# Ln Oracles

> Design verification strategy: diagnose observability, select oracle families, map to loop tiers, surface blind spots. Use after ln-plan when frontier items or scoped slices need oracle design — especially for LLM, visual, or compositional work — or when verification coverage has drifted.

- Skill: `hashintel/ln-oracles` (Agent Skill, multi-file: 3 files)
- Install (CLI): `npx skillmds@latest add hashintel/ln-oracles`
- Raw SKILL.md: https://api.skillmd.com/api/skills/hashintel/ln-oracles/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: hashintel (https://skillmd.com/u/hashintel)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/hashintel/ln-oracles

---


# Ln Oracles

Design what proves the system works before choosing how to build it.

The best oracle removes the most bad degrees of freedom per unit time (Regehr). A system without feedback is open-loop -- it cannot correct errors (Wiener). Verification is first-class work, not accessory: second only to building the product itself. A slice without an oracle strategy is not scoped.

Not every slice needs a full oracle-design pass. For trivial, purely structural slices, `ln-scope` may name the inner-loop checks directly. Use `ln-oracles` when the verification strategy itself is uncertain or materially shapes implementation order.

Do not create standalone oracle-planning docs without explicit permission. Oracle design reconciles back into `memory/SPEC.md` and `memory/PLAN.md`.

Read the [diagnostic framework](assets/diagnostic-framework.md) and [oracle taxonomy](assets/oracle-taxonomy.md) before starting.

## Input

The frontier items or scoped slices to design oracles for: $ARGUMENTS

Read `memory/SPEC.md` (invariants, assumptions, decisions, verification design) and `memory/PLAN.md` (frontier definitions, sequencing, acceptance criteria). If `memory/SPEC.md` already has a §Verification Design section, this is an update -- read it as prior state to evolve, not preserve uncritically.

## Procedure

This is an **interactive process** -- each step involves presenting analysis and grilling the user for pushback, corrections, and design input. Do not produce a finished oracle design in one pass. Walk the design tree like `ln-grill`: present your assessment, ask pointed questions, incorporate answers, then move to the next step.

### 1. Diagnose

Score **Observability**, **Reproducibility**, and **Controllability** (see the [diagnostic framework](assets/diagnostic-framework.md)). Present the scoring table to the user with specific notes per dimension. Low scores constrain which oracle families are feasible and must be addressed before oracle selection proceeds.

**Grill**: For each dimension scored below `high`, ask: is this a deliberate deferral, a blind spot, or something we should address now? What would change the score?

### 2. Extract verification claims

From `memory/SPEC.md` invariant bundles, acceptance criteria, `memory/PLAN.md` frontier definitions, and any in-hand scope-card slices -- list what must be proved. Distinguish:

- **Structural claims** (schema conformance, DB round-trips, type safety) -- oracle-able programmatically
- **Behavioral claims** (LLM output quality, UX judgment) -- require human assessment or statistical thresholds
- **Compositional claims** (graph integrity over time, cumulative drift) -- may need property-based or model-based oracles

**Grill**: For behavioral claims, ask: is schema validation sufficient for the inner loop, with qualitative assessment deferred to manual testing? Where is the boundary between "structurally valid" and "qualitatively good"?

### 3. Select oracle families

Using the [oracle taxonomy](assets/oracle-taxonomy.md), select families ranked by ROI for this project's verification needs. Apply the combination principle: the best oracle is a pair of independent artifacts. Prefer pairs when they compound; don't force them when a single oracle suffices.

**Grill**: For each selected family, present: what it proves, what it costs, and what it misses. Ask the user which tradeoffs are acceptable given timeline and confidence levels.

### 4. Map to loop tiers

Assign each selected oracle to inner (ms, agent-autonomous), middle (seconds-minutes, regression/fitness), or outer (slow hardening). Apply verification economics: cheapest checks first, expensive checks less often.

**Boundary with ln-spec**: ln-spec owns project-wide inner-loop verification commands, policy, and fast automated checks. ln-oracles owns the middle and outer loops, plus strategic framing (diagnostic, stance, blind spots), and may recommend slice-specific inner-loop oracle families when they affect implementation strategy. When updating, preserve ln-spec's command/policy content and extend with middle/outer strategy.

**Grill**: For middle-loop oracles that require external resources (API calls, fixtures), ask: how will fixtures be created? What bootstraps ground truth? Is single-shot measurement sufficient or do we need multi-run variance?

### 5. Design per-frontier / per-slice verification approach

For each in-scope frontier item in `memory/PLAN.md`, specify: which oracles apply, what they prove, and which loop tier they belong to. This becomes the `Verification` annotation in the frontier definition. If a scope-card slice is already available, add slice-level oracle notes there without promoting detailed card history into `memory/PLAN.md`.

**Grill**: For each slice, ask: does this oracle strategy cover the slice's acceptance criteria? What's the gap between "oracle says pass" and "slice is actually correct"?

### 6. Surface blind spots

Name what this verification design does NOT cover and why. Categories: cost exceeds value, observability gap, deferred to later milestone, not oracle-able. A verification design with no blind spots is incomplete.

**Grill**: For each blind spot, ask: is this an acceptable deferral or a risk that should promote to "needs an oracle now"? What would trigger revisiting?

## Output

Update `memory/SPEC.md` §Verification Design:

- **Verification Stance** -- the meta position on verification as first-class work
- **Diagnostic Assessment** -- O/R/C scoring table with notes
- **Oracle Strategy by Loop Tier** -- middle and outer loop oracle families, mapped to what they prove (preserve ln-spec's inner loop content)
- **Design notes** -- project-specific oracle design decisions (e.g. observer history projection, fixture bootstrapping strategy)
- **Acknowledged Blind Spots** -- table with blind spot, reason, mitigation, and revisit trigger

Update `memory/PLAN.md` frontier annotations:

- Add or refresh the `Verification` line in each in-scope frontier definition with oracle family, loop tier, and cross-reference to `memory/SPEC.md` sections
- Keep slice-level oracle detail in the current `ln-scope` card or `memory/CARDS.md` queue unless it changes the frontier definition

### Cross-reference integrity

After writing, verify:
- Every `memory/SPEC.md` invariant has at least one oracle assigned (inner, middle, or outer)
- Every in-scope `memory/PLAN.md` frontier definition has a verification approach annotation
- The blind spots section is non-empty
- Middle/outer loop oracles cross-reference the invariants or assumptions they prove

## Routing

After filing the oracle design, present these options to the user (use `tool-ask-question`):

| #   | Label          | Target       | Why                                              |
| --- | -------------- | ------------ | ------------------------------------------------ |
| 1   | Scope a slice  | `ln-scope`   | Oracle design is complete, ready to scope work   |
| 2   | Spike first    | `ln-spike`   | Oracle feasibility is uncertain for a key claim  |
| 3   | Revise spec    | `ln-spec`    | Oracle design revealed spec gaps                 |
| 4   | Revise plan    | `ln-plan`    | Oracle design changes slice ordering or priority |
| 5   | Back to triage | `ln-consult` | Direction needs reassessment                     |

Recommended: **1** unless oracle feasibility is uncertain (then **2**).