py --check
 
   python-package:
     runs-on: ubuntu-latest

.github/workflows/release.yml

@@ -41,7 +41,7 @@ jobs:
           tag="${GITHUB_REF_NAME#v}"
           py=$(python -c "import tomllib,pathlib;print(tomllib.loads(pathlib.Path('python/pyproject.toml').read_text())['project']['version'])")
           ts=$(node -p "require('./typescript/package.json').version")
-          init=$(python -c "import re,pathlib;print(re.search(r'__version__ = \"([^\"]+)\"', pathlib.Path('python/src/anjo_core/__init__.py').read_text()).group(1))")
+          init=$(python -c "import re,pathlib;print(re.search(r'__version__ = \"([^\"]+)\"', pathlib.Path('python/src/affect_kernel/__init__.py').read_text()).group(1))")
           echo "tag=$tag python=$py typescript=$ts __version__=$init"
           test "$tag" = "$py" && test "$tag" = "$ts" && test "$tag" = "$init"
 

.gitleaks.toml

@@ -1,4 +1,4 @@
-title = "Anjo Core secret-scanning policy"
+title = "Affect Kernel secret-scanning policy"
 
 [extend]
 useDefault = true

CHANGELOG.md

@@ -11,15 +11,78 @@ to a pinned vector is called out here.
 
 ## [Unreleased]
 
+### Added
+
+- `AffectDynamics` and `RetrievalWeights`: the numeric coefficients are now
+  caller-owned data, on the same principle that already made stage names,
+  expectation cues, turn-shape rules, and presence labels replaceable. Inertia
+  terms, the resting-dominance coefficient, the baseline blend, per-emotion
+  carry decay and floor, the recency horizon and floor, the episode bonus, and
+  the mood-congruence threshold and asymmetry can all be changed without
+  forking. Defaults reproduce the pinned fixture exactly, so the 225 shared
+  vectors are unchanged.
+- `docs/foundations.md`: per-constant provenance — literature, production-tuned,
+  or arbitrary-but-bounded — with the departures from the cited work stated,
+  and a list of results that would falsify the current choices.
+- `CITATION.cff`, validated against CFF schema 1.2.0.
+- `docs/threat-model.md`: assets, the one boundary the kernel actually enforces,
+  the ways an adapter can silently undo it, and an explicit list of what the
+  kernel does not defend against.
+- Seeded property and fuzz suites in both runtimes covering clamping over
+  200-turn adversarial walks, determinism, non-mutation of caller state, ranking
+  total order, Unicode handling, and pickle/deepcopy round trips. No new
+  dependency: both use a fixed-seed PRNG so a failure is re-runnable from the
+  seed alone.
+- `bench/`: a seeded, dependency-free retrieval benchmark over five regimes,
+  comparing the scorer against plain similarity and against the additive form
+  used by Generative Agents. `bench/RESULTS.md` is generated and drift-checked
+  in CI, so no document can quote a stale number. It reports results against the
+  current design, including that the additive form wins wherever salience
+  carries signal and that mood congruence is worth +0.012 MRR in a regime built
+  to favour it.
+
+### Fixed
+
+- Retracted a claim in `docs/foundations.md` and in the `recency_weight`
+  docstring that linear-to-a-floor recency was "the least defensible" curve in
+  the module. At a matched 30-day half-life it out-ranks both the exponential
+  and the power-law curve.
+
+### Changed
+
+- **Breaking (pre-release):** `CompanionState` is now `AffectState` and
+  `CompanionEngine` is `AffectEngine`, in both runtimes, along with
+  `createAffectState`, `AffectStateInput`, `ResolvedAffectState`, and
+  `AffectEngineOptions`. The last product-specific noun in the public API is
+  gone: a game NPC no longer instantiates a "companion". The presence vector's
+  `source` field changes from `"companion_state"` to `"affect_state"`, which
+  updates 8 expected values in `shared/golden/kernel_golden.json`. This is a
+  reviewed fixture change under `docs/parity-contract.md`; no version was ever
+  published, so nothing installed is affected.
+- Renamed from `anjo-core` / `@anjo-ai/core` to `affect-kernel` on both
+  registries, and the Python module from `anjo_core` to `affect_kernel`. No
+  version was ever tagged or published under the old name.
+
+### Documentation
+
+- `docs/algorithm.md` now specifies the ambiguous-intent valence amplification
+  (`x1.10` negative, `x1.04` positive above `|v| >= 0.20`), which was
+  implemented but undocumented.
+
 ## [0.1.0]
 
 First public release: the deterministic kernel extracted from
 [Anjo](https://anjo.love) and generalized beyond conversation.
 
+The repository was briefly public as `anjo-core` before this release and was
+renamed to `affect-kernel` to name the library by what it does rather than by
+the application it came from. No version was ever tagged or published under the
+old name, so no installed artifact is affected.
+
 ### Added
 
-- Behaviorally aligned Python (`anjo-core`) and TypeScript (`@anjo-ai/core`)
-  kernels with no runtime dependencies.
+- Behaviorally aligned Python and TypeScript kernels with no runtime
+  dependencies, both published as `affect-kernel`.
 - OCC-inspired appraisal, PAD mood dynamics, and Big Five N/E-conditioned affect
   inertia.
 - Bounded memory relevance, recency, salience, and mood-congruence scoring.
@@ -49,7 +112,7 @@ First public release: the deterministic kernel extracted from
 
 - `FrozenMapping` is picklable. The default `dict` pickle protocol restores
   items by mutating a fresh instance, which the class refuses, so every state
-  object holding one — `CompanionState` with a non-empty `occ_carry`,
+  object holding one — `AffectState` with a non-empty `occ_carry`,
   `TurnShapePolicy`, `PromptPolicy` — raised `TypeError` on `pickle.dumps`.
   That broke any `StateStore` serializing with `pickle` and any use across a
   process boundary. Restored values remain immutable.
@@ -59,5 +122,5 @@ First public release: the deterministic kernel extracted from
   construction step, so a lower-case declaration silently failed to suppress
   there while working in Python.
 
-[Unreleased]: https://github.com/kevindechang/anjo-core/compare/v0.1.0...HEAD
-[0.1.0]: https://github.com/kevindechang/anjo-core/releases/tag/v0.1.0
+[Unreleased]: https://github.com/kevindechang/affect-kernel/compare/v0.1.0...HEAD
+[0.1.0]: https://github.com/kevindechang/affect-kernel/releases/tag/v0.1.0

CITATION.cff

@@ -0,0 +1,107 @@
+cff-version: 1.2.0
+message: "If you use this software, please cite it as below."
+title: "affect-kernel: a deterministic affect-state kernel for long-lived AI characters"
+abstract: >-
+  A dependency-free Python and TypeScript library that turns appraised events
+  into bounded PAD mood, OCC-flavored emotion carry, mood-aware memory ranking,
+  response controls, and presence signals. The deterministic surfaces are pinned
+  by a shared cross-runtime fixture so that both runtimes reproduce the same
+  numbers. Constant provenance — published work, production tuning, or bounded
+  arbitrary choice — is documented per constant in docs/foundations.md.
+type: software
+authors:
+  - family-names: Chang
+    given-names: Chia Wei
+    alias: kevindechang
+repository-code: "https://github.com/kevindechang/affect-kernel"
+url: "https://github.com/kevindechang/affect-kernel"
+license: Apache-2.0
+version: 0.1.0
+keywords:
+  - affective computing
+  - appraisal theory
+  - PAD model
+  - OCC model
+  - agent memory
+  - character agents
+  - deterministic simulation
+references:
+  - type: book
+    title: "The Cognitive Structure of Emotions"
+    authors:
+      - family-names: Ortony
+        given-names: Andrew
+      - family-names: Clore
+        given-names: Gerald L.
+      - family-names: Collins
+        given-names: Allan
+    publisher:
+      name: Cambridge University Press
+    year: 1988
+    doi: 10.1017/CBO9780511571299
+  - type: article
+    title: >-
+      Pleasure-arousal-dominance: A general framework for describing and
+      measuring individual differences in temperament
+    authors:
+      - family-names: Mehrabian
+        given-names: Albert
+    journal: Current Psychology
+    volume: 14
+    issue: 4
+    start: 261
+    end: 292
+    year: 1996
+    doi: 10.1007/BF02686918
+  - type: conference-paper
+    title: "ALMA: A Layered Model of Affect"
+    authors:
+      - family-names: Gebhard
+        given-names: Patrick
+    collection-title: >-
+      Proceedings of the Fourth International Joint Conference on Autonomous
+      Agents and Multiagent Systems (AAMAS '05)
+    start: 29
+    end: 36
+    year: 2005
+    doi: 10.1145/1082473.1082478
+  - type: article
+    title: >-
+      Feelings change: Accounting for individual differences in the temporal
+      dynamics of affect
+    authors:
+      - family-names: Kuppens
+        given-names: Peter
+      - family-names: Oravecz
+        given-names: Zita
+      - family-names: Tuerlinckx
+        given-names: Francis
+    journal: Journal of Personality and Social Psychology
+    volume: 99
+    issue: 6
+    start: 1042
+    end: 1060
+    year: 2010
+    doi: 10.1037/a0020962
+  - type: conference-paper
+    title: "Generative Agents: Interactive Simulacra of Human Behavior"
+    authors:
+      - family-names: Park
+        given-names: Joon Sung
+      - family-names: O'Brien
+        given-names: Joseph C.
+      - family-names: Cai
+        given-names: Carrie J.
+      - family-names: Morris
+        given-names: Meredith Ringel
+      - family-names: Liang
+        given-names: Percy
+      - family-names: Bernstein
+        given-names: Michael S.
+    collection-title: >-
+      Proceedings of the 36th Annual ACM Symposium on User Interface Software
+      and Technology (UIST '23)
+    start: 1
+    end: 22
+    year: 2023
+    doi: 10.1145/3586183.3606763

CONTRIBUTING.md

@@ -1,4 +1,4 @@
-# Contributing to Anjo Core
+# Contributing to Affect Kernel
 
 Thank you for helping make long-lived character systems more inspectable and
 portable.

GOVERNANCE.md

@@ -1,6 +1,6 @@
 # Governance
 
-Anjo Core currently uses a maintainer-led model.
+Affect Kernel currently uses a maintainer-led model.
 
 - Maintainers set scope, merge changes, cut releases, and resolve security issues.
 - Significant public API or parity-contract changes should begin as an issue or
@@ -14,3 +14,23 @@ As sustained contributors emerge, maintainership can be granted based on review
 quality, reliability, and care for the public boundary. Governance changes are
 made through pull requests to this file.
 
+## Bandwidth
+
+Worth knowing before you invest time in a contribution: this is maintained by
+one person, alongside other work. Expect a first response to an issue or pull
+request within about a week, and longer for anything touching the parity
+contract, which needs a reviewed fixture change in both runtimes.
+
+Things that get looked at fastest, in order:
+
+1. A reproducible bug in a deterministic transform, with a failing case.
+2. A benchmark or evaluation result — including one that contradicts something
+   this repository claims. `docs/foundations.md` lists the results that would
+   falsify its own choices, and the linear-recency claim has already been
+   retracted on evidence once.
+3. A new storage or model adapter, or a domain preset.
+4. API surface changes, which need a design discussion first.
+
+If something here is stalled and you need it, say so on the issue rather than
+assuming it was rejected.
+

Makefile

@@ -1,4 +1,4 @@
-.PHONY: setup test lint package check examples
+.PHONY: setup test lint package check examples bench
 
 setup:
 	./scripts/setup.sh
@@ -8,8 +8,8 @@ test:
 	npm test --prefix typescript
 
 lint:
-	python -m ruff check python scripts examples/python-headless
-	python -m ruff format --check python scripts examples/python-headless
+	python -m ruff check python scripts examples/python-headless bench
+	python -m ruff format --check python scripts examples/python-headless bench
 	python -m mypy --config-file python/pyproject.toml python/src
 	npm run typecheck --prefix typescript
 
@@ -23,3 +23,6 @@ check:
 examples:
 	python examples/python-headless/main.py
 	npm run example --prefix typescript
+
+bench:
+	python bench/run.py --write