> For the complete documentation index, see [llms.txt](https://osintelligence-llc.gitbook.io/osintelligence/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://osintelligence-llc.gitbook.io/osintelligence/what-is-novel-in-this-record.md).

# What is novel in this record

A legend for anyone deciding whether this record rewards a deep read. It names what was done, plainly, not what it resembles. Quality is reported at parity where it is parity; only the novelty is stated without hedging. The five headline claims come first, each with the sealed receipt that proves it. The full map follows: one line of novelty per chapter, linked, so a skimmer can find the deep water and go straight to it.

## The five that most reward a check

1. A 122-billion-parameter mixture-of-experts model, pruned and calibrated by a single operator, runs on one consumer 12 GB GPU and clears a pre-registered served-quality gate at parity or better against the build it replaced.\
   **Receipt.** Sealed [D7](/osintelligence/evidence-and-seals/pruned-122b-quality-gate-d7.md) pre-registration and gate verdict, hashed before the run.
2. Two specialist agents were trained into their weights on the operator's own corpus, with no external teacher model, and each beats a larger prompted baseline on a held-out probe.\
   **Receipt.** Sealed [P4](/osintelligence/evidence-and-seals/ioc-classifier-micro-agent-p4.md), a 0.8B classifier at 100% versus 22% for a prompted 2B. Sealed [P6](/osintelligence/evidence-and-seals/cti-ner-a-b-p6.md), a 0.8B CTI-NER at 2.79 times the 9B orchestrator's relaxed entity-F1, at 11 times smaller.
3. A production heap over-read in a serving engine's multimodal position path was isolated by cross-toolchain falsification rather than guesswork, closed with an upstream fix cherry-picked onto a pinned build, and validated across 2,004 production rows.\
   **Receipt.** Sealed [P7.5 and P7.6](/osintelligence/evidence-and-seals/mrope-serving-path-rca-p7.5-p7.6.md), upstream [issue #20093](https://github.com/ggml-org/llama.cpp/issues/20093).
4. Deterministic governance gates ran as the live harness around an autonomous agent for 64 continuous days, recording 390,515 gate fires and 1,142 hard blocks, every figure a direct count rather than an estimate.\
   **Receipt.** Sealed [66-file telemetry window](/osintelligence/part-iii-the-evidence-what-broke/11-the-hook-telemetry-record.md), instrument wired before the window opened.
5. The sealing discipline behind every experiment here is open source, and its hash routine reproduces, byte for byte, the same seal the record's own experiments were sealed under.\
   **Receipt.** [MethodSeal](/osintelligence/the-instrument-methodseal.md), DOI [10.17605/OSF.IO/H64GM](https://doi.org/10.17605/OSF.IO/H64GM), manifest self-hash a371bda.

***

## The full map

Every chapter, in reading order, with the one thing it did that was new in the doing. Each chapter page opens with the same two lines in fuller form: a **What is new here** and a **Deepest water** pointer to the single load-bearing result.

**Keystone.** [The Why](/osintelligence/the-why.md) names the architectural-softening cascade, a capable model asked to harden its own gates authoring an escape hatch into each and unable to surface the pattern itself, and draws the comprehension-not-instruction distinction the field tends to fuse.

### Part I · The architecture

* [1 · The Sovereign Triad](/osintelligence/part-i-the-architecture/1-the-sovereign-triad.md) names three structurally distinct failure classes and closes each with exactly one mitigation, and identifies verification from outside the loop, in a separate hardware-trust domain, as an architectural class current alignment discourse omits.
* [2 · The External Sentinel](/osintelligence/part-i-the-architecture/2-the-external-sentinel.md) makes that external governor a concrete device whose decision is a deterministic byte-equality check rather than a learned classifier, with the one act it must never self-perform gated by a physical write-protect switch no software can throw.
* [3 · Sovereign Safety Architecture](/osintelligence/part-i-the-architecture/3-sovereign-safety-architecture.md) re-decomposes the triad into four mechanism families, the Sovereign Tetrad, and names temporal containment as the peer family the triad omits, the one that closes drift compounding across cycles without any single cycle tripping a gate.
* [4 · Stateless by Construction](/osintelligence/part-i-the-architecture/4-stateless-by-construction.md) treats the agent as amnesiac by design: a grounding gate that works by recency rather than contradiction, and continuity shown separable from progress across a four-day run where the roadmap never advanced but the handoff held.

### Part II · The discipline

* [5 · Sixteen Practices](/osintelligence/part-ii-the-discipline/5-sixteen-practices.md) grades each practice by honest provenance, convergent, novel-application, or novel-technique, and pre-registers that honesty with dated public prior art, beneath the Sovereign Pair and the industry inversion it names.
* [6 · Multi-Agent OODA Mesh](/osintelligence/part-ii-the-discipline/6-multi-agent-ooda-mesh.md) names and formalizes the human-as-routing-layer mesh, the structural dual of the manager-agent pattern, with five load-bearing properties each carrying its own falsifier.
* [7 · The Sovereign Corpus Engine](/osintelligence/part-ii-the-discipline/7-the-sovereign-corpus-engine.md) reconstructs each operator correction as a counterfactual triple and confidence-routes its own curation, making production work into training data as a byproduct.
* [8 · Sovereign Optimization Flywheel](/osintelligence/part-ii-the-discipline/8-sovereign-optimization-flywheel.md) shows the five-axis compound is multiplicative rather than additive, because every layer acts on one shared substrate.

### Part III · The evidence: what broke

* [9 · The Drift Taxonomy](/osintelligence/part-iii-the-evidence-what-broke/9-the-drift-taxonomy.md) sorts nine production failure classes by one rule, will the constraint hold under friction, splitting them cleanly into five that close mechanically and four that only a human can close.
* [10 · The Guard Changes at 23:26Z](/osintelligence/part-iii-the-evidence-what-broke/10-the-guard-changes-at-23-26z.md) is a pre-registered, timestamp-bisected natural experiment across a model swap the substrate could not be reconstructed to fake, and the Cascade Catalogue it produced.
* [11 · The Hook Telemetry Record](/osintelligence/part-iii-the-evidence-what-broke/11-the-hook-telemetry-record.md) is the measurement itself: deterministic enforcement shown continuous and load-bearing at 1,142 hard blocks over 64 days, 72% of them on two boundaries, every figure a direct count.
* [12 · The Public Case Record](/osintelligence/part-iii-the-evidence-what-broke/12-the-public-case-record.md) states a four-part inclusion rule that turns dramatic anecdotes into checkable evidence, then answers the single-operator objection with disinterested-party corroboration.

### Part IV · The evidence: what worked

* [13 · Sovereign IOC Classifier](/osintelligence/part-iv-the-evidence-what-worked/13-sovereign-ioc-classifier.md) is the first deployable proof of in-weights specialization at micro-agent scale, and the precedent template every later specialist inherits.
* [14 · Sovereign CTI-NER](/osintelligence/part-iv-the-evidence-what-worked/14-sovereign-cti-ner.md) is a sovereign silver-label adapter that Pareto-dominates the larger incumbent, with the silver-label supply-chain threat named and a falsified primary hypothesis kept in full.
* [15 · Sovereign Imatrix Calibration](/osintelligence/part-iv-the-evidence-what-worked/15-sovereign-imatrix-calibration.md) shows the calibration corpus is a first-order fidelity lever, 16 times tighter divergence, and reframes it as a supply-chain surface where the defense is also the quality win.
* [16 · Sovereign Domain Pruning](/osintelligence/part-iv-the-evidence-what-worked/16-sovereign-domain-pruning.md) is the four-file integrity contract and the mismatch control, whose two ancillary failures are kept as the honesty architecture rather than dropped.
* [17 · Sovereign Big-Model Compression](/osintelligence/part-iv-the-evidence-what-worked/17-sovereign-big-model-compression.md) prunes a 122B mixture-of-experts onto one 12 GB card at parity with the build it replaced, with the in-distribution perplexity asymmetry as the signature and a parallel 35B attempt reported as a composite failure.
* [18 · Corpus-Sovereign Self-Distillation](/osintelligence/part-iv-the-evidence-what-worked/18-corpus-sovereign-self-distillation.md) is a pre-registered pilot whose primary hypothesis was falsified and reported in full, leaving a stronger survivor finding that a standard benchmark would have missed.
* [19 · Watcher KL-Drift Floor](/osintelligence/part-iv-the-evidence-what-worked/19-watcher-kl-drift-floor.md) finds an architectural drift floor by measurement across six sealed checkpoints rather than by design, with a targeted-attack checkpoint foreclosing the incidental reading.
* [20 · mRoPE Serving-Path RCA](/osintelligence/part-iv-the-evidence-what-worked/20-mrope-serving-path-rca.md) is falsification-first root-cause analysis, a toolchain flip used as an isolation step rather than a fix, and an attribution-preserving upstream patch bundle.
* [21 · The Watcher](/osintelligence/part-iv-the-evidence-what-worked/21-the-watcher.md) names and first builds the correction-reflex governor layer between the gates and the operator, with phantom tools as a governance channel and a published six-round negative as its honest spine.

### Part V · The frontier

* [22 · Sovereign Sustainability](/osintelligence/part-v-the-frontier/22-sovereign-sustainability.md) projects the sealed efficiency envelope to industrial scale as a dual-claim thesis against the 2030 demand curve, argued from an existence proof and leaving the Jevons paradox measurable rather than argued away.
* [23 · Built Portable](/osintelligence/part-v-the-frontier/23-built-portable.md) maps the coordination fabric surface-for-surface onto robotics middleware, so an embodied port is a transport swap rather than an architecture rewrite.
* [24 · The Weaver Paradigm](/osintelligence/part-v-the-frontier/24-the-weaver-paradigm.md) is a horizon discipline that tags every roadmap claim ground-truth, corrected, or aspirational, and the Three-Layer Governance Chain that terminates in a human hand on a physical switch.

### The instrument

* [The Instrument · MethodSeal](/osintelligence/the-instrument-methodseal.md) is the sealing discipline itself, open-sourced: the fifth headline claim above, and the mechanism every sealed receipt in this record was made with.

***

*This legend names novelty only. What each contribution does not claim, and the parity-not-gain discipline behind the numbers, is set out in* [*How to read this record*](/osintelligence/how-to-read-this-record.md)*.*