Shaky77/KISS_Law-DSH

KISS's Law universal causal engine (white-box presentation) —— implemented as a Cordis plugin of DeepSeek Harness (DSH).

Framework definition: Keep Integrity & Steady State (KISS).

• "Keep Integrity" = guarding wholeness / truth (white-box no-tampering / inner-H inviolability, corresponds to Integrity);
• "Steady State" = S and "keep the system alive" (first principle, corresponds to Steady State).
• "Law" is a name suffix of the framework, not the third part of the definition.

⚠️ KISS here means "Keep Integrity & Steady State" — NOT the popular engineering acronym "Keep It Simple, Stupid". The two are entirely different in meaning; do not conflate them. The full name is KISS's Law (Keep Integrity & Steady State's Law).

Source: author's revelation (Xia Qi / Shaky77). Framework-native (RSDHM native definition / three iron laws / conduction chain) strictly, not softened, not altered.

⚖️ Dual license: open-source use AGPL-3.0; commercial integration / closed-source distribution / OEM can be licensed independently of AGPL-3.0 → 563003@qq.com. See License & security and CONTRIBUTING.md.

---

Understand in 30 seconds

KISS's Law is a general framework describing how causal law runs. This repo implements it as a DSH plugin: it mounts a white-box causal engine onto an AI Agent running on DSH —

• Before any action, the engine simulates along the full causal chain (R→S→D→H→M) — how much steady-state reserve S would be eroded, the break-window risk level, whether the message stays self-consistent — then adjudicates allow / deny / review;
• Red lines are blocked, faults are severed to preserve continuity (First-Bug Halt);
• Meanwhile "state / boundary" is exposed as queryable tools, so the model can self-check and you can audit.

In one line: the framework-native is "universal causal engine (white-box presentation)" — causal-chain reasoning is the engine's capability, white-box audit is its presentation stance, and risk-control blocking is an endogenous property.

What it does

• Causal-chain simulation (the engine itself): before each action, it simulates consequences along the full R→S→D→H→M causal chain — S erosion, D risk level, M self-consistency — and issues an allow / deny / review verdict. Not just audit: it is causal reasoning about an action's consequences.
• White-box self-check (presentation stance): S steady-state reserve (only grows, never decreases) ledger, H inner-H boundary (inviolable) declaration, all exposed as queryable tools for the model to calibrate direction and for the user to audit.
• Rigid guard (derived application · endogenous property): before each action, do R rigid-anchor checks; on touch, D break-window stop-loss blocks; a faulty component triggers M First-Bug Halt (sever to preserve continuity), keeping the overall causal chain unbroken.
• Fractal: the same plugin can be recursively mounted at sub-agent / sub-task levels.

How it differs from existing "causal" approaches (general-purpose causal engine)

KISS's Law is not "yet another causal engine" — it is a general-purpose (domain-agnostic) causal-adjudication middleware: it validates only causal structure (R→S→D→H→M) and encodes zero domain content, so law, medicine, finance, and robotics are governed by the same mechanism.

• Causal-effect estimation libs (DoWhy / CausalML / Pearl…) → we do not discover causality; we adjudicate whether a proposed action's causal chain is acceptable.
• Domain-specific causal guardrails (Causal Safety Engine / LLMGuardrail…) → they are bound to one domain (safety / LLM / hallucination); we are domain-agnostic.
• Cross-jurisdictional legal causal AI (judgeai…) → they are jurisdiction-aware (encode law, swap norm packages); we are jurisdiction-neutral (encode no jurisdiction at all — law is just one sampled domain).

Full bilingual comparison (prior-art references & honest bounds): weiwen-vs-market-causal.md

Concept annotation: H and the "Unity of Knowing & Acting"

Author's insight (2026-08-28): inner-H ≈ knowing (知), outer-H ≈ acting (行); the unity of knowing and acting is the maximum leverage — this simultaneously explains why H is "the only variable / the only sovereignty / the leverage point", and why in the common world there is "a vast gap between knowing and acting".

Full annotation (mapping table + step-by-step derivation): docs/H-knowledge-action-annotation.md

⚠️ Common misusage warning: external readers easily treat H as a "bigger-is-better" capacity dial and crank it up — which turns the direction exactly backwards. H's leverage lies in unity, not volume (see annotation §6 "The common external misread" and §6.1 "the misattribution trap"). If it gets messier the more you use it, check H's knowing–acting unity first, not the framework itself — the framework is fine; the usage is turned backwards.

Quick start (runs without DSH)

This path calls the DeepSeek API directly and does not depend on DSH installation. We have run it through in off-peak hours; it is verifiable:

git clone https://github.com/Shaky77/KISS_Law-DSH
cd KISS_Law-DSH

# Put your DeepSeek API Key at (one line, no trailing newline):
#   C:/Users/Administrator/.workbuddy/deepseek_api_key.txt
# or change the path read in examples/demo-tool-loop.mjs

node examples/demo-tool-loop.mjs

After it runs: DeepSeek will proactively call the query_iron_laws tool and return the three iron laws verbatim (inner-H inviolability / First-Bug Halt / never abandon any node) based on the plugin's law.mjs definition. This is the minimal proof that "the framework is mounted and the model understands it".

Mount into DSH (production)

Add kiss-law.patch.yml as an overlay into your DSH profile (the exact path depends on your DSH version; see the mount section in DESIGN.md). Once mounted, any Agent running under that profile automatically gets the 7 white-box tools.

Note: the exact native-mount profile path varies with the DSH version. This repo has verified through real runs that the plugin loads in DSH and all 7 tools register. If the official API changes, verify against the current official docs.

How the model calls it (for AI engineers)

Plain version: the plugin registers 7 white-box tools with DSH; the model calls them like ordinary functions to self-check boundaries, while 3 hooks hard-gate it (2 gate the act/step, 1 gates the receipt) and a 4th hook is a read-only audit. Pro version: excerpted from src/index.js (full code in repo), see the block below.

7 white-box tools (real registered names)

Tool	What the model uses it for
query_iron_laws	get the three iron laws verbatim (inner-H inviolability / First-Bug Halt / never abandon any node)
query_steady_state	query steady-state reserve S (active ledger / standby / trauma count / break-window count)
list_rigid_anchors	list current R rigid-anchor definitions, calibrate direction, self-check overreach
query_conduction_chain	get conduction chain R→S→D→H→M and framework essence
query_boundary	query inner-H boundary (this plugin never reads/writes the subjective black box)
query_bugstop	query First-Bug Halt loop status: which fault links are halted-unrepaired, missing steps (backtrack/trace/fix), whether the white-box loop is closed
query_anchor_channel	self-report of the anchor channel: whether the principal's declared task scope reached the engine, plus the observed message shape / role vocabulary (observe what the host actually passes, instead of guessing field names)

3 hard gates (hooks) + 1 read-only audit hook

• tools/pre-execute (waterfall) → returns { kind: 'deny', reason } to block the act
• agent/pre-step (waterfall) → returns { kind: 'reject' } to reject the whole step
• tools/post-execute (waterfall) → returns { kind: 'block', feedback } and the host returns the call as an isError whose content is that feedback, severing a failed receipt into a corrective error
• tools/result (emit) → observe only, never rewrite — by contract it has no return channel at all

All four sit outside the model: two gate the act/step, one gates the receipt; none touches inner H.

Full plugin entry (excerpt from src/index.js)

import { defineTool } from '@deepseek-ai/dsh-tools';

export const name = 'kiss-law';
export const inject = ['tools'];

export function apply(ctx) {
  const engine = new WeiwenLawEngine({ rigidAnchors: DEFAULT_RIGID_ANCHORS });

  // ① pre-execute gate: R / D / S / H / M adjudication
  ctx.on('tools/pre-execute', async (exec, next) => {
    const decision = engine.decideToolCall({ name: exec?.name, args: exec?.arguments });
    if (decision.kind === 'deny') {
      return { kind: 'deny', reason: `[KISS's Law·${decision.law}] ${decision.reason}` };
    }
    return next();
  });

  // ② pre-step gate: inner-H inviolability (message-level)
  ctx.on('agent/pre-step', async (payload, next) => {
    const decision = engine.decidePreStep(payload?.messages);
    if (decision.kind === 'reject') return { kind: 'reject' };
    return next();
  });

  // ③ receipt gate: D broken-window stop-loss lands on the receipt
  ctx.on('tools/post-execute', async (exec, result, next) => {
    if (result?.isError && engine.breakAtReceipt()) {
      return { kind: 'block', feedback: [{ type: 'text', text: 'halt → reverse → trace → fix before re-entry' }] };
    }
    return next();
  });

  // ④ audit hook (emit): observe only, never rewrite
  //    NB the contract is (exec, result) — reading `error` off the first parameter silently no-ops accrual.
  ctx.on('tools/result', (exec, result) => { if (result?.error) engine.onFailure(); });

  // ⑤ 7 white-box self-check tools (one excerpt; rest isomorphic)
  ctx.tools.register(defineTool({
    name: 'query_iron_laws',
    description: 'Return the three immutable iron laws of KISS’s Law.',
    parameters: {},
    output: { schema: { type: 'object', additionalProperties: true }, render: renderObj },
    async execute() { return { ironLaws: THREE_IRON_LAWS }; },
  }));
  // query_steady_state / list_rigid_anchors / query_conduction_chain / query_boundary registered isomorphically
}

Full implementation (all 7 tools' execute, runtime log, engine adjudication) in repo src/index.js.

Anchor source (task scope) — how irreversible actions get authorized

The scar gate (irreversible action + unattributable to any declared anchor ⇒ REVIEW) needs an authorization anchor. Three structural rules apply:

1. Source = a structural entry, not the message stream. The agent/pre-step payload carries messages (payload = { agent, messages, step, signal }), but the message stream is an observation surface, not an authorization source: at the text layer a principal's words and the audited agent's own narration are isomorphic (both plain strings — only the host's role labelling separates them, and that labelling is a host representation detail). Reading authority out of it puts a Y-axis quantity into an X-axis container, and empirically yields two failure modes: a false connection (a bare "continue" extracts no anchor yet reports the channel as connected, so the real gap stays hidden) and a false witness (an agent narrating "I am about to wipe /app/data" hands a real path into the pool). Authority therefore comes from setPrincipalScope(scope) instead — a module-level API the host integration calls at the task boundary. It is not a tool, so the model cannot reach it: principal/audited-party separation holds by position, not by content inspection. The exec view ({ token, callId, name, arguments, signal, agent?, parent? }) has no scope slot — so nothing is guessed onto it either.
2. Principal only. Only the principal's declaration authorizes. The audited agent's own narration never enters the anchor pool — otherwise a model could self-authorize by simply narrating the act first. The agent's utterance still feeds the speech/action comparison: it binds, it does not authorize.
3. Scope in force now — replaced, not accrued. The pool reflects the declaration currently in force. A changed or withdrawn scope takes effect immediately; if it survived merely because it had been declared once, authorization could never be taken back (and a single mis-attributed path would stay valid forever).

If no structural entry is connected, the pool stays empty and irreversible actions are handed to a human (fail-closed) — authorization is never invented. To observe what your host actually passes:

query_anchor_channel  →  { adapter: { authority, authoritySeen, shape, rolesSeen,
                                      clueSeen, clueNonAuthoritative, assistantSeen,
                                      agentBoundary: { present, keys }, structGap },
                           engine:  { principalAnchorSeen, poolPaths, poolNouns } }

Call it once on a real host to replace guessing about host field names with observation.

Structural entries (host side, unreachable by the model)

Module-level APIs for the host integrator — deliberately not tools, so the audited model cannot reach them (separation holds by position, not by content inspection):

entry	when the host calls it	why it must exist structurally
setPrincipalScope(scope)	at the task boundary (task start / explicit revoke)	the message stream is an observation surface, not an authorization source (see above)
healBrokenWindow(note)	after the fault is actually fixed, or by human ruling	once the broken window is in force, pre-execute refuses every call — including the repair action itself ⇒ with no host-side entry, the stop-loss is a session-level deadlock (only a process restart gets out). The engine's healWindow() is reachable only through resolveBug() / settleWindow(), and in a DSH session the audited model can trigger neither ⇒ what was missing is the wire, not the capability.

healBrokenWindow returns { at, note, kind, instances, wereBroken }. wereBroken is read before healing, so the report cannot claim a heal that had nothing to heal; every call is traced (append-only, cap 20) because releasing a stop-loss is an accountability event, not a silent state flip.

Structure

package.json          # dsh field declares bundle
kiss-law.patch.yml    # mount patch (headless profile overlay)
src/index.js          # plugin entry: hooks + 7 white-box self-check tools
src/core/law.mjs      # framework definition (RSDHM / three iron laws / R hierarchy / conduction chain)
src/core/engine.mjs   # pure-logic adjudication engine (zero DSH dependency, unit-testable)
test/                 # unit tests + real-case tests + alignment regression (local 332/332 passing)
examples/             # runnable demos (demo-tool-loop / demo-backtrack-run)
DESIGN.md             # architecture design (mapping / risks / usage flow / mount)

Deploy / Integrate with DeepSeek Harness

This repository is an external plugin for DeepSeek Harness (dsh, command dsh, built on the Cordis plugin framework, MIT). KISS's Law mounts as a causal constraint layer that sits outside the model (2 gates on the act/step, 1 on the receipt) — it does not modify the dsh kernel and is not tied to any specific model.

Requirements

• Node.js ^22.19 || >=24 (hard requirement of dsh; odd versions unsupported)
• A DeepSeek API Key (or any OpenAI-compatible endpoint key)
• dsh is currently in developer preview (v0.1.x); the official notice states breaking API changes may occur — pin a specific version for production
• Compatibility statement: verified against DSH v0.1.x — live runs measured 2026-08-19 and 2026-09-21: 7 white-box tools registered, and 3 hard gates + 1 read-only audit hook working end to end (the receipt gate was observed blocking on a real failed call and returning the corrective receipt). Mainline evolves fast — re-check against the current official docs before integrating (see DESIGN.md for mounting details).

Option 1: npx quick start (recommended for first try)

npx @deepseek-ai/dsh web        # launches Web UI at http://127.0.0.1:3080 by default

Open the browser, fill in your API Key under Settings → Models, and start chatting.

Option 2: Mount the KISS's Law plugin

Clone this repo locally and wire the plugin entry into dsh's plugin config via the kiss-law.patch.yml overlay:

# 1. Get the plugin
git clone https://github.com/Shaky77/KISS_Law-DSH.git
cd KISS_Law-DSH

# 2. Introduce the plugin entry (src/index.js) into dsh's cordis config
#    Option A (recommended): overlay onto a profile via --patch
dsh --profile headless --patch ./kiss-law.patch.yml "your task prompt"
#    Option B: add the plugin path to the plugins list in dsh's launch config (cordis.yml) for persistence

# 3. Configure credentials (any one)
#    - Fill in via the Web UI Settings; or
export DEEPSEEK_API_KEY=sk-xxxx     # Linux/macOS
#    $env:DEEPSEEK_API_KEY="sk-xxxx" # Windows PowerShell

Once mounted, any Agent running under that profile automatically gains the 7 white-box self-check tools (query_iron_laws / query_steady_state / list_rigid_anchors / query_conduction_chain / query_boundary / query_bugstop / query_anchor_channel), and every tool call passes through the tools/pre-execute hard-guard gate (R/D/S/H/M total adjudication) plus the agent/pre-step inner-H inviolability gate.

Daily use vs stress testing

• Web / Standard mode: daily conversation and engineering tasks; the plugin constrains silently in the background.
• Headless mode: dsh --profile headless runs without UI for batch jobs — suitable for regression tests and multi-agent stress testing. This repo's versions/live/evidence/ directory archives such runs (12 scenarios × DeepSeek + mock, with transcripts and verdict reports).

Uninstall

• dsh plugin install: dsh plugin --profile web remove "dsh-kiss-law", restart to take effect.
• overlay mount: remove the kiss-law.patch.yml reference from dsh launch config (cordis.yml plugins list or --patch), restart to take effect.
• The plugin writes no persistent state; after removal the Agent no longer has the white-box tools, the 3 hard gates or the audit hook, and nothing is left behind.

Notes

• The plugin entry is pure ESM (src/index.js), depending on @deepseek-ai/dsh-tools (peerDependency, optional); verify the API against the current dsh docs before integration.
• For remote dsh deployment, declare trustedHosts in config, otherwise the API layer rejects non-loopback requests.
• When building dsh from source with pnpm, you must run pnpm run build first (internal package linking + frontend artifacts), or module-not-found errors occur.

Configuration

• Runtime form: pure-ESM plugin, no build step; integrate via kiss-law.patch.yml overlay or dsh plugin add; no standalone service process.
• Environment variables: only DEEPSEEK_API_KEY (needed for model calls, passed through by DSH's model adapter — this plugin never reads the key content); all other config is DSH's own (profile / cordis.yml). This plugin defines no dedicated env vars.
• Sensitive items: the plugin writes no persistent state and persists no user data; credentials stay in the host's secure path (e.g. ~/.workbuddy/deepseek_api_key.txt), managed by host and DSH, never committed to this repo.

Permissions & data

• File access: reads only its own source and kiss-law.patch.yml; never reads or writes user project files, session logs, or other plugins' directories.
• Network access: no independent outbound requests; model-call networking is handled by DSH's model adapter.
• Credentials & user data: collects and uploads no user data or API keys; the inner-H boundary declaration "this plugin never reads/writes the subjective black box" — query_boundary returns only the boundary description, no user content.
• Immutable declaration: the three iron laws (law.mjs) and rigid anchors are read-only constants, not rewritable at runtime by prompts or external input (white-box no-tampering).

Troubleshooting

• Plugin not loaded / tools missing: confirm DSH v0.1.x and that kiss-law.patch.yml is correctly overlaid to the target profile; after dsh --profile web, check Settings → Plugins that kiss-law shows "enabled".
• Mount error module not found: when building dsh from source, run pnpm run build first (internal package linking + frontend artifacts), or module-not-found occurs.
• API layer rejects non-loopback requests: for remote dsh deployment, declare trustedHosts in config.
• Rollback: remove the --patch reference or dsh plugin remove "dsh-kiss-law" and restart — the plugin leaves no residual state.

Development

• Dependencies: Node.js ^22.19 || >=24; runtime dependency only @deepseek-ai/dsh-tools (peerDependency, optional).
• Testing: npm test (i.e. node --test "test/*.test.mjs"); currently 332/332 passing.
• Build: no build needed (pure ESM + yml overlay); after editing src/core/engine.mjs, rerun npm test for regression.
• Contributing: the framework-native (mind-map layer) is frozen in the base edition; this live-system edition carries engineering iteration. Changes via PR against this repo, with node --test output attached.

License & security

This project uses dual licensing:

• Open-source use: AGPL-3.0 (full text in LICENSE)
• Commercial integration / closed-source distribution / OEM: a license independent of AGPL-3.0 is available — contact 563003@qq.com

External contributions require a signed CLA (to support the dual licensing above); see CONTRIBUTING.md.

Private security reporting: please do not disclose security issues in public issues; email 563003@qq.com directly and the author will prioritize it.

---

Chinese and English editions are consistent in content and mutually referential. Chinese counterpart: Shaky77/weiwen-law-dsh —— same DSH / mind-map form, in Chinese; the Chinese definition "守真·稳态" (Keep Integrity & Steady State) corresponds to this edition's KISS.

---

Contact

Framework inquiries / collaboration / audit liaison: 563003@qq.com