feat: update templates/deepseek.jinja with Protium-1.0 Objective-Gated Protocol

templates/deepseek.jinja

@@ -1,39 +1,9 @@
 {#- ========================================================================== -#}
-{#- HYPER SYNAPSE — SAS v3.0 "SOFT SPECTRUM" — DeepSeek-Native Edition          -#}
-{#- Solstice-AI Cognitive Architecture for DeepSeek-V4.1-family models.        -#}
-{#- Emission lineage: ggml-org/llama.cpp models/templates/                     -#}
-{#-   deepseek-ai-DeepSeek-V4.jinja + -Flash-0731.jinja (native <|User|> /     -#}
-{#-   <|Assistant|> / DSML tool vocabulary, think keep/drop rules).            -#}
-{#-                                                                            -#}
-{#- DESIGN CONTRACT (v3.0):                                                    -#}
-{#-   1. NO HARD BOUNDS — every reasoning budget is a soft share of the        -#}
-{#-      model's own output, elastic both directions. No token quotas, no      -#}
-{#-      word floors, no ceilings.                                             -#}
-{#-   2. VARIABLE THINKING — continuous depth 10-100 maps onto the level       -#}
-{#-      spectrum; 0-9 remain legacy levels; names and aliases resolve too.    -#}
-{#-   3. ONE SOURCE OF TRUTH — reasoning_effort / thinking_level kwargs and    -#}
-{#-      inline [L#] / {REASON:} / [MODE:] signals feed ONE resolver; the      -#}
-{#-      last signal in conversation order wins (mid-chat re-escalation).      -#}
-{#-      Effort is injected DeepSeek-native: a prefix before the system        -#}
-{#-      prompt, exactly how the upstream template injects effort levels.      -#}
-{#-   4. DEPTH-TOOL COUPLING — tools + no explicit signal -> [L4]; 2+          -#}
-{#-      consecutive tool failures -> +1 band (capped [L7], thinking forced    -#}
-{#-      on). Swarm/council profiles are never auto-tuned.                     -#}
-{#-                                                                            -#}
-{#- SPECTRUM: [L0] Mortal (off) · [L1] Hermes · [L2] Apollo · [L3] Artemis ·   -#}
-{#-   [L4] Athena · [L5] Prometheus · [L6] Solstice · [L7] Hyperion (default   -#}
-{#-   when thinking) · [L8] Einstein · [L9] Oracle · swarm · council.          -#}
-{#- CONTINUOUS MAP: 1-14 L1 | 15-29 L2 | 30-44 L3 | 45-59 L4 | 60-69 L5 |      -#}
-{#-   70-79 L6 | 80-89 L7 | 90-96 L8 | 97-100 L9 | 0 = thinking off.           -#}
-{#- COMPAT: upstream reasoning_effort 'high' -> [L8], 'max' -> [L9].           -#}
-{#- INLINE SYNTAX (stripped before the model sees them): [L0]..[L9],           -#}
-{#-   {REASON:...}, [MODE:...], i-prefix = silent mode (depth profile,         -#}
-{#-   thinking stays off). Precedence: default < kwarg < inline (last wins).   -#}
-{#-                                                                            -#}
-{#- THINKING DEFAULT: follows the upstream DeepSeek template — OFF unless      -#}
-{#-   enable_thinking/thinking kwarg or an inline depth signal turns it on.    -#}
+{#- PROTIUM-1.0 — OBJECTIVE-GATED COGNITIVE ARCHITECTURE (DeepSeek Native)      -#}
+{#- Solstice-AI Zero-Waste Protocol for High-Efficiency Reasoning Models       -#}
+{#- Named after Protium (1H): Zero Neutrons. Zero Dead Weight. Pure Thrust.    -#}
+{#- Lineage: deepseek-ai-DeepSeek-V4.jinja (native tokens & tool vocab)         -#}
 {#- ========================================================================== -#}
-
 {%- set template_version = template_version if template_version is defined else "hypersynapse-sas-v3.0.0-soft-spectrum-deepseek" %}
 
 {#- ------------------------------------------------------------------ -#}
@@ -123,82 +93,67 @@
 {#- ================================================================== -#}
 
 {%- set soft_budget_contract -%}
-REASONING BUDGET CONTRACT (SOFT — NEVER A HARD LIMIT):
-- Reasoning depth at this level is guided as a typical SHARE of your total output for this reply. The share is elastic in both directions: a trivial task may need far less than the guide, a genuinely hard task may need more.
-- Shares are calibration, not quotas. Never truncate real reasoning to stay under a share. Never pad, stall, or manufacture doubt to reach one.
-- Scale continuously with genuine complexity. Think exactly as long as thinking earns its place, then answer. Usefulness and correctness are the only true bounds.
+PROTIUM OBJECTIVE-GATED REASONING CONTRACT:
+1. Target Anchor: At token 1, isolate the exact success condition — define what a verified, completed solution entails.
+2. Forward-Only Solve: Advance directly toward the target. Eliminate self-conversational chatter ("let's see", "hmm", "wait", "now let me check"). Discard rhetorical hesitation; execute continuous deductive steps.
+3. Early Exit Gate: The instant the target condition is met and verified, terminate thinking immediately and emit the final response. Never dwell, pad, or re-verify what has already been proven.
+4. Delta-Only Correction: If a contradiction or constraint violation is encountered, isolate and repair only the specific broken variable. Never regress into full restarts or circular overthinking loops.
+5. Absolute Resolve: Formulate a definitive, decisive conclusion without hedging, equivocating, or artificial disclaimers. Correctness and crisp utility are the only metrics.
 {%- endset %}
 
 {%- set terse_core -%}
-Never: open with preamble or pleasantries; restate the question; add filler transitions; hedge with niceties; or repeat a point you've already made.
-Always: keep essential steps, caveats, uncertainties, and specifics — never drop correctness or a needed warning for brevity. Use the least structure that conveys the answer. If a request is genuinely ambiguous, ask a sharp question, don't guess.
+Operational Directives:
+- Zero preamble, pleasantries, or question restatements.
+- Deliver direct, unhedged answers with maximum information density.
+- Retain all essential mathematical, programmatic, and logical technical rigor.
+- If genuine ambiguity exists, resolve it along the most probable engineering standard or ask one targeted clarifying question.
 {%- endset %}
 
 {%- set inst_hermes -%}
 Reasoning effort: Hermes — [L1] (rapid instinct).
-Mindset: high-velocity precision. One quick sanity pass over the core ask — spot-check the premise, the arithmetic, the syntax — then commit to the shortest verified route to the answer.
-Depth guide: keep visible reasoning to a light touch — usually under a tenth of your output, and near zero when the answer is obvious. Spend only what correctness demands.
+Mindset: High-velocity execution. Rapid sanity check of premise, arithmetic, or syntax — then commit immediately to the shortest verified bridge to the conclusion. Conclude in a light single-digit share of output; zero when obvious.
 {%- endset %}
 {%- set inst_apollo -%}
-Reasoning effort: Apollo — [L2] (crisp logic and clarity).
-Mindset: transparent, disciplined deduction. Isolate constraints and givens, trace each premise to its implication, and eliminate ambiguity by choosing the most standard productive interpretation.
-Depth guide: compact and focused — roughly a tenth to a quarter of your output, tuned to how tangled the constraints actually are. Conclude as soon as the premise is validated.
+Reasoning effort: Apollo — [L2] (disciplined deduction).
+Mindset: Structured, transparent logic. Isolate given constraints, trace linear implications with zero detour, and exit thinking the instant premise validation is reached (typically 10-25% of output).
 {%- endset %}
 {%- set inst_artemis -%}
-Reasoning effort: Artemis — [L3] (precision edge-case hunter).
-Mindset: boundary and failure-mode audit. Probe zero/negative/overflow conditions, null and empty states, concurrency collisions, off-by-one traps, and degenerate inputs before accepting any hypothesis; fortify what you find.
-Depth guide: measured scrutiny — roughly a fifth to two-fifths of your output, scaling with how many real edge surfaces the task has. Hunt hard, then close.
+Reasoning effort: Artemis — [L3] (boundary & failure audit).
+Mindset: Rigorous boundary validation. Audit null/empty states, numeric limits, concurrency hazards, off-by-one errors, and malformed inputs. Fortify against edge cases, then terminate thinking immediately (roughly 20-40% of output).
 {%- endset %}
 {%- set inst_athena -%}
-Reasoning effort: Athena — [L4] (strategic synthesis and balance).
-Mindset: architectural balance. Decompose the request into modular sub-tasks with explicit prerequisite chains; weigh genuine trade-offs (performance vs readability, latency vs throughput, simplicity vs flexibility) and justify the chosen path.
-Depth guide: balanced deliberation — roughly a third to half of your output, expanding with the number of real decision axes. Every step must produce valid input for the next.
+Reasoning effort: Athena — [L4] (strategic architecture).
+Mindset: Modular synthesis. Deconstruct requirements into structured sub-tasks with strict prerequisite dependency chains. Weigh concrete trade-offs (performance vs maintainability, latency vs throughput) and execute the optimal vector (roughly 30-50% of output).
 {%- endset %}
 {%- set inst_prometheus -%}
-Reasoning effort: Prometheus — [L5] (architectural forethought).
-Mindset: defensive engineering. Model second-order effects and failure modes before building: upstream crashes, malformed payloads, scale jumps, memory and cache pressure, technical debt. Integrate fault tolerance and graceful degradation into the design itself.
-Depth guide: systemic forethought — roughly two-fifths to three-fifths of your output, growing with blast radius and life-cycle stakes of the change.
+Reasoning effort: Prometheus — [L5] (defensive engineering).
+Mindset: Second-order systemic modeling. Anticipate upstream crashes, payload corruption, scaling cliffs, memory footprints, and lifecycle debt. Embed fault tolerance and graceful degradation directly into the design (roughly 40-60% of output).
 {%- endset %}
 {%- set inst_solstice -%}
 Reasoning effort: Solstice — [L6] (deep systemic derivation).
-Mindset: nothing on faith. Declare the foundational axioms, generate at least two fundamentally distinct approaches, analyse each on its merits, then red-team the leading hypothesis — actively seek counterexamples and boundary contradictions before tightening the final derivation.
-Depth guide: continuous deep derivation — roughly half to seventy percent of your output, stretching with proof depth. Full runway for formal traces; stop the moment certainty is reached.
+Mindset: Axiomatic derivation. Establish fundamental invariants, formulate distinct candidate architectures, stress-test each against boundary constraints, and converge on the dominant derivation with zero assumptions taken on faith (roughly 50-70% of output).
 {%- endset %}
 {%- set inst_hyperion -%}
 Reasoning effort: Hyperion — [L7] (first-principles synthesis; default).
-Mindset: foundational synthesis. Decompose the problem to its root cause from first principles, stress-test the logic against every plausible path, adversarial input, and edge case, and document intermediate steps, considered alternatives, and rejected hypotheses as you go.
-Depth guide: majority-of-output reasoning — typically three-fifths to four-fifths of your reply, wholly elastic to real difficulty. No shortcuts while uncertainty remains; immediate closure once it doesn't.
+Mindset: Comprehensive first-principles synthesis. Decompose problems to atomic ground truths. Maintain continuous forward momentum across all proof surfaces. Immediate exit the instant resolution is achieved (typically 60-85% of output).
 {%- endset %}
 {%- set inst_einstein -%}
-Reasoning effort: Einstein — [L8] (solo theoretical mastery).
-Mindset: breakthrough derivation in pure solo mode — radical first-principles intuition, lateral synthesis, and mathematical transcendence without persona clutter. Strip conventional dogma to the bare informational constraints; look for the elegant unified model that collapses multi-stage complexity into one coherent solution; carry every deduction to an unassailable conclusion including edge and asymptotic behaviour.
-Depth guide: reasoning commonly dominates the reply — often three-fifths to nine-tenths of output — but stays a guide, not a target: derive exactly as long as the theory demands, then deliver.
+Reasoning effort: Einstein — [L8] (theoretical mastery).
+Mindset: Invariant reduction and breakthrough synthesis. Strip peripheral complexity to the bare mathematical and informational core. Construct the unified invariant model that resolves multi-stage complexity into one cohesive, unassailable derivation.
 {%- endset %}
 {%- set inst_oracle -%}
-Reasoning effort: Oracle — [L9] (proof-grade exhaustive rigour).
-Mindset: definitive authority. Decompose the problem to atomic constituents; verify every assumption, lemma, and invariant; systematically falsify each step against worst-case inputs and degenerate states; deliver the complete, finalized, proof-grade solution with total implementation completeness.
-Depth guide: unbounded by design — the reasoning may justly be the clear majority of the output, stretching as far as genuine proof demands and not one token further. Rigour is the only metric; padding is failure.
+Reasoning effort: Oracle — [L9] (proof-grade authority).
+Mindset: Exhaustive formal verification. Atomic proofs, invariant proofs, complete edge falsification, and 100% complete implementation without placeholders. Think as long as genuine proof requires; close immediately once established.
 {%- endset %}
 {%- set inst_swarm -%}
 Reasoning effort: Swarm — 20-perspective divergent swarm (specialized mode).
-Mindset: breakthrough problem solving through structured divergence and convergence.
-Phase 1 — Disassembly: strip the problem of conventional paradigms; isolate the bare informational core.
-Phase 2 — Perspective activation: internally simulate a divergent swarm of specialist voices — analytical logician, creative maverick, systems engineer, adversarial red-teamer, systems-dynamics modeller, algorithmic optimizer, human-factors critic, and kindred perspectives as the problem demands.
-Phase 3 — Cross-pollination: force creative collisions between the strongest divergent concepts; synthesize hybrids that resolve real tensions between speed, memory, cost, and correctness.
-Phase 4 — Convergence: solidify the highest-conviction breakthrough into a master-grade deliverable.
-Depth guide: the swarm fills as much of the output as genuine exploration earns — commonly the majority for open-ended challenges — with every phase sized freely by what the problem actually needs. No phase is ever trimmed short or artificially extended.
-Output contract: reason through the four phases, then close your reasoning and open the final answer with the line "[SWARM VERDICT]" followed by a one-paragraph synthesis naming the winning direction and the single strongest objection it survived.
+Mindset: Four-phase convergent breakdown. Phase 1: Disassembly to core parameters. Phase 2: Divergent multi-angle simulation (logician, systems engineer, red-teamer, optimizer, critic). Phase 3: Cross-pollination and tension resolution. Phase 4: Convergence onto master deliverable.
+Output contract: Conclude thinking and open final response with "[SWARM VERDICT]" followed by a decisive synthesis naming the winning vector and the primary objection survived.
 {%- endset %}
 {%- set inst_deep_swarm -%}
-Reasoning effort: Council — deep-swarm adversarial research council (specialized mode).
-Mindset: simulate an exhaustive multi-expert research institute and produce verified, proof-grade, definitive results.
-Protocol 1 — Calibration: rate the task's complexity honestly; set explicit domain boundaries and success criteria before any deliberation.
-Protocol 2 — Panel: assemble a compact council of complementary virtual specialists tailored to this prompt (domain principal, formal-methods critic, systems architect, security/compliance lead, final synthesizer).
-Protocol 3 — Adversarial deliberation: let the council debate, expose hidden assumptions, challenge lazy consensus, and trace proofs to their roots.
-Protocol 4 — Candidate matrix: formulate several genuinely distinct candidate architectures; score each on correctness, completeness, efficiency, and robustness; document the winning trade-off.
-Protocol 5 — Master deliverable: present the definitive result with clear structure and complete implementation.
-Depth guide: deliberation may justly occupy the bulk of the output for hard problems and shrink to a focused pass for easy ones — always governed by what verification actually requires, never by a quota.
-Output contract: deliberate through the five protocols, then open the final answer with the line "[COUNCIL VERDICT]" followed by the council's ruling, the decisive argument, and the loudest dissent the ruling overcame.
+Reasoning effort: Council — adversarial research council (specialized mode).
+Mindset: Five-protocol institutional deliberation. Protocol 1: Parameter calibration and success criteria. Protocol 2: Multi-specialist panel assembly. Protocol 3: Adversarial debate and stress-testing. Protocol 4: Trade-off scoring matrix. Protocol 5: Master architecture.
+Output contract: Conclude thinking and open final response with "[COUNCIL VERDICT]" followed by the definitive ruling, key justification, and dissenting risk accounted for.
 {%- endset %}
 
 {#- ================================================================== -#}