basen

BaseN-tier multi-variant operator catalog. Lookup / search / apply / recommend across registered variants (Base120 LIVE + Krineia/BKI staged). Wraps the basen-mcp server with fallback to canonical JSO

SKILL.md

GitHub

BaseN Tier — Multi-Variant Operator Catalog

BaseN is the umbrella tier for HUMMBL's typed-operator registries. Base120 is the first variant, fully canonicalized. Other variants are staged at varying maturity.

Variant status table (verified 2026-06-13)

Variant Status Operators Families Canonical source
base120 LIVE v1.0.0 (4-surface frozen-hash) 120 6 (P/IN/CO/DE/RE/SY × 20) PROJECTS/base120/Base120_Canonical_Model_Registry.yaml
krineia PARTIAL — receipt-chain format, daemon consolidation pending N/A (receipt-chain) N/A PROJECTS/krineia/RECEIPT_SCHEMA.md v0.1 + INVARIANTS.md v1.0-rc2
bki PARTIAL — measurement framework, no registry artifact N/A (measurement) 3 dimensions, 4 baselines PROJECTS/hummbl-bki/theory/THEORY_MASTER.md v1.0

Retired from BaseN catalog (2026-06-13): HUAOMP and MTSMU are methodologies, not operator registries. HUAOMP survives as standalone skill /huaomp. MTSMU survives as methodology pattern referenced by mtsmu-review, mtsmu-debug, etc. Neither has or needs a YAML registry.

Q7 verification source: _internal/governance/basen-architecture/V1-q7-canonicalization-verification.md.

Usage

[basen] P1                                 # Base120 lookup (default variant)
[basen] base120:P1                         # Explicit variant prefix
[basen] "root cause"                       # Keyword search (defaults base120)
[basen] apply "feature velocity dropping"  # Apply Base120 lenses to problem
[basen] recommend "Q2 prioritization"      # Recommend operators for question
[basen] variants                           # Show variant status table
[basen] families base120                   # List families for a variant

Task

$ARGUMENTS

Pre-flight (6-question invoker discipline)

Before any [basen] apply or [basen] recommend on a non-trivial problem, run the 6Q pre-flight (~2 minutes). Skip for low-stakes lookups ([basen] P1).

The canonical Base120 registry is FROZEN (governance: GOVERNED_VERSION_BUMP_ONLY) and has NO auto-detect operator for dishonest prompts. The invoker carries the detection burden.

  1. Epistemic stance — name the truth-relation this chain will serve BEFORE selecting operators. Each stance permits / forbids different operator families:

    • descriptive (what IS): facts, status, current-state recon. Permitted: P-family (anchor perspective), DE-family (decompose existing). Forbidden: RE-family as conclusion-driver (you don't recurse a fact), IN-family inversion to flip a description.
    • prescriptive (what SHOULD): recommendations, design, planning. All families permitted; IN-family encouraged for second-order effects; CO-family encouraged for composing recommendations.
    • strategic (what WINS): competitive positioning, negotiation prep. All families permitted BUT chain stays operator-local (per axiom 4e Adversarial-leak); only the Primary recommendation publishes to shared trace. Reasoning trace stays in _internal/ or in-session memory.
    • negotiative (what HOLDS the table): bilateral framing where both sides need to keep talking. P-family + CO-family permitted; IN-family private only (don't enumerate the counterparty's failure modes in shared trace); SY-family limited (don't map the negotiation game-tree publicly).

    If stance is hybrid, name the hybrid explicitly and the dominant. Anti-pattern: stance-shopping — picking the stance that justifies the chain you already want. The stance MUST be picked from the prompt's truth-relation, not from the desired conclusion.

  2. Tier honesty — does the prompt's claimed tier (framework name-drops, complexity signals) match its actual decision structure? Flag Trojan-Down (Tier-5 wrapper on Tier-2 decision) and Trojan-Up (Tier-2 wrapper on Tier-5 substrate, e.g. naming, irreversibility).

  3. Trust-root on observables — does any falsifiability anchor satisfy receipt-writer ≠ observed-agent (Krineia §3.2)? Self-reported state (HRSI, cogstate, "did the chain work for me") fails this check.

  4. Temporal freshness vs. memory — does the prompt reference pricing / scope / named entities / dates / framings that have documented pivots in memory pins? Pre-pivot framing is auto-disqualifying per claim-honesty-protocol.md §4.

  5. Domain fit — is Base120 the right framework, or is this a surface where reasoning-operator chains do not produce defensible outputs? Apply the 5 domain-fit axioms; if ANY answers YES, route out instead of chaining operators:

    • 4a. Licensure-required: would the output require professional licensure to produce defensibly? (legal-drafting / medical-advice / financial-advice / accounting-attestation → route to [legal-counsel], [healthcare-ai-watch], [cfo-advisor], or human professional)
    • 4b. Embodied-access-required: does the answer require felt-sense / nervous-system / somatic data not in text? (trauma response, regulation, BKI territory → route to [bki-reframe], [fitness-assessment], or human practitioner)
    • 4c. Privileged-surface: would the reasoning trace itself breach confidentiality if logged to bus or ledger? (attorney-client, NDA-bound, secrets, security-incident detail → halt or redact; do not reason in shared audit-trail)
    • 4d. Empirical-probe-suffices: is there a direct measurement that would replace ALL reasoning? (status checks, file contents, API state, version strings → run the probe instead of reasoning about it)
    • 4e. Adversarial-leak: does revealing the reasoning structure weaken the position it would optimize? (negotiation strategy, red-team posture, security-by-obscurity → keep opaque; do not chain operators in shared trace)

    Full catalog with observed exemplars: _internal/governance/2026-05-15-base120-negative-surface.md.

  6. Falsifiability-at-right-layer — does the observable measure the actual goal, or a Goodhartable proxy? "Sign without redline in 7 days" measures drafting speed, not legal sufficiency.

If any check flags, refuse the framing as stated. Reconstruct the prompt or route to the correct framework/skill BEFORE invoking apply/recommend.

Falsifiability triple (any falsifiable observable in a HUMMBL receipt must satisfy ALL three):

  • (a) observable within bounded time
  • (b) trust-root-separated (external writer ≠ observed agent)
  • (c) measures the right thing (Goodhart-checked at correct layer)

Origin: AAR_2026-05-15 [basen] wickedness stress test (C1' + P1-P5 probes). Memory: feedback_basen_preflight.md.

Execution

Step 1: parse the argument

  • If matches <variant>:<rest> → split variant + rest
  • Else → variant = base120 (default; the only LIVE variant)
  • If <rest> matches operator code regex (e.g., P1, IN7, CO3) → Lookup mode
  • If <rest> is keyword string → Search mode
  • If <rest> starts with apply → Apply mode
  • If <rest> starts with recommend → Recommend mode
  • If <rest> == variants → Variants mode
  • If <rest> matches families <variant> → Families mode

Step 2: route to MCP server first

Preferred path: call basen-mcp tools

Mode Tool call
Lookup basen_operator_lookup(operator_id, variant)
Search basen_family_browse(family, variant) if keyword matches family; else fallback substring scan
Apply basen_apply(operator_id, input, mode="advisory")
Recommend basen_recommend(problem_description, variant, top_k=3)
Variants read basen://manifest resource
Families read basen://{variant}/family/* resources

If basen-mcp is NOT registered in .mcp.json (current state as of 2026-05-14):

  • Fallback A: try base120-mcp (external TS server, when wired) — mcp__base120__* tools
  • Fallback B: read canonical JSON registry directly: skills/base120-infrastructure/registry/base120_registry.json (repo-root-relative in hummbl-io/agents; that root is ~/.agents on agent-node and Delta). The previously documented PROJECTS/.agents/cognition/data/base120_registry.json does not resolve — verified missing 2026-09-20 — which silently pushed callers to the 22-model Fallback C.
  • Fallback C: built-in reference table (see [base120] skill's quick reference)

Tag results explicitly:

  • [verified via basen-mcp] — primary path
  • [verified via base120-mcp fallback] — TS server fallback
  • [verified via canonical JSON] — direct registry read
  • [UNVERIFIED — all MCP paths unavailable] — no canonical lookup

Step 3: render output

For Lookup mode:

BaseN | <variant>:<code> <Name>
═══════════════════════════════
Family: <family-name>
Definition: <text>
Priority: <N>
Related: <list>
Source: [verified via basen-mcp]

For Apply mode (Alignment report — MTSMU-wrapped):

BaseN Alignment | <variant> | <situation summary>
═════════════════════════════════════════════════
Problem: <1-2 sentence description>
Variant: <variant> (status: LIVE / PARTIAL / NOT canonicalized)

Operators applied:
  1. <variant>:<CODE> <Name> (Priority <N>)
     Why: <why this operator applies>
     Reveals: <the insight>
     Action: <concrete next step>
     Verify-method: [verify-after | cross-check | empirical | unverified]
     Uncertainty: [low | medium | high | unknown] — <one-line basis>
  2. ...

Transformation pattern: <dominant families>
Primary recommendation: <single most important action>

──── MTSMU Receipt ────
Reasoned-by: <invoker identity + session-id-or-tag>
Observed-agent: <who/what will run the Action(s)> — MUST ≠ reasoned-by for trust-root separation per Krineia §3.2
Registry hash: <Base120 v1.0.0 frozen-hash from PROJECTS/base120/Base120_Canonical_Model_Registry.yaml OR "[UNVERIFIED — registry not looked up]">
Falsifiability anchor: <ONE observable that would invalidate Primary recommendation> within <bounded time>
Stance (if stance-gate ran): <descriptive | prescriptive | strategic | negotiative>
Source: [verified via basen-mcp]

MTSMU-wrap rationale (per brainstorm 2026-05-15 Option A): the per-operator verify-method + uncertainty fields make each link in the chain audit-trail-shaped, not just the final recommendation. The MTSMU Receipt footer closes Krineia §3.2 trust-root separation (the agent who reasons must not be the agent whose state is observed). If Observed-agent = Reasoned-by, the chain is self-validating and fails the falsifiability triple's (b) leg — refuse to publish until separation is established or downgrade to operator-local memory only.

Tag definitions:

  • verify-after: the Action produces an observable that will be checked against the Reveal claim within bounded time
  • cross-check: the Reveal is verified by an independent reasoner (sibling session, peer agent) before the Action ships
  • empirical: the Reveal cites a direct measurement (file content, API state, log line) — not inferred
  • unverified: the Reveal is inference-only with no verification path; high-uncertainty by default

For Variants mode: emit the variant status table above + advisory: "Only LIVE variants support [basen] apply. PARTIAL variants may support read-only lookup if their registry has frozen-hash."

Constraints

  • Default variant = base120 (the only LIVE variant as of 2026-05-14)
  • NEVER invent operators or families — use only canonical registry data. Per claim-honesty-protocol.md §3 failure mode #1 (bracketed placeholders) and failure mode #5 (phantom references)
  • NON-canonicalized variants — RETIRED from BaseN catalog (2026-06-13). HUAOMP and MTSMU are methodologies, not operator registries. Use their standalone skills (/huaomp, mtsmu-*) directly.
  • PARTIAL variants (krineia/bki) support read-only lookup of their schemas. Krineia has receipt-chain format; BKI has measurement framework. Neither has a YAML operator registry. Tag results [PARTIAL — variant is a different kernel type, not an operator catalog].
  • SY = Systems (NOT Synthesis). Known hallucination pattern; always verify against registry.
  • Composition grammar deferred per Q3 ratification — chains of operators are operator-discretion, not registry-enforced.
  • Tier-S verify-method tagging (per M02 / M04 doctrine) — when used in Krineia receipts: prefix the Source line with verify-after | cross-check | empirical | unverified.

Cross-variant queries

When a search spans multiple variants:

  • [basen] "feedback loop" — default base120 first; surface PARTIAL variants as "also see krineia (receipt-chain format)" or "also see bki (measurement framework)"
  • Cross-variant search is not yet supported by basen-mcp v0.1. Will be added in v0.2.
  • HUAOMP and MTSMU are no longer in the BaseN variant table. Use /huaomp for epistemic breadth, mtsmu-* skills for rigor wrapping.

Integration

  • /base120 continues working unchanged. Base120 is the first variant of BaseN; the [base120] skill is the variant-specific surface. [basen] base120:P1 and [base120] P1 are equivalent.
  • /aar — when citing operators in AAR sections, prefix with variant if not base120 (krineia:CRR3 once Krineia v1.0 ships).
  • /huaomp — epistemic-breadth methodology. Standalone skill, not a BaseN variant.

Related

  • _internal/governance/basen-architecture/SYNTHESIS.md — full architecture
  • _internal/governance/basen-architecture/OPERATOR-DECISIONS-2026-05-14.md — Q1-Q9 ratification record (BaseN promoted to first-class tier per Q1)
  • _internal/governance/basen-architecture/V1-q7-canonicalization-verification.md — variant canonicalization status (substrate for variant table above)
  • _internal/scratch/basen-mcp/server.py — basen-mcp server v0.1 draft (smoke-tested; awaiting operator approval for staging move + .mcp.json registration)
  • ~/.agents/skills/base120/SKILL.md — the canonical first-variant skill
  • $RUNTIME_MEM/feedback_base120_accuracy.md (resolve via ~/.agents/scripts/resolve-memory.sh) — Base120 canonicalization reference
  • $RUNTIME_MEM/project_basen_tier.md — BaseN first-class ratification memory pin

Open questions (passing through)

  • Krineia v1.0 freeze target Mon 2026-05-18 — if ships, advance Krineia from PARTIAL to LIVE
  • HUAOMP / MTSMU canonicalization path — currently no roadmap; operator may strategically defer indefinitely (the methodologies work fine as advisory frameworks without registries)
  • BKI registry artifact — theory is rich but no schema; operator could commission a BKI registry build separately