Skill Runtime: activation beats re-derivation, acquisition stays sovereign
Ran Tao (Octoryn Research)
Abstract
Skill Runtime activates a learned, parameterized action pattern instead of re-deriving it, removing the planning round-trips re-derivation pays: a structural four-to-zero reduction. One single-run small-model case was byte-correct, its large wall-clock win latency-dominated and non-generalizable. A model-free, embedding-free inverted index over intent tokens keeps matching flat and precise as the library grows. Acquisition stays sovereign: a model-free miner is producer of record; a teacher only proposes behind a model-free judge. Distillation by a too-small model failed. A smoke test.
1. Contribution
Claimed. This report establishes a procedural-memory runtime ("Skill Runtime") for agents in which a learned, parameterized action pattern (a "Skill") is activated rather than re-derived. The validated results are: (i) activation eliminates the planning round-trips that re-derivation pays, a structural 4 -> 0 reduction; under a small real model the resulting single-case wall-clock ratio is large and byte-correct, but it is dominated by that model's per-round-trip latency and reported single-run; (ii) data-flow skills thread their own intermediates (single-producer scalar bindings) and activate from external inputs alone; (iii) an embedding-free, document-frequency-pruned inverted index over intent tokens keeps library matching flat (microsecond-scale) to a large library at full precision — the "Utility Problem" smoke test; (iv) acquisition is sovereign: a model-free pattern-miner is the producer of record, and a larger teacher only proposes behind a model-free judge, on which it failed on output discipline; (v) stale/retracted skills are never confidently activated, and every step passes a pre-dispatch governance gate.
NOT claimed. This is a correctness-and-engineering validation, not a novel algorithm and not a benchmark-scale evaluation. We do not claim generation quality of any model, throughput at fleet scale, or that the single-run wall-clock ratio generalizes (it is dominated by one model's per-round-trip latency and is single-run with no variance reported). We do not claim distillation works with an under-capable small model — it failed. We do not claim expressiveness beyond single-producer scalar bindings (no fan-in/branches/loops). We do not claim the trajectory set represents a live recurring-task fleet. The altitude is: a sovereign, audited activation/acquisition mechanism demonstrated on a small, mostly-synthetic stimulus set.
2. Methods
| Field | Value |
|---|---|
| Runtime model (acquisition/activation) | Model-free pattern-miner; zero model on activation path |
| Real model for round-trip timing | A small real model (sub-1B class) |
| Teacher (proposal-only) | A larger teacher (tens-of-billions class), behind a deterministic format-normalizer and a model-free judge |
| Distillation control | Small-model distiller front-end; rule-distiller control |
| Library matching baseline | Naive O(N) scan vs. document-frequency-pruned inverted index over intent tokens |
| Procedure | Distill skill from a successful trajectory; activate instead of re-derive; gate every step pre-dispatch |
| Metric + gate | Round-trip count (4->0); wall-clock ratio; match latency; precision; byte-correctness; producer/judge pass-rate |
| Stimulus / inputs | Representative-synthetic trajectories (with a real small-model forward) plus one real-task end-to-end acquisition; a simple data-flow example mapping an input key to derived scalar outputs |
| Index scale | Up to a large library (ten-thousand-skill order) |
| Precision/correctness | Byte-correct activation; a cryptographically chainable, byte-deterministically-replayable event log |
| Hardware | Consumer Apple Silicon laptop |
| Stack/versions | Not recorded in source notes |
| Sovereignty | No third-party ML framework; platform numeric primitives only |
3. Results
| Measurement | Value |
|---|---|
| Planning round-trips, re-derive -> activate | 4 -> 0 |
| Wall-clock win (single small-model case) | Large, byte-correct, single-run, latency-dominated |
| Naive O(N) match cost crossover (Utility Problem) | Order of several-thousand skills |
| Inverted-index match latency to a large library | Flat, microsecond-scale |
| Inverted-index precision | 100% |
| Pattern-miner producer pass-rate | Passed every trial |
| Larger-teacher pass-rate (output discipline) | Failed every trial |
| Small-model distiller usable-skill rate | Failed every trial |
| Rule-distiller control | Passes every time |
| Stale/retracted false-confident dispatches | 0 |
| Governance-gate side-effects (over-privileged injected steps) | 0 |
4. Ablations / Negative controls
| Control | Result | Interpretation |
|---|---|---|
| Larger teacher vs. model-free miner | Teacher fails, miner passes | Teacher proposes; the model-free judge disposes — teacher is not the authority |
| Small-model distillation front-end | Not usable | Reflection front-end needs a capable model |
| Rule-distiller control | Passes every time | The learning step itself is sound; failure is model-capacity, not mechanism |
| Stale / retracted skill activation | 0 dispatches | Never confidently activated |
| Over-privileged injected step | 0 side-effects | Pre-dispatch governance gate denies |
Measured utility curve
Freshly measured on consumer Apple Silicon with a model-free harness, median of repeated reps, microsecond scale. The re-derive wall is N-independent (four planning round-trips). Naive linear-scan matching grows roughly O(N) and crosses the re-derive wall at a several-thousand-skill library size (the Utility Problem, measured): beyond that point a naive scan is slower than simply re-deriving. The document-frequency-pruned inverted index stays flat at microsecond scale (it visits a single candidate, not N) with no crossover through the largest library tested — at the high end it visits one skill versus the scan's full library (orders of magnitude fewer). Match precision was perfect: the target was picked over all distractors at every library size.
5. Threats to validity / Limitations
- Sample size is a smoke test, not a benchmark. Producer/judge and distillation rates are over a handful of trials; index precision is over a synthetic intent-token set. These are correctness smoke tests, not benchmark-scale evaluations.
- Single-run, no variance. The wall-clock ratio is one timing case dominated by a single small model's per-round-trip latency. No variance, confidence interval, or repeated trials are reported; the structural 4->0 is the model-independent claim, the ratio is not.
- Stimulus is mostly synthetic. Trajectories are representative-synthetic (with a real small-model forward) plus exactly one real-task end-to-end acquisition — not a live fleet of recurring tasks.
- Parity/correctness != quality. Byte-correct activation says nothing about the generation quality of any model in the loop.
- Expressiveness bound. Data-flow is single-producer scalar bindings only — no fan-in, branches, or loops; results do not extend beyond that shape.
- Distillation capacity bound. A sub-1B model is insufficient for the reflection front-end. The result holds only where distillation uses a capable model; the larger teacher is proposal-only.
- Index trade-off is deliberate. Flat matching avoids embeddings to preserve model-free sovereignty; it is not compared against a learned retriever and makes no claim of recall on fuzzy/paraphrased intents.
- evidenceLevel. Retained as "experimental"; the single-run perf ratio and small sample sizes would not support a stronger claim.
6. Reproducibility / Provenance
| Item | Value |
|---|---|
| Identifier | TR-2026-0006 |
| Producer of record | Sovereign model-free pattern-miner, zero model on activation |
| Library | Event-sourced, cryptographically chainable, byte-deterministically-replayable log |
| Validate-first caveat | Validate the model-free judge before trusting any teacher proposal — the judge, not the teacher, is the authority |
| Build / run | Standard release build and test of the project |
| Hardware / stack | Consumer Apple Silicon laptop |
| Sovereignty | No third-party ML framework; platform numeric primitives only |
Claim boundary
The author's explicit scope — what this work does and does not establish — carried over from the Octoryn Research publishing model.
Proves
- Activating a learned skill eliminates planning round-trips, a structural reduction from four to zero; one single-run timing case with a small real model is byte-correct.
- An embedding-free inverted index over intent tokens keeps library matching flat and high-precision as the library grows large, without learned retrieval.
- Acquisition is sovereign: a model-free pattern-miner is the producer of record while a larger teacher only proposes behind a model-free judge that holds final authority.
- Stale or retracted skills are never confidently activated, and every step passes a pre-dispatch governance gate with no observed side-effects.
Does not prove
- That distillation by an under-capable small model is viable; in this work it failed, so the reflection front-end needs a capable model.
- That the single-run wall-clock win generalizes; it is one timing case dominated by one model's per-round-trip latency, with no variance reported. Only the four-to-zero structure is model-independent.
- Expressiveness beyond single-producer scalar bindings; no fan-in, branches, or loops are demonstrated.
- Generation quality, fleet-scale throughput, or recall on fuzzy or paraphrased intents; the flat index makes no recall claim against a learned retriever.
Applies when
- Recurring tasks with a stable tool-chain shape, where activation is fully sovereign with zero models on the activation path and every step is audited.
Does not apply when
- One-shot non-recurring tasks; skill distillation performed by an under-capable model; or workloads requiring fuzzy or paraphrased intent recall the flat index does not claim.
Authors
- Ran Tao — Investigation, Writing
Cite this
Citation
Tao, R., Octoryn Research. (2026). Skill Runtime: activation beats re-derivation, acquisition stays sovereign (TR-2026-0006). Octopus Research Institute.
BibTeX
@techreport{oritr20260006,
title = {Skill Runtime: activation beats re-derivation, acquisition stays sovereign},
author = {Tao, Ran and {Octoryn Research}},
institution = {Octopus Research Institute},
year = {2026},
note = {Permanent ID TR-2026-0006. Not peer reviewed.}
}Disclosures
- Funding
- Hardware and infrastructure provided by Octoryn / Octopus Core Pty Ltd.
- Conflicts of interest
- Octoryn ships commercial inference and governance tooling; findings are reported independently.