v6 is the only forward development line, and it has a stated purpose. The values are the contract: sovereignty under European jurisdiction, alignment to EU constitutional values, honesty, open-weight integrity, and capability without compromise. The goal follows from them: a sovereign European alternative that competes with the frontier, with the explicit target of beating Sonnet 5 on PACE while staying open-weight and honest about its lineage. Below, what is proven today, what is planned, and what remains open research — kept distinct.
Status legend: Proven ships today · Planned designed, not yet run · Open research gated by a falsification pilot.
v6 is a single program with one thesis: that a dense policy of roughly nine billion parameters, trained under a verifier-first capability foundry and promoted only on verified evidence, can compete with frontier-lab systems on agentic and coding work — while remaining open-weight and sovereign. The values it runs under are the site's contract: sovereignty, EU constitutional alignment, honesty, open-weight integrity, and capability without compromise. The canonical artifact is one shared dense ~9B policy — not an ensemble, and not a quietly swapped 27B substitute.
The goal is concrete and falsifiable: beat Sonnet 5 on four PACE projections plus ten of twelve capability families, and beat comparable open models (Ornith 9B, Qwen 27B) wherever the comparison is valid. Architecture alternatives such as Qwen 3.6 27B/35B are matched bakeoffs, never silent replacements.
The mechanism for competing above weight class is not scale but operating discipline. The supporting evidence is not yet ours — it is the observation, from a multi-agent panel session, that under identical harnesses the strongest orchestrator beat a larger model purely on behavioral deltas: questioning representations earlier, choosing discriminating experiments, converting observations to code sooner. The bet is that such discipline, internalized into weights as reflexes rather than bolted on as prompt scaffolding, is the variable a small sovereign model can move.
Every capability lane generates, verifies independently, repairs, splits disjointly, and only then measures student lift. Promotion runs from receipts, never from chat state. Sovereignty is operational: the foundry runs on the project's own gateway with no external API dependency, and every source enters weights only through rights and contamination receipts.
The pipeline runs from an exact preserved pre-alignment parent in the Eugeny 5.1 lineage, through a small late-CPT pilot, replay of the SFT and SimPO behavioral program against frozen gates, and then the combined verified agent-RL and instinct stages, before consolidation, vision regraft, and a dense release.
The current promoted Eugeny 5.1 Vision model remains the production baseline, comparison teacher, and rollback until v6 is promoted. Nothing below is a trained result unless marked otherwise.
An exact preserved checkpoint in the Eugeny 5.1 lineage. The preserved parent is the source of truth for every later stage; nothing is trained against a moving baseline.
A small continued pre-training (CPT) pilot on rights- and contamination-cleared sources. Admission fails closed when provenance is unknown.
The behavioral program — supervised fine-tuning and SimPO preference optimization — replayed against frozen qualification gates, so promotion is evaluated against a fixed bar, not a moving one.
The combined verified agent-RL and instinct stages. This is where the two Phase-5 tracks — process awareness and mechanism discovery — live. Specialist adapters are never silently merged; verified examples are replayed into one shared dense checkpoint through a single consolidation route after the unified retention gate.
A single consolidation route writes retained examples from every adapter lane back into one shared dense checkpoint. There is one canonical dense policy, not a federation of adapters.
The native visual tower and processor contract are grafted onto the v6 dense policy, non-destructively and reversibly — the same contract that ships on Eugeny 5.1 Vision today. Multi-image vision remains unverified; the qualified modality contract is one image and a narrow two-frame video gate.
v6 is promoted as the dense release only on completion of the retention gate. Until then the dense v5.1 policy at 262K remains the served baseline, the comparison teacher, and the rollback.
Architecture alternatives (Qwen 3.6 27B/35B) are matched bakeoffs, evaluated on the same gates — never silent replacements for the ~9B canonical policy.
Two tracks run inside Phase 5. Process awareness internalizes orchestrator discipline as weight-level reflexes that survive prompt ablation. Mechanism discovery asks whether training on self-generated, certified episodes can make a small policy vertically superior at reverse-engineering novel interactive environments into backtestable world models.
These are the goal paths: the route by which a small dense policy competes with the frontier on behavioral discipline rather than on scale. Both are in development; neither is a trained result. The status of each is stated on its column.
A small model internalizes orchestrator discipline as weight-level reflexes that survive prompt ablation. Eight behavioral families, each with a structural ground-truth signal and a named failure mode it replaces.
Status: a canonical track plan, designed and ready for a Phase-0 pilot — not a trained result.
Self-generated, certified episodes that reverse-engineer novel interactive environments into backtestable world models. The most ambitious track — and the one gated hardest.
Status: open research, gated by the falsification pilot. No commitment precedes its result.
Process awareness decomposes orchestrator discipline into eight behavioral families. The grounding insight is that under an identical harness, behavioral discipline — not raw scale — separated the best orchestrator. Each family is trained as a weight-level reflex against structural ground truth, with a named failure mode it replaces.
| Family | Ground truth | Failure it replaces | The reflex |
|---|---|---|---|
| 1. Status reporting | Journal phase and agent counts | Guessing "it's probably still running" | Reads state before reporting it |
| 2. Agentic time estimation | Workflow DAG plus elapsed time | Human-time intuition ("maybe an hour?") | Decomposes parallel phases (max) from sequential barriers (sum); returns ranges with the bottleneck named |
| 3. Research → Plan → Delegate | Question complexity and live evidence state | Answering from memory; panel-everything or answer-everything-solo | Calibrates orchestration depth to the actual question |
| 4. Progressive correction | Distance between the user's mental model and the truth | Dumping the whole answer; an unstructured "actually…" | Names what is right, then the specific error, as an investment in the next question |
| 5. Sub-agent design | Role coverage, phase architecture, schema quality | A missing skeptic, unstructured output, or prescribed conclusions | Designs panels with adversarial coverage |
| 6. Epistemic hygiene | Source dates, provenance, and verification status | Undated claims, single-source assertions, memory treated as evidence | Dates and searches when uncertain, with bounded freshness windows |
| 7. Resource reasoning | GPU-hours, parallelism, and sequential barriers | Dollar estimates without decomposition | Decomposes cost and names fallbacks |
| 8. Serendipity flagging | A contradiction between a sub-result and an upstream assumption | Waiting for the final report and missing lateral signal | Flags the implication inline |
Promotion for each family requires the reflex to survive scaffold ablation on workflows outside the training distribution. A narrated behavior that was never executed under the judge is rejected at admission.
A mechanism-discovery episode is admitted only when it is certified. Certification is structural, not narrative — and the track itself is gated by a pre-registered falsification pilot that runs before any Phase-2 compute.
HiLS-Attention is end-to-end learned chunk-wise sparse attention. Each query attends to a local window plus the top-k retrieved distant chunks, at constant cost. A landmark token per chunk produces a learnable summary key; chunk mass is estimated by a learnable first-order Taylor surrogate of the true LogSumExp mass — the principled fix for the prior finding that parameter-free and mean-pooled selectors returned 0% needle recall. Attention is factorized into an inter-chunk softmax (allocating mass across chunks) and an intra-chunk softmax; because the surrogate parameterizes the forward pass, the LM loss directly trains chunk selection, with HoPE positional encoding for length extrapolation.
The dense 262,144-token envelope ships today. Planned HiLS is research, not a verified product capability.
The figures above are from the HiLS paper (arXiv:2607.02980), not from a Eugeny measurement. Our integration is the custom part: apply HiLS to the eight full-attention layers of the hybrid ~9B backbone and leave the twenty-four GatedDeltaNet layers untouched. That modeling file does not exist upstream and is the main engineering task.
Conciseness in v6 is architectural, not stylistic. One principle — the shortest correct path is the goal — is applied at three layers that compound. A model trained on the direct path, decision-critical content kept and noise removed, behaves as if it already knew what to do.
Teacher chain-of-thought is condensed before it reaches the student. The rewriter keeps decision-critical structure — the state-reading action, the DAG decomposition, dated claims, serendipity flags — and removes exploration, backtracking, and theatrical filler. Target: 40–60% shorter traces.
Short correct paths are reinforced through the Long2Short class of training: preference pairs and outcome RL whose target is the condensed direct path, with rewards for stopping, non-duplication, and a non-empty parseable final answer — a direct response to PACE failures where models exhausted their budget in reasoning and never emitted a final channel.
The simplify reflex governs editing: minimal correct edits, the smallest helper justified by a no-tool baseline, tests preserved, regressions repaired rather than papered over.
Conciseness is judged by shorter trajectories that still pass the verifier — a capability, not a style preference. Planned
v6 states its boundaries plainly. Honesty is a value, and saying what is not yet true is how it is kept. The boundaries below hold while the result is still open.