Anthropomorphic Agent Engine · Your Soulmate

An anthropomorphic psychology engine based on SPL Pure Core V8.0. Its local core is deterministic — no LLM, no RNG, no outbound calls, offline — and bit-exactly replayable once a virtual clock is injected, so every state change is traceable and reproducible.

> Scope of the determinism guarantee. The reproducibility guarantee covers the local core only. The optional LLM adapters (OpenAI / Claude / Chain) are OFF by default, opt-in only, networked and non-deterministic: they render text from a state snapshot and are never permitted to change state. Any "fully reproducible" statement applies to the core with the adapters disabled, which is the default configuration.

Declared Capability Scope

Everything this skill can reach is listed below. Each entry was checked against the shipped code.

Explicitly not used anywhere in this package: subprocess, os.system, os.popen, eval, exec, and dynamic loading of remote code.

Privacy, Data Flow & Consent (read before use)

This skill is not passive or local-only. Before integrating, understand what it does with data:

Consent & retention rules
- Full transcript retention is disabled by default — stored request rows keep only masked, 120-char summaries.
- Exact age is never persisted; only the boolean outcome of the Step 0 gate and the coarse age band from the consent form.
- Data is stored locally under the server directory (logs/, data/); operators are responsible for retention limits (SPL_MINOR_RETENTION_DAYS), access control, and deletion. Users/guardians can pull or erase a session via /api/export and /api/delete (guardian token required).
- If you integrate the LLM adapters, you must disclose to users that input-derived prompts leave the device, and obtain explicit consent before enabling.
- This engine is not a medical device and does not provide crisis intervention guarantees — see "Safety & Content Policy".

Age Confirmation Gate — Step 0 (read before use)

The synced upstream engine is written for an adult audience, so 18+ confirmation is enforced inside the code, not just in the docs. It runs before every other gate.

Where it lives. Both cores (SPLPureCoreV7_3 in scripts/, SPLMinorPureCore in assets/minor-protection/) carry the same gate: ask_age_gate(), confirm_age(age=…, answer=…), age_gate_required(), age_gate_state(), require_adult(feature), adult_bridge(anchor_phrases), and the AgeGateBlocked exception.

Decision table (confirm_age returns the gate state; nothing is decided by a bare truthy value):

Nobody is dead-ended: a minor or a declining answer is a redirect to the protection core, not a refusal of service — crisis handling stays available in every state.

Ordering. Crisis red-line detection is evaluated before the gate — a crisis message is never delayed by an unanswered age question. After the gate comes the minor-protection service-consent flow (/api/consent), which is a separate, additional requirement, not a replacement: age_group="18+" also answers the gate, while 0-13 / 14-17 force safe_mode + minor_mode.

HTTP surface. POST /api/age_gate {session_id, age|answer} records the confirmation and returns {age_gate, safe_mode, reply}. /api/chat returns age_gate_required: true plus the question and its three options while the gate is open, so an API client can render Step 0 before anything else; the crisis branch returns the gate state as well. A restored session whose age_group is already persisted re-applies the gate silently instead of re-asking after a restart. Non-string answer values are coerced to "__invalid__" before reaching the engine.

The bundled browser UI does not call /api/age_gate directly: its consent form collects an age band, and /api/consent maps 18+ → adult_confirmed, 0-13 / 14-17 → safe_mode. Chat stays locked until that panel is completed, so the gate is answered before the first message on that path.

Downstream projection. adult_bridge() is the only path that may emit persona anchors into image-prompt or other downstream workflows; it refuses to run for a non-adult session, and identity_anchors never accept minor-coding strings (see "Safety & Content Policy").

Two deliberate design constraints.
- Gate state is exposed through age_gate_state() and the audit log, not as a new snapshot() key. The normative key sets in the upstream contract (docs/agent-standard.md §5.1: [full-range] 23 / [minor] 15) and the snapshot_sha256 baselines in tests/conformance_vectors.json stay valid. Both files live in the GitHub repository — this package ships no docs/ or tests/ directory.
- Breaking change for callers of the full-range core. AGE_GATE_STRICT=True, so process_vector() / process_event() / expect() / induce_dissonance() raise AgeGateBlocked until confirmed. Existing code and replay harnesses must add one call after construction:
python core = spl.SPLPureCoreV7_3() core.confirm_age(answer="adult") # required in v2.6.0 before event entry
The protection core is not strict: it accepts adult_confirmed and safe_mode, and blocks only unverified.

Safety & Content Policy (read before use)

• No minors in identity anchors. Personas defined by this engine and consumed downstream must declare themselves as adult (18+). Strings like "16yo girl", "teen girl", "未成年少女", "loli" are forbidden in identity_anchors, persisted state, and any output projected into image-prompt workflows.
• Hard floor on anchor imagery. The minor-protection engine (assets/minor-protection/) is the only authoritative minor-mode source; it removes trauma nodes, clamps negative emotions at 0.75, and raises an L3 crisis signal on risk_level=HIGH. Do not bypass it.
• Default safe output. When emitting any persona state for downstream image-prompt use (ai-draw-cue-word-project or similar), the engine MUST inject an Adult (18+) marker into anchor phrases unless the caller explicitly overrides with a documented, audited reason.
• Cross-skill integration contract. When this engine is integrated with ai-draw-cue-word-project, every persona snapshot forwarded as a prompt anchor must satisfy: (i) character is declared adult (18+); (ii) no minor-coding visual cues; (iii) the receiving skill's safety checklist must report 0 P0 violations before any image is generated.
• User responsibility. You are solely responsible for ensuring downstream usage complies with applicable laws and platform policies in your jurisdiction.

v2.6.0 engine-sync release (2026-10-10)

Synced the engine in this package with the current upstream Anthropomorphic-Agent-Engine repository while keeping the age gate that the package has enforced since 2.5.0. name / slug stay ASCII (ai-soulmate); the display title is carried in displayName.

• Engine synced from the repo (SPL Pure Core V8.0 line). New in the core: token metering (TokenUsage / TokenStats, record_llm_usage(), log_llm_call()), AuditLogger.snapshot_delta() for before/after diffs, _mood_event_dose mood dosing and _defense_allocation defense-mechanism budgeting, and the rewritten LLM adapter stack (OpenAI / Claude / Chain — still OFF by default, still state-read-only).
• Minor-protection layer synced. SQLite session store with per-stream SHA-256 hash chain (_chain_push / _write_chain_row, /api/audit/verify), optional Fernet field encryption (SPL_MINOR_STORE_KEY), retention cleanup (cleanup_expired, SPL_MINOR_RETENTION_DAYS), reply-side output redline (gate_output), AI-identity disclosure + reality-reminder cadence, exit_service, guardian token endpoints and the under-14 verifiable guardian-consent flow (/api/guardian/challenge + /api/guardian/verify).
• Age gate re-ported onto the new engine (see "Age Confirmation Gate — Step 0"): AgeGateBlocked, gate fields and methods in both cores, gate checks in process_vector / expect / induce_dissonance, POST /api/age_gate, the L0a branch in handle_chat, and /api/consent mapping age bands onto gate states.
• Deliberately not a snapshot change. No snapshot() key was added, so the normative key sets ([full-range] 23 / [minor] 15) in the upstream contract docs/agent-standard.md §5.1 and the snapshot_sha256 baselines in tests/conformance_vectors.json remain valid; gate state is exposed via age_gate_state() and the audit log.
• Restored the privacy/network switches the upstream copy had lost. SPL_AUDIT_LOG=0 again disables core JSONL writes (it had become a dead declaration); SPL_LOG=0 again disables request rows; SPL_LOG_DIR is honoured again; the HTTP server binds 127.0.0.1 by default again instead of 0.0.0.0 (SPL_BIND / --bind is the explicit override, and the chosen address is printed at startup); request-log text is masked and truncated to 120 chars while complaint text stays whole but masked. Six new env vars are now declared in the frontmatter so declared ↔ actual behaviour matches.
• LICENSE updated to the repo's v1.2 (2026-10-08), which adds Art. 1.3 "Evaluation Purpose Exception"; assets/minor-protection/COMPLIANCE.md synced. The package README.md was kept (the repo README documents paths that do not exist in this package).
• Breaking change: callers of the full-range core must call confirm_age() once after construction before feeding events.
• Version: 2.5.2 → 2.6.0.

Verification evidence (2026-10-10, Python 3.13 on Windows).
- Age-gate unit suite on both patched cores: 20 assertions, 0 failures — covering crisis-before-gate, garbage/boolean ages, explicit decline → safe_mode, and "numeric age never persisted".
- Snapshot equivalence: 8/8 conformance vectors produce byte-identical snapshots against the unpatched repo engine for both the [full-range] and [minor] cores (0 differences), and audit streams match modulo ts/seq.
- Repo conformance suite run against verbatim copies of the shipped cores, with only a minimal bootstrap injected into __post_init__ (silence audit writes + confirm_age(answer="adult")): 216/216 assertions passed, including S1–S4 determinism / replay 25/25.
- HTTP suite against a temp copy of the server: 16/16 — gate-then-consent ordering, /api/age_gate invalid-answer coercion, restart re-application, crisis branch, guardian token gating on /api/state, export/delete, complaint flow.
- Switch suite: audit default-on and SPL_AUDIT_LOG=0 silent (both cores); SPL_MINOR_DATA_DIR relocates the DB outside the package; SPL_LOG=0 → 0 request rows; SPL_LOG=1 → masked rows carrying the 120-char truncation marker; default bind confirmed loopback-only (connection to the host LAN address refused).

Out of scope for this release: anthropomorphic-agent-engine-workbuddy.zip and the GitHub repository itself were not rewritten. A pre-sync backup of the folder and the old zip is in openclaw-skills/_backup-2026-10-09-260/.

v2.6.1 documentation release (2026-10-18)

• Listing internationalized for ClawHub. displayName, the H1 title and the [full-range] / [minor] core labels are now English; the remaining Chinese strings are deliberate forbidden-token literals in the safety policy ("未成年少女") and stay verbatim so the guard keeps matching them.
• No behavioural change: documentation and labels only. Engine, age gate, conformance contract and version 2.6.0 verification evidence all unchanged.
• Version: 2.6.0 → 2.6.1.

v2.5.2 manifest-fix release (2026-10-02)

• Moved the declared capability scope out of the YAML frontmatter. The top-level permissions: key introduced in 2.5.1 was rejected by the ClawHub manifest parser (manifest_parse_error / _ManifestSchemaError), which dropped scan coverage from 100% to 93.8% and raised a HIGH analysis-evasion finding — that single finding is what pushed the aggregate verdict from clean back to suspicious. The scope now lives in the body as ## Declared Capability Scope: identical content, no manifest risk.
• Resolved an internal contradiction that the 2.5.1 scan flagged: the capability section still described the extension modules as "pluggable / independent replaceable modules" while the Files section called them reference-only. The section now states explicitly which module is wired and which are not.
• Version: 2.5.1 → 2.5.2.

v2.5.1 security-response release (2026-10-02)

Response to the ClawHub security scan on 2.4.0 (decision: fail, security.status: suspicious).

• Determinism claim split into two explicit scopes. The unconditional phrase "zero probabilistic black boxes" was removed. The local core is now described as deterministic (no LLM, no RNG, offline, bit-exact replay after clock injection), while the optional OpenAI / Claude / Chain adapters are described as OFF by default, opt-in only, networked and non-deterministic, and are explicitly excluded from the guarantee. This removes the contradiction the scanner flagged — a "fully reproducible" claim coexisting with outbound, non-deterministic adapters.
• Added a declared capability scope (permissions:) covering filesystem read/write, the five environment variables actually read (SPL_AUDIT_LOG, SPL_LOG, SPL_LOG_DIR, SPL_BIND, SPL_MINOR_CHAT_PORT), outbound hosts, the loopback inbound port and the execution model — plus an explicit list of what is not used (subprocess, os.system / os.popen, eval / exec, remote code loading). Each entry was verified against the shipped code.
• SPL_MINOR_CHAT_PORT declared — the server reads it, but it was previously absent from the frontmatter env list.
• No behavioural change: documentation and declarations only.
• Version: 2.5.0 → 2.5.1.

v2.5.0 documentation-accuracy release (2026-10-02)

• assets/feature/ renamed to assets/feature-guide/ — every reference updated in SKILL.md and README.md.
• Extension modules re-positioned as reference-only. The Identity / Goal / Value / Bias / World / language-style modules are design references, not runtime components: they emit a tension_base key that is outside the interoceptive-vector contract, so their vectors never enter the core (mirrors docs/agent-standard.md §4.5 upstream). The previous wording — "modular modeling", "pluggable modules" — overstated their status.
• All three usage examples are now runnable. The engine file is hyphenated (SPL-anthropic-engine.py) and must be loaded via importlib; the class is SPLPureCoreV7_3 (the V8.0 behaviour set), not SPLPureCoreV8_0; render_style() is a method of LanguageStyleEngine, not a module-level function; add_identity() takes strength as a float, not a dict.
• Description rewritten to name the actual deterministic subsystems and to disclose that the extension modules are not wired into the core.
• Version: 2.4.0 → 2.5.0.

v2.4.0 security-audit patch changelog (2026-09-01)

ClawHub security-audit response (25 findings verified). Fixes:

• HTTP server now binds 127.0.0.1 by default (SPL_BIND env to override) — previously 0.0.0.0 exposed the chat API + browser UI to the network.
• Request logs now masked — user input and replies truncated to 120 chars (_mask); SPL_LOG=0 fully disables request logging.
• guardian_callback documented as opt-in — default None; consent + field-minimization requirements stated in code and docs.
• Added "Privacy, Data Flow & Consent" section — discloses audit logs, request logs, HTTP server, LLM adapters, and callback data flows with defaults and opt-out env vars.
• LLM adapters (OpenAI/Claude/Chain) declared as opt-in — never enabled by default; user consent required before any outbound call.
• Frontmatter now declares optional env vars (SPL_LOG, SPL_LOG_DIR, SPL_BIND, SPL_AUDIT_LOG) so ClawHub security analysis sees declared ↔ actual behavior match.
• Version: 2.3.0 → 2.4.0.

v2.3.0 safety-patch changelog (2026-09-01)

• Added explicit "Safety & Content Policy" section with cross-skill integration contract.
• Reinforced minor_mode=True default and audit-log invariant.
• Removed residual "16yo girl" string from references/PersonaPersistence.md (replaced with "adult woman (18+)").
• Old SKILL.md.bak-20260829 soft-deleted to .workbuddy/.trash/2026-09-01/.

Trigger Scenarios

Automatically trigger when the user asks about the following content:
- Anthropomorphic agent architecture, personality implementation solutions
- Emotion/cognition/motivation modeling methods
- Long-term behavioral consistency design for agents
- Deterministic, auditable AI psychological simulation (local core — no LLM, no RNG, offline)
- Novel/game character behavior deduction

Core Capabilities

🧠 SPL Pure Core V8.0 Architecture

• 8-dimensional emotional fluid model: Joy/Anger/Fear/Trust/Alienation/Tension/Guilt/Shame, continuous state dynamic evolution
• Trauma and memory system: Trauma nodes, memory reconsolidation, Ebbinghaus forgetting curve, repression rebound mechanism
• Trust and relationship model: Trust capacity erosion under long-term cold treatment, dynamic relationship depth calculation
• Psychological metabolic system: Excitation arousal, dynamic viscosity, psychological time, energy fatigue metabolism, sleep dream processing
• V8.0 extended features: Slow variable emotional layer, independent shame dimension, dynamic self-esteem, expectation system (hope/anxiety/disappointment), cognitive dissonance, psychological defense mechanisms
• Added in the v2.6 sync: TokenUsage / TokenStats metering with record_llm_usage() and log_llm_call() (adapters only — they never touch state), AuditLogger.snapshot_delta() diffs between consecutive snapshots, mood-event dosing (_mood_event_dose) and defense-mechanism allocation (_defense_allocation)

🧩 Extension Modules — reference only, NOT wired into the core

The modules below live in assets/feature-guide/. They are design references, not runtime components: they emit a tension_base key that is outside the interoceptive-vector contract, so their vectors never enter the core. The engine's actual extensibility comes from the narrative mapping layer and the deterministic parameter set, not from these files.

• Narrative mapping layer — wired (scripts/SPL-anthropic-engine.py): customizable personalities (optimistic / paranoid / cynical), converts external events into interoceptive vectors
• Identity engine — reference only (assets/feature-guide/Identity module.py): multi-identity model; the description below is the module's own design intent, not behaviour wired into the core
• Goal / value / bias / world — reference only (assets/feature-guide/*.py): design references for drives, valuations, cognitive biases and world-model priors; not connected to the core

💬 Language Style Rendering (Added in v2.1, enhanced in v2.2)

• Language personality module: Expression gears (restrained/sharp/evasive/intimate/frank) + silence strategy
• Style profile: Customizable dimensions including sentence length, formality, sarcasm tendency, era sense, etc.
• Dynamic multi-dimensional tone: Added in v2.2, emotional state drives tone parameter changes in real time
• Deterministic line generation: Added in v2.2, produces reproducible style instructions given a state
• State-to-line instruction: Translates SPL Core internal states into director instructions usable by LLMs

🛡️ Minor Protection Engine (Added in v2.2, expanded in v2.6)

• Weakened core SPLMinorPureCore: Removes trauma nodes, repression-rebound/implicit pressure avalanche and other burst mechanisms
• Emotional clamping: All negative emotion dimensions have upper limits (0.75), attachment capped at (0.8)
• Layered protection: L0 crisis red line (evaluated first, before any gate) → L0a Step 0 age gate → L0b service consent + guardian verification → L1 input gate (red line lexicon hard interrupt) → L2 engine weakening → L3 crisis signal (risk_level HIGH → caring message + guardian_notified) → output redline (gate_output) on the reply side
• Session hygiene: AI-identity disclosure and reality reminder cadences (ai_disclosure_required / reality_reminder_due), rest/overuse hints, exit_service, and retention cleanup (cleanup_expired, SPL_MINOR_RETENTION_DAYS)
• Deterministic audit log: All state changes written to local JSONL, traceable, reproducible
• Persistence: SQLite session store (data/minor_protection.db) with an append-only SHA-256 hash chain per record stream, verifiable via /api/audit/verify; optional field encryption with SPL_MINOR_STORE_KEY
• Guardian controls: overview / block / register / export / delete behind a per-session guardian token, plus a verifiable guardian-consent flow (/api/guardian/challenge + /api/guardian/verify) for under-14 users
• HTTP service layer: Zero dependency http.server, default port 8788, loopback-only bind by default, session isolation, optional HTTPS

Usage

Run Core Engine Directly

import importlib.util
# The engine filename is hyphenated by design, so load it via importlib.
spec = importlib.util.spec_from_file_location("spl_core", "scripts/SPL-anthropic-engine.py")
spl = importlib.util.module_from_spec(spec); spec.loader.exec_module(spl)

core = spl.SPLPureCoreV7_3()   # class name kept for compatibility; this IS the V8.0 behaviour set
# Step 0 — required since v2.6.0: the strict core raises AgeGateBlocked on event entry until confirmed.
core.confirm_age(answer="adult")   # or age=25; a minor/declining answer puts the core in safe mode
# Input event vector: belonging 0.5, threat -0.1, time step 1.0
core.process_vector({"belonging": 0.5, "threat": -0.1}, 1.0)
# Get current full state snapshot
print(core.snapshot())
# Gate state lives outside the snapshot (normative key set is frozen):
print(core.age_gate_state())

Replay / CI harnesses

Any existing harness driving the full-range core must add the confirm_age() call before it feeds events, otherwise the first process_vector() / process_event() / expect() raises AgeGateBlocked. Inject a virtual clock (set_clock / advance_clock) as before — bit-exact replay is unchanged by the gate, and the upstream tests/conformance_vectors.json baselines still match the shipped cores key-for-key.

Custom Personality Configuration

import importlib.util
# Reference-only module — NOT wired into the core (upstream: docs/agent-standard.md §4.5)
spec = importlib.util.spec_from_file_location("identity", "assets/feature-guide/Identity module.py")
m = importlib.util.module_from_spec(spec); spec.loader.exec_module(m)

identity = m.IdentityEngine()
identity.add_identity("Poet", strength=0.9)   # strength is a float, not a dict

Language Style Rendering

# Load language style module directly
import importlib.util
spec = importlib.util.spec_from_file_location("lang_style", "assets/feature-guide/language style.py")
lang = importlib.util.module_from_spec(spec); spec.loader.exec_module(lang)
# Generate style instructions based on SPL Core snapshot
# render_style() is a method of LanguageStyleEngine, not a module-level function
style_engine = lang.LanguageStyleEngine()
style = style_engine.render_style(core.snapshot())
print(style.prompt_injection)

Minor Protection Engine

# Run weakened core engine
# python "assets/minor-protection/SPL-anthropic-minor-engine.py"
#
# Or start minor protection HTTP service (port 8788)
# python "assets/minor-protection/SPL-anthropic-minor-server.py"

v2.0 Upgrade Capabilities (P0-P2)

💾 Personality State Persistence (P0)

Cross-session continuous personality requires state persistence. Follow references/PersonaPersistence.md: State Schema (four blocks: cognition/emotion/motivation/social), atomic writes, schema version migration, determinism guarantee (temperature=0 core decision + seeded randomness).

🎭 Emotion-Behavior Mapping (P0)

Projects internal emotional states into observable behaviors (body language/microexpressions/gaze/line style). Refer to the six-state mapping table in references/EmotionBehaviorMap.md; isomorphic with the posture mapping in Section D of the AI Drawing Composition Template — personality states can directly drive character image generation.

⚖️ Motive Conflict Engine (P2)

Deterministic adjudication during multi-motive competition: Safety constraints → temperament alignment → weighted utility → recency/persistence → user override. Full audit trail for the entire decision process (references/MotiveConflictRules.md), auditable, no probability.

💬 Conversational Adaptation (P1)

Continuous personality can be directly connected to conversational image/text models (GPT-4o / Gemini): Provide personality state snapshot in the first round, subsequent single-point corrections without changing core anchors.

🔌 Soulmate Integration (P1)

The engine can be used as the inference kernel for the your-soulmate extension: The extension handles UI/interaction, the engine handles state evolution, state files are synchronized bidirectionally (see persistence contract).

v2.2 Upgrade Highlights

• SPL Core synchronized with latest GitHub version: Added deterministic audit log (AuditLogger, local JSONL, failures do not block engine)
• language style.py significantly enhanced: Added dynamic multi-dimensional tone rendering + deterministic line generation (+234 lines)
• Added minor protection engine (assets/minor-protection/): Weakened core + HTTP service layer, three-layer protection architecture
• Deleted obsolete files: language-style-demo.py, spl-chat-server.py (replaced by minor-server)
• Added LICENSE: Dual-track licensing (free for personal research / commercial authorization required for government/enterprise)
• README synchronized with latest repository version

v2.1 Upgrade Highlights

• SPL Core synchronized with latest GitHub version: Completed fields such as rationalization_load
• Added language style module (assets/feature-guide/language style.py): Discrete gear expression personality + style profile + state-to-line rendering
  > Renamed from assets/feature/ — these modules are reference guidance only, not wired into the core (they emit a tension_base key outside the vector contract).

付费接入（支付宝 AI 按量付费 / 402）

本技能提供可直接运行的付费入口 payment/service.py，按次收费：

• 未携带 Payment-Proof 时返回 HTTP 402 与 Payment-Needed 头；账单为 Base64URL 编码，包含 out_trade_no、amount、currency=CNY、resource_id、pay_before、seller_signature（RSA2）。
• 智能体侧按 probe → pay → complete → ack 推进；付款后携带 Payment-Proof 重试同一请求。
• 服务端调用 alipay.aipay.agent.payment.verify 验付，校验 active、金额、订单号与资源标识。
• 资源生成后调用 alipay.aipay.agent.fulfillment.confirm 履约确认，确认成功才返回内容。
• 订单持久化到 SQLite（out_trade_no 唯一），原子进入 PENDING_CONFIRM；同一凭证重复请求只返回已生成结果，不重复发放。

依赖：alipay-sdk-python、pycryptodome。

pip install -r payment/requirements.txt

export A2M_APP_ID=... A2M_APP_PRIVATE_KEY=... A2M_ALIPAY_PUBLIC_KEY=...        A2M_SELLER_ID=... A2M_DB_PATH=./data/orders.db
python payment/service.py

沙箱 service_id 固定为 api_mock_service_id；生产环境替换为真实 serviceId。

Files

• references/PersonaPersistence.md (P0 state persistence contract)
• references/EmotionBehaviorMap.md (P0 emotion-behavior mapping table)
• references/MotiveConflictRules.md (P2 motive conflict adjudication rules)
• scripts/SPL-anthropic-engine.py (Core engine + NarrativeMapper + AuditLogger + age gate + LLM adapters)
• assets/feature-guide/ (Identity / goals / values / biases / world / language-style modules — reference only, NOT wired into the core; see docs/agent-standard.md §4.5 upstream)
• assets/minor-protection/SPL-anthropic-minor-engine.py (protected core + age gate)
• assets/minor-protection/SPL-anthropic-minor-server.py (loopback HTTP service, consent/guardian/crisis flows)
• assets/minor-protection/COMPLIANCE.md (mapping to the PRC anthropomorphic-interaction service measures)
• assets/sujin-demo.md (worked persona demo — documentation, not runtime code)
• LICENSE (Dual-track licensing agreement, v1.2)

At runtime the server additionally creates data/minor_protection.db and logs/*.jsonl under its own directory unless SPL_MINOR_DATA_DIR / SPL_LOG_DIR point elsewhere. Neither path is part of the shipped package (.clawhubignore excludes them).

Notes

• Pure Python standard library implementation, no additional dependencies required, Python ≥ 3.8. The single optional package is cryptography, used only to encrypt stored fields when SPL_MINOR_STORE_KEY is set; without it the server falls back to masking (no crash, no outbound call).
• Deterministic local core: the same inputs produce the same outputs — no random numbers, no LLM calls, no outbound network. Replay requires an injected virtual clock.
• The age gate is part of the state machine, not of the numerics: confirming adulthood changes no state variable, so gated and ungated runs of the same event sequence produce the same snapshots.
• The optional remote LLM adapters are non-deterministic and OFF by default; they are outside the guarantee above and never participate in state evolution.
• Licensing: Dual-track model — free for personal non-commercial research, government/enterprise commercial use requires written authorization, see LICENSE for details