> 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/evidence-and-seals/matched-corpus-compression-chain-d6.md).

# Matched-Corpus Compression Chain (D6)

*Evidence & seal for* [***18 · Corpus-Sovereign Self-Distillation***](/osintelligence/part-iv-the-evidence-what-worked/18-corpus-sovereign-self-distillation.md)*. This is the **matched arm** to the* [*35B mismatch control*](/osintelligence/evidence-and-seals/35b-prune-experiment-p3-mismatch-control.md)*: the identical MoE-Slimming compression chain, same banks, same binary, same bars, with the sovereign architect corpus substituted at every corpus-touching layer (prune calibration, recovery-FT, imatrix). It was pre-registered as the answer to a falsifiable bet, and it returned a **composite FAIL whose cause was a reframe of the question, not the data**. Reading that verdict correctly is the whole point of publishing it.*

*New to how these seals work, read* [***Verifying a Seal***](/osintelligence/evidence-and-seals/verifying-a-seal.md) *first.*

**Experiment id** `osint-suite-dev-2026-07-12-7e42` · **pre-registered** 2026-07-12 · **concluded (Mode D)** 2026-07-14 · **verdict composite FAIL** (H1 FAIL · H2 PASS · H3 INCONCLUSIVE; premise invalidated by reframe).

## The seal (quoted from the sidecar, verify it yourself)

```
5381685d0f8c6587aa738ceee52de0ff9fe379fc17e890f17c2381e3d20f4bc1  PRE_REGISTRATION.md
prefix-bytes-hashed: 26703 · raw-file-bytes: 28564 · marker-position-lf: 26704
computed AT 2026-07-12T11:38:21Z (canonical-prefix per seal_prereg.py; hex in the sidecar + marker, never inlined)
```

At Mode D conclusion the post-run integrity envelope re-computed this prefix hash in-session and it was an **exact match** against the experiment marker, the methodology was unchanged from Stage G through Stage R (15 frozen artifacts held fresh). The pre-registration carried **three pre-data amendments** (DEV-1, DEV-2, DEV-3), each ruling on the minority-corpus source before any result bytes existed; their strikethrough corrigenda are preserved in §3/§4.1 of the frozen text below, and each amendment's pre-amendment self-hash is preserved in the amendment itself.

## Verdict (final seal, adjudicated on the frozen bars)

**D6 composite FAIL** (PASS required H1 ∧ H2 ∧ H3; H1 FAIL forecloses). The incumbent `[architect-apex]` UD-Q5\_K\_XL was retained.

| Gate                                                 | Rule (sealed §9)                                                        | Measured (summary `d1660cd7…`)                                                                                                         | Verdict          |
| ---------------------------------------------------- | ----------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------- | ---------------- |
| **H1** pooled math (vs deployed apex)                | Δ ≥ −5.0 pp (n=80 paired)                                               | 56.25% vs 67.5% → **Δ −11.25 pp** (McNemar p=0.0225; MATH-500 −14.0 p=0.039, AIME24 −6.67 n.s.)                                        | **FAIL**         |
| **H2** GPQA-198 (vs deployed apex)                   | Δ ≥ −6.0 pp                                                             | 47.98% vs 29.80% → **Δ +18.18 pp** (p=8.7e-5)                                                                                          | **PASS**         |
| **H3** matched vs frozen mismatch control (THE bet)  | Δm > 0 on GPQA-198 AND one-sided p < 0.05, pooled math not sig-negative | GPQA Δm **+2.53 pp in-direction** (discordant 40/35, one-sided p=0.322); pooled math Δm −10.0 pp (two-sided p=0.096, not sig-negative) | **INCONCLUSIVE** |
| **SERVE row** (§13.2, mechanical, separate sentence) | resident, no spill, ≤ 12 GB, ≥ 63.8 tok/s                               | IQ4\_XS 11,095 MiB with spill signature, 9.12 tok/s; Q4\_K\_M dominated                                                                | **SERVE-FAIL**   |
| H4 drafter · H5 persona                              | —                                                                       | **NOT RUN** (concluded out by operator direction, §3 below)                                                                            | —                |
| **Composite**                                        | H1 ∧ H2 ∧ H3                                                            | —                                                                                                                                      | **FAIL**         |

The corpus sentence (H3 INCONCLUSIVE) and the serving sentence (SERVE-FAIL) are pre-registered as **separate results, never collapsed**, and no multiplicativity claim is made from a whole-chain binary comparison (the interpretation boundary is pre-registered in §2 of the frozen text).

## Why the FAIL is not what it looks like (the sealed framing guardrail)

The experiment was **concluded at the honest FAIL without running the remaining stages, because its core premise was invalidated by a reframe, not by the data.** The final seal records three framing errors in the pre-registration's own design: it gated on the **wrong axis** (capability/math retention, where specialization is properly measured on a tool-calls-to-objective and typed-restraint axis), pushed a **role corpus into the base slot** (a mis-layering against the base/LoRA/behavioral split), and used the **wrong model** for that corpus's layer. A broken premise cannot be repaired by amendment, so concluding cleanly at the FAIL is the pre-registration discipline working as designed.

The seal carries an explicit guardrail against misreading, reproduced here because it governs how this evidence must be cited: the incumbent stays as base **because it is the correct broad general-reasoning substrate for the Architect, not because "sovereign-corpus lost a bake-off."** The D3-mix cells were never a production candidate; they are the pre-registered mismatch control. **Any reading of this record as "sovereign-corpus underperforms" is false.** The true finding is that a role-flavored corpus was mis-layered into the base slot and H1 correctly caught the base going too narrow, the domain-shift signature (GPQA up versus both baseline and control, math down versus both) is the fingerprint of that mis-layering, not of corpus weakness. The redraw onto the correct head-to-head framework (in-weights versus in-context crossed with governance on/off, anchored on tool-calls-to-objective) is authored as a full rewrite, not an amendment of this experiment.

## Integrity envelope (10 artifacts, SHA-256, computed at-claim 2026-07-14)

| Artifact                                                                                                                         | sha256\[:16]       |
| -------------------------------------------------------------------------------------------------------------------------------- | ------------------ |
| `d6_results.jsonl` (556 per-item records, 639,301 B)                                                                             | `d7fe6d12fadfd301` |
| `d6_retention_summary.json` (H1/H2/H3\_corpus source)                                                                            | `d1660cd72f1f972f` |
| `d6_bank_manifest.json` (**byte-identical to the P3 and D7 bank manifests** — one literal bank surface across three experiments) | `22531c4d84c8fbcd` |
| `telemetry/d6_eval35…/run_manifest.json` (Stage R; both-arm GGUF shas + H3 control binding)                                      | `707c60eaa766d836` |
| `telemetry/d6_quant35…/quant_receipt.json` (Stage Q; SERVE rows + corpus\_identity)                                              | `f98935a2caf56614` |
| `telemetry/d6_recover35…/recover_receipt.json` (Stage P recovery; corpus\_identity)                                              | `b7a51cd02927ad06` |
| `telemetry/d6_score35…052821…/score_receipt.json` (Stage P capture 128/128)                                                      | `fe6ab66f0f423280` |
| `telemetry/d6_score35…083625…/score_receipt.json` (Stage P score 40/40)                                                          | `8a862531af11c9d2` |
| `telemetry/d6_slim35…/slim_receipt.json` (Stage P prune-apply m25p)                                                              | `a21706cd9fc5b13d` |
| `INFRA/…/architect_d6_matched/v1/manifest.json` (Stage C identity: draw `3d1d53a5…` ×3 byte-identical + calibration `6e745e6c…`) | `5ce66c889afe879e` |

The Stage R manifest binds the H3 control at `ec9b664b…` (the P3 frozen `p3_results.jsonl`, verified at preflight) and the matched arm at `10c5a14b…`, the pairing chain of custody is closed end to end. Stages M and H5 produced no receipts because they were not run; the envelope covers every artifact the experiment actually produced.

## The frozen pre-registration, verbatim

Reproduced exactly as sealed (the bytes the self-hash above is computed over). The strikethrough passages in §3 (R1) and §4.1 are the DEV-1 / DEV-2 / DEV-3 corrigenda, preserved as the record carries them, the minority-corpus source ruled three times before any data, landing at 100:0 architect-primary.

```md
---
doc_class: attestation
status: FROZEN
status_as_of: 2026-07-12
frozen_at: "2026-07-12 (SHA-256 canonical-prefix self-hash via /sidecar Mode C at seal moment; hex lives in the .sha256 sidecar + experiment marker — NEVER inlined into this body per the 2026-04-18 Gate-D 815e0a35 anti-propagation lesson)"
verified_against_code_on: 2026-07-12
verified_via: |
  Authored 2026-07-12 at CYC41 P3 [D6] step 5 via /experiment Mode A, post-P7 seal
  (27602b5). Grounded in: the P7 [D4] blend proposal (fresh-draw receipts 45,967
  rows; sandboxed), the sealed sovereign draw v20260712 (sha at-claim matched by
  profile_compiler this session), the P3 FINAL_SEAL mismatch-control rows (frozen),
  and the realignment ruling 4a3508c16d7642a9 + sequence eb60030e2e454a7b.
  Hypotheses H1-H5 + Rulings R1-R3 await operator ratification BEFORE seal
  (§16 carries no hash until ratified).
signed_at: "2026-07-12T09:45:00Z"
signed_by: "Claude Fable 5 (claude-fable-5) + Jamey Kistner"
authors:
  - Jamey Kistner
  - Claude Fable 5 (claude-fable-5)
  - Claude Opus 4.8 (web collaborator — H3 design review, operator-relayed 2026-07-12)
bid_relevance: high
related_docs:
  - RESEARCH/Roadmaps/ROADMAP_CYCLE41_SOVEREIGN_BIG_MODEL_COMPRESSION.md
  - RESEARCH/SPECS/BLEND_PROPOSAL_CYC41_ARCHITECT_CORPUS.md
  - RESEARCH/SPECS/RESULTS/cyc41_p3_35b_prune_drafter/PRE_REGISTRATION.md
  - RESEARCH/SPECS/RESULTS/cyc41_p3_35b_prune_drafter/FINAL_SEAL.md
  - REFINERY_MEMORY_SERVICE/SRC/corpus_engine/profile_compiler.py
phase_anchor: "CYC41 P3 [D6] step 5 — matched-corpus compression chain re-run (the answer arm to the pre-registered mismatch control). Experiment-id osint-suite-dev-2026-07-12-7e42."
arc_status: active
changelog:
  - {date: "2026-07-12", change: "Authored at CYC41 P3 [D6] step 5 via /experiment Mode A (DRAFT). H1-H5 + Rulings R1-R3 pending operator ratification; seal follows ratification."}
  - {date: "2026-07-12", change: "Collaborator design review entered into the record (Claude Opus 4.8, web; operator-relayed): H3 either-bank-uncorrected disjunction KILLED (primary bank = GPQA-198 pre-registered a priori; pooled math confirmatory) + multiplicative-vs-directional interpretation boundary pre-registered + R1 minority-source doctrinal-leak tension named + corpus-identity clause extended to hash the POST-blend training bytes. Full narrative in decisions.md."}
  - {date: "2026-07-12", change: "H1-H5 + R1-R3 + §9 bars operator-ratified pre-data (verbatim 'Proceed. Excellent work.', post-review form). R1 = 90:10 D3-mix-reasoning minority (tension standing; pre-data Mode C flip to sovereign-general available until Stage C). status DRAFT→FROZEN; canonical-prefix self-hash sealed via /sidecar Mode C (hex in .sha256 sidecar + experiment marker); freeze-gate armed."}
  - {date: "2026-07-12", change: "DEV-1 (Mode C, PRE-DATA, AMENDMENT_1): R1 minority source FLIPPED D3-mix → SOVEREIGN-GENERAL (operator structured-choice ruling at the collaborator-flagged decision point; the pre-registered escape valve exercised before any Stage C bytes). §3 R1 strikethrough-corrigendum; pre-amendment self-hash preserved in AMENDMENT_1 §4; post-corrigendum hash recomputed at DEV-1 resolution."}
  - {date: "2026-07-12", change: "DEV-2 (Mode C, PRE-DATA, AMENDMENT_2): R1 minority category set WIDENED to all outside-selector sovereign conversation categories (knowledge+decision+feedback+tool_summary) — the Stage C ruling-vs-bytes tripwire (collaborator forward-flag #3) found conversation_reasoning EMPTY in the sealed v20260712 bytes; operator structured-choice 'Widen categories' preserves the ratified 90:10 + 100%-sovereign properties. Pre-amendment self-hash preserved in AMENDMENT_2 §4; post-corrigendum hash recomputed at DEV-2 resolution."}
  - {date: "2026-07-12", change: "DEV-3 (Mode C, PRE-DATA, AMENDMENT_3 — supersedes DEV-2): R1 minority DROPPED → 100:0 architect-primary. The dedup census falsified the minority premise (conversation multi-tagging → 61 unique rows vs 5,107 required; only codebase-docs at scale = wrong forgetting mode). H1/H2 bars carry the forgetting guard by measurement; H3 contrast pristine. Operator + collaborator conferred; operator 'Confirmed and proceed' on an explicit stop-restate-ask checkpoint (the DEV-2 §1.2 violation lesson applied). §3 R1 + §4.1 corrigenda; pre-amendment self-hash preserved in AMENDMENT_3 §4."}
---

# PRE_REGISTRATION — CYC41 D6: Matched-Corpus Compression Chain (sovereign draw at every layer)

> **Companion docs:** [`ROADMAP_CYCLE41`](../../../Roadmaps/ROADMAP_CYCLE41_SOVEREIGN_BIG_MODEL_COMPRESSION.md) §1 Phase 3 [D6] + §2 ADDENDUM (realignment) · [`BLEND_PROPOSAL`](../../BLEND_PROPOSAL_CYC41_ARCHITECT_CORPUS.md) (Rulings input) · [`P3 PRE_REG`](../cyc41_p3_35b_prune_drafter/PRE_REGISTRATION.md) + [`P3 FINAL_SEAL`](../cyc41_p3_35b_prune_drafter/FINAL_SEAL.md) (the frozen MISMATCH CONTROL this experiment answers) · [`EXPERIMENT_LEDGER.md`](EXPERIMENT_LEDGER.md) · [`decisions.md`](decisions.md)
> **Reading discipline:** DRAFT until operator-ratified; immutable once sealed per §11. Post-seal changes ONLY via /experiment Mode C. The `.sha256` sidecar is the tamper-evidence envelope (canonical-prefix per `seal_prereg.py`).
> **Spec authority:** `~/.claude/specs/EXPERIMENT_DISCIPLINE.md` v1.1 + `~/.claude/specs/SOVEREIGN_OPERATIONS_SPEC.md` §1.1/§4.3/§4.6/§4.7/§X.
> **Rules:** [`attestation-class-files`](../../../../.claude/rules/attestation-class-files.md) · [`brain-query-first`](../../../../.claude/rules/brain-query-first.md) · [`phase-claim-discipline`](../../../../.claude/rules/phase-claim-discipline.md) · [`skill-enforcement`](../../../../.claude/rules/skill-enforcement.md)

> **Ground-Truth Attestation**
> - **Doc class:** attestation (pre-registration; SHA-256 canonical-prefix self-hash at seal)
> - **Status as of:** 2026-07-12 — FROZEN (H1-H5 + R1-R3 operator-ratified pre-data, verbatim "Proceed. Excellent work."; sealed)
> - **Verifier:** Claude Fable 5 (claude-fable-5) + Jamey Kistner
> - **Ledger:** [`.claude/verified/RESEARCH__SPECS__RESULTS__cyc41_d6_matched_corpus_chain__PRE_REGISTRATION.verified.json`](../../../../.claude/verified/RESEARCH__SPECS__RESULTS__cyc41_d6_matched_corpus_chain__PRE_REGISTRATION.verified.json) + this file's `.sha256` sidecar at seal
> - **Scope:** Pre-registers the corpus-identity clause, blend/shaping/negatives rulings, matched-arm chain (governed draw → prune attribution + recovery-FT → imatrix/quant → retention → drafter row), hypotheses, banks, thresholds, frozen set, deviation policy.
> - **Re-verify cadence:** IMMUTABLE mid-experiment per §11 (once sealed).

---

## 1. Abstract

The realignment ruling (`4a3508c16d7642a9`) established the doctrine: for the 35B-and-below sovereign slots, **one sha-sealed sovereign corpus draw is cited at EVERY flywheel layer** (prune calibration → recovery-FT → imatrix → SSD → persona-LoRA → drafter head). P3 (`osint-suite-dev-2026-07-06-764b`, FINAL-SEALED) is now the complete pre-registered **MISMATCH CONTROL**: every layer measured on D3-mix, frozen BEFORE this matched arm existed. This experiment runs the **MATCHED arm**: the same MoE-Slimming chain (same method, same banks, same binary, same bars) with the sovereign architect blend substituted at every corpus-touching layer. The falsifiable bet (Less-is-MoE prediction, our own paired evidence: nothink table MATH500 −8.0 n.s. vs GPQA −24.7/AIME −53.3 cratering on domain-thin banks): **matched corpus improves the compression chain's quality retention over the mismatched control**. Verdict = H1 ∧ H2 ∧ H3. The resident-serving gate (P3 H3's FAIL class — WDDM spill) is carried as the DEPLOYMENT gate, mechanically separate from the corpus claim (§13.2 — corpus match cannot fix a VRAM spill). SSD/persona-LoRA/DFlash layers ride P4/P5 but are BOUND to this experiment's corpus identity (§5).

## 2. Hypotheses (operator-ratified 2026-07-12 pre-data, verbatim "Proceed. Excellent work." — post-collaborator-review form: H3 primary-bank design + interpretation boundary in place at ratification)

| ID | Statement | Primary statistic | Decision rule | Anchor |
|---|---|---|---|---|
| **H1 — math/reasoning retention (deployment framing)** | Matched pruned-35B (+matched recovery-FT, 4-bit) retains generative math vs deployed `[architect-apex]` UD-Q5_K_XL | Pooled AIME24 (30) + MATH-500 (50 stratified) = 80 paired items; Δ = acc(matched) − acc(baseline) | **PASS** Δ ≥ −5.0 pp · **FAIL** Δ < −5.0 AND McNemar p < 0.05 · else **INCONCLUSIVE** → operator | Inherited verbatim from P3 H1 (comparability) |
| **H2 — knowledge retention (deployment framing)** | Matched pruned-35B retains knowledge-MC vs the same baseline | GPQA-diamond n=198 paired; Δ = acc(matched) − acc(baseline) | **PASS** Δ ≥ −6.0 pp · **FAIL** Δ < −6.0 AND p < 0.05 · else **INCONCLUSIVE** | Inherited verbatim from P3 H2 |
| **H3 — matched-vs-mismatched corpus-match direction (THE bet)** | The matched arm beats the frozen mismatched control at matched ratio (same banks, seed, binary, thinking mode) | Δm = acc(matched) − acc(mismatched frozen cell), paired per-item vs the P3 frozen JSONL. **PRIMARY BANK = GPQA-diamond n=198** (pre-registered a priori, collaborator review 2026-07-12: mismatch damage is domain-structured — GPQA −24.7 pp in the frozen nothink table, the domain-thin bank where corpus-match has room to move; also the best-powered paired bank). Pooled math (80) = **CONFIRMATORY** (reported, one-sided, Holm-adjusted; carries no independent pass authority) | **PASS** Δm > 0 on GPQA AND one-sided McNemar p < 0.05 on GPQA, PROVIDED pooled-math Δm is not significantly negative (two-sided p < 0.05) — a significant math regression downgrades to INCONCLUSIVE → operator · **FAIL** Δm ≤ 0 on GPQA · else **INCONCLUSIVE** | The realignment bet made falsifiable; either-bank-uncorrected disjunction removed (≈10% family-wise FP surface, collaborator flag); disconfirm anchor AIMER 2603.18492 stands |
| **H4 — drafter-layer row (report + directional)** | Matched trunk drafts ≥ the mismatched trunk (Stage M banked: τ 1.86, accept 62.7%) | τ (mean accepted length/verify step) + accept% on the H-serve cell, §4.3 battery, same binary | **CONFIRMED** if τ(matched) ≥ 1.86 AND accept ≥ 62.7% · else **DRIFT-NEGATIVE** (documented; no phase-gate weight) | Stage M mismatch row = the drafter-layer control (AMENDMENT_8 lineage) |
| **H5 — persona-domain evidence table (report-graded)** | The matched arm shows persona-domain capability the control never trained toward | Architect held-out test split (governed draw, §4.2b): contract-parse rate + janitor-graded correctness on a 100-item stratified sample, matched arm vs baseline vs mismatched cell | Report table only — NO gate weight v1 (shaping maturity per R2; parse-rate on raw-passthrough tranche is descriptive) | First persona bank; honest v1 |

**Verdict rule (frozen at seal):** PASS = H1 ∧ H2 ∧ H3. Any FAIL → documented, incumbent stays, doctrine re-examined (an H3 FAIL **falsifies the matched-corpus direction bet** — that is the point of pre-registering it). **DEPLOYMENT gate (separate, mechanical):** cutover additionally requires the P3 H3-class resident-serve receipt (`-ngl 99`, no `--n-cpu-moe`, no WDDM spill, ≥ 63.8 tok/s) — reported as SERVE-PASS/SERVE-FAIL, orthogonal to the corpus verdict (§13.2).

**Interpretation boundary (pre-registered; collaborator review 2026-07-12):** an H3 PASS confirms the corpus-match DIRECTION at whole-chain level. It does **NOT** by itself isolate multiplicative-vs-additive compounding — a main effect of "sovereign corpus is better data" also passes H3, and the binary comparison cannot distinguish the two. The multiplicativity claim (§5.34 class) requires per-layer decomposition — matched-vs-mismatched receipts at EACH layer (prune, recovery, imatrix, SSD, LoRA, drafter), which the corpus-identity clause makes possible and which is DEFERRED to the P4/P5 layer receipts citing this experiment. A PASS is claimed as directional, never as the multiplicativity result. (The P3 GPQA "win" over-claim class, one level up.)

## 3. Rulings R1-R3 (operator-ratified 2026-07-12 pre-data, verbatim "Proceed. Excellent work." on the recommendation column as presented; values frozen at seal)

| Ruling | Question | Recommendation (blend proposal §3-§5) | **RATIFIED VALUE** |
|---|---|---|---|
| **R1 — blend ratio + minority source** | architect-primary : general-reasoning minority; minority source | 90:10; minority = D3-mix reasoning subset (anti-forgetting slice; doctrinally distinct from D3-mix-as-primary). **⚠ Collaborator tension (Opus 4.8, pre-registered):** a D3-mix minority re-introduces 10% of the CONTRASTED corpus into the matched arm — the arms then differ in match *minus a 10% leak*, and the matched arm is not 100% sovereign in the very experiment meant to prove the sovereign doctrine. Cleanliness-preferred alternative = **sovereign-general minority** (conversation `reasoning`/`knowledge` categories OUTSIDE the architect selector — blend proposal option (b)). Operator rules CONSCIOUSLY between contrast-cleanliness (sovereign minority) and anti-forgetting strength (D3-mix minority); whichever is ruled, the tension stands pre-registered | ~~**90:10; minority = D3-mix reasoning subset** (ratified 2026-07-12 on the recommendation as presented, tension standing; a pre-data flip to sovereign-general minority remains available via Mode C amendment BEFORE Stage C — after Stage C the ruling is frozen)~~ **[corrigendum DEV-1, AMENDMENT_1, 2026-07-12 PRE-DATA: 90:10; minority = SOVEREIGN-GENERAL — conversation reasoning/knowledge categories OUTSIDE the architect selector, cut from the SAME sealed v20260712 draw. Operator structured-choice ruling at the collaborator-flagged decision point; the pre-registered escape valve exercised exactly as written, before any Stage C bytes existed. Arms now differ ONLY in corpus match (zero D3-mix leak; matched arm 100% sovereign); H1/H2 bars = the honest forgetting guard; the Stage C HF-fetch dependency is eliminated.]** **[corrigendum DEV-2, AMENDMENT_2, 2026-07-12 PRE-DATA: minority category set WIDENED to ALL outside-selector sovereign conversation categories (~~reasoning/knowledge~~ → `knowledge`+`decision`+`feedback`+`tool_summary`) — pre-Stage-C ruling-vs-bytes reconciliation found `conversation_reasoning` EMPTY in the sealed v20260712 bytes (0 rows; declared-but-unpopulated), capping the named minority at ~5.5%. Operator structured-choice "Widen categories" 2026-07-12: deterministic cut of the widened pool (7,620 available) to hit the ratified 90:10 EXACTLY. Sovereignty property unchanged (100% sovereign, zero D3-mix); ratio unchanged; only the category enumeration corrected to match the sealed bytes.]** **[corrigendum DEV-3, AMENDMENT_3, 2026-07-12 PRE-DATA — SUPERSEDES the DEV-2 widening: ~~90:10~~ → **100:0 architect-primary; the minority is DROPPED**. DEV-1's sovereign-general minority premise was falsified at Stage C by the dedup census: the conversation harvester multi-tags one pair under several categories, so the outside-selector conversation pool is ~98% byte-redundant with the primary (61 unique rows vs 5,107 required at 90:10), and the only at-scale sovereign pool (codebase-docs, 30,537) guards the WRONG forgetting mode (codebase knowledge the primary already saturates, not reasoning/instruction-following). Minority dropped; **H1/H2 retention bars carry the forgetting guard by measurement** (inherited verbatim; catastrophic forgetting registers as FAIL). The H3 contrast is thereby pristine — the arms differ in exactly one thing, corpus match, with no minority-composition confound. Pre-registered caveat: had a prior round shown THIS cut forgets instruction-following, the ruling would flip to a memory-pool minority (1,251 rows) at whatever ratio the pool supports; no such evidence exists. Operator + collaborator conferred; operator verbatim "Confirmed and proceed. Peer-reviewed diligence is the bar."]** |
| **R2 — contract-pair shaping method** | rule-templating vs janitor-assisted vs two-tranche | two-tranche contract+raw (contract tranche = SAT trail + engine-loop accrual + templated exemplars; raw tranche = passthrough volume) | **two-tranche contract+raw** (ratified 2026-07-12) |
| **R3 — negatives floor at 45k scale** | how the 0.25 floor applies | floor applies to the CONTRACT tranche only; raw tranche = substrate; engine-loop reject verdicts accrue as labeled negatives | **contract-tranche-only floor** (ratified 2026-07-12) |

## 4. Materials + variables

### 4.1 Corpus identity (the clause — every layer cites BOTH hashes)

| Identity | Value |
|---|---|
| Sovereign merged set | `DATA/DOCTRINE/training_sets/sovereign/v20260712/sovereign_pairs.jsonl` — 97,002 pairs, sha256 `321a18f48e4c5fd50384614a7c1d729b1d94fa440f4a6fecba61e5c9add3a954` (sidecar at-claim; `corpus.draw_seal` ledger seq 149) |
| Architect training draw | CUT AT PREFLIGHT G2 via the governed chain (`compile.propose → compile.approve HUMAN → compile.dispatch`) with `profile_compiler` under R1-R3; its `draw_sha256` + manifest recorded in `decisions.md` + the `corpus.compile` ledger record AT cut — NEVER propagated symbolically into this body |
| **Blended training draw — THE cited bytes (collaborator rider, Opus 4.8)** | The Stage C output: POST-blend, POST-shaping training set (including the ~~R1 minority tranche +~~ R2 tranche structure + R3 negatives) — sha256-sidecar'd AT assembly. **This hash is what every training layer cites** — hashing only the pre-blend sovereign draw would prove identity of the wrong artifact (clause technically-present-but-hollow). Lineage recorded alongside it: {merged-set sha above, architect compile draw_sha256~~, minority-source manifest~~} **[corrigendum DEV-3, 2026-07-12 PRE-DATA: R1 = 100:0 — no minority tranche, no minority manifest; the cited bytes = the SHAPED architect-primary draw (R2 two-tranche structure incl. engine-loop contract renders + R3 negatives — still distinct bytes from the raw G2 cut, so the clause keeps its force); lineage = {merged-set sha, architect compile draw_sha256}]** |
| Layer bindings | prune attribution/calibration text · recovery-FT set · imatrix calibration text · H5 persona bank · (P4/P5: SSD tokens + persona-LoRA set + any A3 distillation set) — each stage manifest records the BLENDED-draw hash + its lineage pair above |

### 4.2 Arms / banks

(a) **General banks + sampling inherited VERBATIM from P3 PRE_REG §4.2** (AIME24 30 · MATH-500 50 stratified seed 20260422 · GPQA-diamond 198; T=0.6/top_p 0.95/top_k 20/seed 20260422; budgets 12288/6144/4096; extractors + janitor fallback; 8088 eval serving; SynthLock; per-item JSONL + resume; thinking-ON primary per P3 frozen verdicts, nothink descriptive). Baseline arm = deployed `[architect-apex]` UD-Q5_K_XL. Mismatched-control per-item data = the P3 FROZEN run artifacts (no re-run of the control — its seal is its evidentiary force). (b) **Persona bank (H5):** the governed draw's `test.jsonl` split; 100-item stratified sample (deterministic seed 20260422) for janitor-graded correctness; parse-rate over the full split. (c) **Drafter battery inherited from P3 §4.3** (8 prompts, T=0, 256 tok; τ from server timings).

### 4.3 Method + artifacts

MoE-Slimming chain per P3 (method already proven on-box: slim25p/slim31p checkpoints exist): matched arm re-runs **attribution + prune + recovery-FT** with the R1 blend at the ratio matching the strongest P3 cell (25%; 30/33% contingency pre-authorized per P3), then convert → **imatrix on calibration text rendered FROM the governed draw** → Q4_K_M (MTP tensors Q8_0-pinned). P3's FROZEN drivers embed D3-mix in `build_corpus()` (`run_p3_score35.py:54,84` · `run_p3_recover35.py:77-83,146`) — **forked drivers** load the sha-sealed LOCAL draw + HARD-FAIL on hash drift; each fork is amendment-co-sealed into `frozen_paths` BEFORE the stage it governs (P3 §5 mechanics). Same b9789 binary all llama.cpp arms.

## 5. Frozen set / gate discipline

`frozen_paths` at seal: **this `PRE_REGISTRATION.md` only** (forks do not exist yet; added via amendment co-seals). The corpus-identity clause EXTENDS beyond this run: P4 (persona-LoRA/SSD) and P5 (drafter) stage manifests MUST cite the §4.1 identity pair — a P4/P5 layer citing a different draw is a deviation against THIS experiment (Mode C required). Run outputs never frozen. Cold-start per EXPERIMENT_DISCIPLINE §3.2.

## 6. Analysis plan + post-hoc policy

1. H1/H2: paired McNemar exact two-sided + Wilson 95% CI (P3 §6.1 verbatim).
2. H3: per-item pairing vs the P3 frozen JSONL (same item ids); one-sided McNemar (directional bet); Δm per bank + pooled.
3. H4: τ + accept% vs the banked Stage M row; descriptive per-prompt variance.
4. H5: parse-rate full-split + 100-item janitor correctness (janitor never sees arm identity); table only.
5. **Post-hoc policy:** anything unnamed = `EXPLORATORY (post-hoc)`, cannot move the verdict. Pre-authorized descriptives: truncation rate, thinking-length, per-domain GPQA, blend-minority ablation notes.
6. No mid-run peeking; no accuracy aggregation until an arm completes.

## 7. Procedure (stages; SynthLock + receipts + telemetry per stage)

1. **Stage G (preflight):** G1 Rulings ratified → G2 governed draw cut (HUMAN approve; draw sealed + ledgered) → G3 minority-source assembly + §1.1 re-check (D3-mix components license-cleared CYC41 P1) → G4 disk/VRAM budget receipt → G5 driver forks authored + amendment-co-sealed. Each outcome in `decisions.md`.
2. **Stage C (corpus):** blend assembly per R1-R3 → training set + rendered calibration text, both hash-sidecar'd + manifest citing §4.1.
3. **Stage P (prune):** matched attribution + channel-prune + recovery-FT → checkpoints + manifests.
4. **Stage Q (quant):** convert → matched imatrix → Q4_K_M (MTP Q8_0-pinned) → serve receipts on 8088 (SERVE row).
5. **Stage R (retention):** paired eval per §4.2 → H1/H2/H3.
6. **Stage M (drafter):** τ battery on the served cell → H4.
7. **Post-run:** Mode D — POST_RUN_INTEGRITY + FINAL_SEAL + verdict; roadmap ticks; brain finding; de-register.

## 8. Sample-size rationale

Inherited from P3/D7 (n=198 GPQA ~70-80% power for 7-8 pp; pooled-math n=80 point-estimate with McNemar guarding FAIL; single-seed primary, 3-seed sensitivity pre-authorized EXPLORATORY on INCONCLUSIVE). H3 pairing reuses the SAME frozen items — no new power cost. H5 100-item sample is evidence-table-grade only.

## 9. Thresholds (frozen at seal)

| Gate | Bar |
|---|---|
| H1 | Δ pooled-math ≥ −5.0 pp vs deployed baseline (FAIL needs Δ < −5.0 AND p < 0.05) |
| H2 | Δ GPQA-198 ≥ −6.0 pp (FAIL needs Δ < −6.0 AND p < 0.05) |
| H3 | PRIMARY = GPQA-198: Δm > 0 AND one-sided p < 0.05; pooled math confirmatory (Holm; sig-negative math → INCONCLUSIVE) · FAIL: Δm ≤ 0 on GPQA |
| H4 | CONFIRMED iff τ ≥ 1.86 AND accept ≥ 62.7% (report row) |
| H5 | none (report-graded) |
| Verdict | PASS = H1 ∧ H2 ∧ H3 · SERVE gate separate (63.8 tok/s resident, mechanical) |

## 10. Blinding / bias notes

No human grading; janitor never sees arm identity; identical prompts/seeds/binary; fresh server process per arm; mismatched control is FROZEN (cannot be re-run toward a preferred outcome); no mid-run peeking.

## 11. Deviations policy

Immutable once sealed. `/experiment --deviate osint-suite-dev-2026-07-12-7e42` → `decisions.md` narrative → ledger row 🔧 → strikethrough-corrigendum → ✅ resolve. Post-data adjudications in `## 99. Corrigendum` below the seal. Document, don't silently patch.

## 12. Wall-clock + infra plan

Stage C: minutes (compile ~3s sandboxed receipt; render + hashing minutes). Stage P: hours-to-a-day per cell (P3 receipts). Stage Q: 1-3h. Stage R: ~3-6h/arm. Disk: 885.3 GB free (post Tier B/C reclaim) vs ~250-350 GB transient peak — clears. All GPU work under SynthLock; eval on 8088; PROD ports untouched.

## 13. Pre-registered deviations from precedent

1. **Recovery-FT + attribution + imatrix corpus = the R1 sovereign blend** (the entire point) — all other method knobs held at P3 values for comparability.
2. **The SERVE gate is separated from the corpus verdict.** P3 H3 FAILED on WDDM spill — a serving-mechanics class a corpus swap cannot address. Folding it into this verdict would let an orthogonal failure mask the corpus finding (or vice versa). Cutover still requires SERVE-PASS.
3. **Mismatched control is not re-run** — its frozen per-item artifacts ARE the comparison arm (pre-registration timestamping is its evidentiary force).
4. **H5 persona bank is report-graded v1** — contract validity is not yet enforced on the raw tranche (R2 maturity); gating on it would manufacture a verdict from an un-ruled seam.
5. **H3 primary-bank design replaces the either-bank disjunction** (collaborator review, Opus 4.8, 2026-07-12): "pass on either bank, uncorrected" carried ≈10% family-wise false-positive on the cycle's load-bearing hypothesis. GPQA-198 pre-registered as primary on domain-structure reasoning (§2); pooled math confirmatory only.
6. **H3 PASS is claimed as DIRECTIONAL, never multiplicative** — interpretation boundary pre-registered in §2; the multiplicativity decomposition rides P4/P5 per-layer receipts.

## 14. Limitations (honest scope)

1. H3 attributes quality deltas to corpus match; recovery-FT stochasticity is a confound (single-seed; ratio + method held constant mitigates but does not eliminate).
2. The persona bank derives from the same distribution the matched arm trains on — held-out split guards leakage at the item level, not the distribution level (named, not hidden).
3. Single-operator, single-box; deployment-retention framing (baseline carries Q5 damage) per P3 §13.1.
4. SSD/LoRA/drafter-head layers are identity-BOUND here but run in P4/P5 — their receipts land in those phases, citing this experiment.

## 15. Cross-references

- Experiment-id: `osint-suite-dev-2026-07-12-7e42` · marker `~/.claude/hooks/experiments/osint-suite-dev-2026-07-12-7e42.experiment.json` (written at seal) · registry `~/.claude/hooks/experiments/REGISTRY.json`
- Brain anchors: realignment `4a3508c16d7642a9` · sequence `eb60030e2e454a7b` · D3 draw seal `781c7a4e57a341b0` · P7 seal `194130a92c224025` · D4 mechanics `1f22e346663b41c0`
- Precedents: P3 PRE_REG + FINAL_SEAL (mismatch control) · D7 PRE_REG (banks/stats lineage) · BLEND_PROPOSAL (Rulings input)

## 16. Seal

Seal state: **SEALED 2026-07-12** — §2 hypotheses H1-H5 + §3 Rulings R1-R3 + §9 bars operator-ratified pre-data (verbatim "Proceed. Excellent work.", on the post-collaborator-review form). The canonical-prefix self-hash (LF-norm bytes `0..^## 16. Seal`, rstrip+`\n`, SHA-256 per `seal_prereg.py`) was computed via `Skill(/sidecar --hash-first)` at this moment; the hex lives in the `.sha256` sidecar beside this file + the experiment marker `~/.claude/hooks/experiments/osint-suite-dev-2026-07-12-7e42.experiment.json` — NEVER inlined here (Gate-D `815e0a35` anti-propagation). Frozen content = everything above the `## 16. Seal` line. Freeze-gate armed on `PRE_REGISTRATION.md` (driver forks added via amendment co-seals per §5). Post-data adjudications land in `## 99. Corrigendum` below this section.
```

***

*Evidence & seals · Matched-Corpus Compression Chain (D6) · experiment `osint-suite-dev-2026-07-12-7e42` · pre-registration reproduced verbatim from the sealed record; self-hash quoted from its sidecar · CC BY 4.0 · © Jamey Kistner, OSINTelligence LLC*