Externalized Metacognition — gzkit's Anti-V.I.B.E.S. Architecture (Design Note)¶

Status: Draft North-Star / precursor capture — conversational, pre-decisional. Date: 2026-06-24. Origin: Operator design dialogue (g0), 2026-06-24. Triggered by Lanham, AI Agents in Action 2e — Table 5.1, Fig 10.1, Fig 10.2. Mode: Recorded at operator request ("record this for later use"). This is the Record terminal of the three-mode interaction contract in §6 — not an ADR, not a pool entry, not an OBPI, not a plan. Nothing here is booked or decided. Precursor to: pre-decisional. The gaps in §3 and §7 may later spawn foundation ADRs. None are booked. A future session promotes specific rows only on operator direction. Reading guard (state-doctrine): this is a Layer-1-adjacent design note, never canon and never a derived view. Do not cite it as a decision. It captures thinking, not law.

1. Why this exists¶

gzkit governs artifacts (vertical traceability) and execution (the OBPI pipeline) well. What it has never formed is an explicit organizing model for the agent's cognition itself — the act of reasoning, not the act of recording. This note names that missing layer and maps it against existing surfaces so a later session can see, in one place, what is built, what is prose-only, and what is simply absent.

The operator's opening wager, verbatim:

"gzkit's design intentions are sound, it just needs re-calibration of execution. I've perhaps not paid enough attention to architectural anti-V.I.B.E.S. design."

The skeptical read this note records: those two sentences are in quiet tension, and the second is the more honest. The enforcement intentions (structural binding) exist and are partly built — that part of the wager holds. But the organizing cognition layer was never formed. This is a gap, not execution drift. (Operator agreed with the gap-not-drift read.)

2. The core thesis¶

gzkit's strategic anti-V.I.B.E.S. bet = externalize metacognition out of the stochastic model into a deterministic harness, so that reasoning-mode selection and self-monitoring fire structurally (enforced) rather than by request (stochastic).

A reasoning step that fires by request is a V.I.B.E.S. surface — it depends on the model choosing, in the moment, to do the right cognitive thing. A reasoning step that fires structurally cannot be vibed away. The whole point is to move metacognition from "the model will remember to self-check" to "the harness will not let the model skip the self-check."

Two distinctions from Lanham¶

Reasoning primitives are static; cognitive architecture is dynamic (Fig 10.2). Chain-of-Thought, ReAct, Tree-of-Thought, Reflexion (and self-consistency, plan-execute, least-to-most, debate) are prompt-defined primitives — fixed strategies. A cognitive architecture is the dynamic process that selects among them: Query → Perception (classify the problem) → Attention (select the reasoning strategy) → branch by class {Simple → Direct; Multistep → Planning/decompose; Ambiguous → Exploration/gather-info; Contradictory → Hypothesis-testing} → Evaluation (monitor quality, re-loop on bad output). The primitives become a selection within the process, not a thing the operator hand-picks per prompt.
Failure modes map to cognitive deficits map to fix-modules (Fig 10.1). Each production failure is a missing cognitive function, repaired by an architecture module.

3. The map — cognitive function → gzkit externalized form → maturity¶

This is the heart of the note. Lanham's internal fix-modules, re-expressed as gzkit's externalized (harness-side, deterministic) forms:

Lanham fix-module (cognitive function)	Production failure it stops (Fig 10.1)	gzkit externalized form	Maturity
Evaluation — monitor output quality, re-loop on bad	Confident-wrong-answer	AIRLOCK-OUT drift-diff / `gz validate` (push-edge "what did I break?")	Mature
Confidence gate — know the knowledge boundary	Overcommitted-guess	Behavior Rule 7 — "<90% sure? ask the human"	Prose-only — not mechanical; fires by request, the V.I.B.E.S. surface
Planning — update the model, don't follow a stale plan	Rigid-plan	Phases-in-conversation + Rule 9 "STOP on inconsistency, don't resolve unilaterally"	Partial
Perception / Compositional reasoning — classify the problem, compose sub-results	Shallow-composition	Seam-map / blast-radius (the two-graph)	Waits on the graph engine (Architectural Boundary 3)
Attention / Stagnation-awareness — detect the broken loop	Broken-record	—	Absent — no stagnation detector exists. This is the one module with zero gzkit form.

The Attention/stagnation row is the sharpest finding: it is a genuine hole, not a weak implementation. The nearest unpromoted candidate is ADR-pool.agent-execution-intelligence CAP-10 (analysis-paralysis / stall detection), but it sits in a capability grab-bag, not a cognition architecture.

4. Gap-not-drift: why the layer was never formed¶

gzkit's anti-V.I.B.E.S. is organized around governance ceremony (gates, ledger, attestation) and a trust/safety failure taxonomy (the SIX-pattern axis from the Opus 4.7 / GPT-5.5 system cards: Safeguard-circumvention, Reckless-action, Fabrication, Skipped-cheap-verification, Correction-fails, Dishonest-when-caught). It was never organized around a cognition model.

These are two complementary, orthogonal axes:

Trust/safety axis (gzkit's SIX patterns): will the agent cheat, fabricate, or hide? — already a first-class organizing taxonomy.
Cognition axis (Lanham's five): can the agent perceive, plan, self-monitor, compose, and detect its own stalls? — never formed as an organizing taxonomy; its functions are scattered across prose rules and validators.

Confirming signal: ADR-pool.agent-reliability-framework (ARF) explicitly draws the boundary and leaves the cognition side empty — "ARF deliberately does not try to solve the hard problem (agent cognition). It solves … artifact verification." The cognition axis is exactly the empty quadrant.

5. What already exists, scattered (overlap inventory)¶

So a future promoter knows the pieces are not zero — they are unorganized:

ADR-pool.interpretability-hardened-agent-surfaces — states the doctrine verbatim ("agent-facing surfaces must bind to structural evidence, not narrative recall … replacing 'agent will remember to do X' with 'the system will not let agent X drift'") but scopes to three named surfaces, not an architecture.
ADR-pool.agent-execution-intelligence — holds the scattered cognition candidates: CAP-10 stall detection (≈ Attention), CAP-21 predictive-failure-match (≈ Perception), CAP-09 goal-backward verify (≈ Evaluation). A capability list, not a unifying model.
ADR-pool.harness-aware-execution-modes — the enforcement substrate (Mode-1 skill-chain vs Mode-2 hooks) the cognition layer would run on — not the layer itself.
ADR-pool.agent-reliability-framework — the verification axis; disclaims cognition (see §4).

No existing ADR takes the cognitive-architecture lens as gzkit's organizing model. That absence is finding (a).

6. Methodological coda — finding (b), the interaction contract, the diet¶

Recording finding (a) surfaced a second finding of the same shape.

The shared shape: the apparatus lacks an explicit home for a kind of work, so that work deforms to fit the nearest recognized container.

(a) Metacognition has no home → it deforms into prose Behavior Rules and scattered validators.
(b) Pre-ADR design has no recognized home → it deforms into pool ADRs. The pool stands at 167 entries, many of them note-shaped design documents wearing ADR-pool costume. Premature crystallization is itself a V.I.B.E.S. failure — ceremony applied where exploration belonged, the mirror image of prose applied where structure belonged.

Operator, verbatim:

"We don't have a design level that predicates ADR creation." "gzkit needs to be on a diet."

The skeptical correction this note records (do not over-read (b)): the precursor tier is not missing — it already exists informally and is heavily used (docs/governance/ holds ~30 strategy/North-Star/research notes, including this one's siblings harness-loop-engineering-strategy-note, four-phases-of-work, external-proving-ground-note). What is missing is recognition: the workflow vocabulary (Constitution → PRD → ADR → OBPI → REQ → TASK) skips it, and the gz-design skill's Hard Gate forbids landing there. The fix is therefore small — recognize, don't invent a tier. Building a new formal tier with CLI/schema/lifecycle would be over-production: answering "you over-crystallize into ceremony" with more ceremony is self-refuting. The diet is the fix for (b): the agent stops defaulting past the tier that already exists.

The three-mode interaction contract (operator holds the triggers)¶

Mode	Default?	Trigger	Terminal
Conversational	Yes	—	Nothing written. Think out loud.
Record	No	Operator says "record / capture this"	A precursor doc in `docs/governance/` (status header, "precursor to:", recognized home). This note is an instance.
Design	No	Operator says "let's design X"	`gz-design` runs; an ADR is the legitimate terminal.

The agent's job is to name the mode and flag the seams — surface when a conversational thread has hardened into something worth recording, or has crossed into real design, and offer the pivot. The operator pulls the trigger; the agent never auto-crystallizes.

The one code-level tooth¶

gz-design's Hard Gate — "All artifacts land in GovZero structures: pool ADRs, canonical ADRs, OBPI briefs" — structurally cannot not yield an ADR. That is finding (b) encoded in the skill. A one-line relaxation (add "foundation/precursor document" as a first-class terminal) closes it. Not proposed here; recorded as the single load-bearing change available when wanted.

7. Considered and not taken (rejected alternatives)¶

Recorded per the interpretability discipline (a destination written backward is indistinguishable from a reasoned one):

Book this as an ADR now. Rejected — premature crystallization; pre-decisional; violates the diet. The whole session's lesson was not to do this.
Invent a new formal precursor tier (CLI verbs, schema, lifecycle, gates). Rejected — over-production; the tier already exists informally; self-refuting to answer ceremony-bloat with ceremony.
Treat (b) as a pure agent-disposition tweak, change nothing. Rejected as under-reaction — the Hard Gate provably forces the ADR terminal and pool=167 shows real deformation; recognition debt would persist and re-deform next session.
Build the Attention/stagnation detector now (finding (a)'s one true hole). Not taken — no decision to build; recorded as a candidate gap only.

8. Open threads — pre-decisional, none booked¶

Gaps a later session may (on operator direction) promote into foundation ADRs:

Attention/stagnation detector — the only cognition module with zero gzkit form (§3).
Confidence-gate mechanization — move Behavior Rule 7 from prose to a structural fire (§3).
Precursor-tier recognition — name the tier in the workflow vocabulary + relax the gz-design Hard Gate (§6). A correction, not a new tier.

No ADR booked. No pool entry created. No OBPI scoped. The floor remains the operator's.