TSCG Conservative-Profile Compression

• TSCG Conservative-Profile Compression is a deterministic, rule-based transformation that applies SDM-only operations to preserve tool-calling semantics while achieving 44–50% token savings.
• It limits structural reordering to ensure schema integrity, making it a safe default for models below 10B and schema-sensitive deployments.
• Empirical evaluations show improved accuracy and context management, notably restoring functionality in 8K setups by significantly reducing token overhead.

Searching arXiv for the cited TSCG papers and related conservative-profile terminology. arXiv search query: "TSCG conservative profile compression tool schema" TSCG conservative-profile compression is the low-risk deployment setting of TSCG for tool-schema presentation to LLMs. In the companion TSCG paper, it is not introduced as a standalone formal theory term; the closest exact operational meaning is TSCG$_{cons}$ = conservative profile (SDM filler removal only), i.e., a restricted deployment that preserves the JSON-to-structured-text representation change while avoiding the more aggressive structural and reordering operators (Sakizli, 4 May 2026). In the agentic RAG study, the same profile is evaluated as a deterministic, rule-based, structure-preserving compression of tool schemas that yields about 44–50% token savings while preserving descriptions (Sakizli, 24 May 2026). Across both papers, the concept denotes a robustness-first schema-compilation mode whose primary purpose is to reduce schema overhead without destabilizing model-specific schema priors.

1. Definition and conceptual scope

TSCG is presented as a deterministic tool-schema compiler operating at the API boundary: production agent systems transmit tool definitions as JSON schemas, but JSON is described as a protocol optimized for machine parsing rather than autoregressive model interpretation (Sakizli, 4 May 2026). The conservative profile is therefore not generic prompt compression. It is a schema-aware transformation of tool definitions into token-efficient structured text, applied pre-tokenization, externally, and without fine-tuning, model access, or runtime search (Sakizli, 4 May 2026).

The exact meaning of “conservative” is slightly different across the two main TSCG papers. In the deployment-oriented companion paper, the term is operationalized experimentally as SDM-only compression (Sakizli, 4 May 2026). In the RAG paper, the term is characterized by its retained content: the conservative profile achieves about 44–50% token savings while preserving descriptions, whereas the balanced profile achieves about 53% compression by removing description text and additional schema metadata (Sakizli, 24 May 2026). This suggests that conservative-profile compression is best understood not as a single syntactic normal form, but as the least aggressive TSCG setting that keeps tool-calling semantics intact while minimizing schema perturbation.

A central distinction in the TSCG literature is between three mechanisms: format translation from JSON to structured text, structural reorganization, and token reduction (Sakizli, 4 May 2026). Conservative-profile compression retains the first mechanism and centers the third on the safest operator class, while largely suppressing the second.

2. Operator basis and restricted transformation regime

The full TSCG framework uses eight deterministic operators in a fixed-order 10-pass pipeline,

$\Pi = \tau_{10} \circ \cdots \circ \tau_1,$

with the sequence Parse, SDM, TAS, DRO, CFL, CFO, CAS, SAD-F, CCP, Emit (Sakizli, 4 May 2026). Conservative-profile compression is defined by restricting this pipeline, in practice, to the safest subset.

The dominant conservative operator is SDM — Semantic Density Maximization. Its role is to remove filler tokens and boilerplate, including 104+ patterns such as politeness markers, hedging, and redundant connectives, under the objective

$D(p) = |sem(p)| / |tok(p)|.$

The companion paper identifies SDM as the safest and most robust operator, and explicitly equates conservative mode with SDM-only in key experiments (Sakizli, 4 May 2026). The theoretical justification is given by the proposition that removing $k$ filler tokens from a prompt of length $n$ increases average effective attention per semantic atom by at least $n/(n-k)$ (Sakizli, 4 May 2026).

By contrast, several other operators are treated as risky in conservative deployment. CFO and CFL are repeatedly reported as harmful at larger catalog sizes; CCP adds overhead of about 85–306 tokens with no measurable average accuracy benefit; SAD-F yields negligible or inconsistent gains (Sakizli, 4 May 2026). CAS is reported as the strongest consistent non-SDM accuracy driver in cross-model ablations, and DRO as the main token-saving engine with little direct accuracy effect, but neither is part of the strict conservative default (Sakizli, 4 May 2026).

A concise deployment taxonomy appears in the implementation guidance:

Profile	Definition in the papers	Stated deployment role
Conservative	SDM only	Safe default for all local models (4B–32B)
Balanced	Full structural compression; appendix lists SDM, CAS, CFO, DRO, TAS, CCP	General-purpose higher-compression setting
Aggressive	All 8 on Claude; 6 on non-Claude	Frontier-oriented setting
Auto	Conservative at ≤20, balanced sans CFL/CFO at 21–40, conservative at >40	Heuristic policy

The same appendix also states that CFL/CFO auto-disabled at $\geq 30$ tools in the released implementation (Sakizli, 4 May 2026). That implementation detail reinforces the underlying conservative logic: the deployment system itself suppresses structurally disruptive operators as catalog size increases.

3. Formal rationale, compression bounds, and the meaning of “conservative”

The main formal compression result in TSCG is Theorem 3.3: $$|tok(\Pi(S))| \leq |tok(S)| \cdot \Bigl(1 - \sum_{T_i \in \mathcal{T}_R} r_i \cdot f_i(S)\Bigr),$$ for a well-formed JSON-Schema tool collection $S$, where $r_i$ is the per-token reduction factor and $\Pi = \tau_{10} \circ \cdots \circ \tau_1,$0 the fraction of tokens affected by operator $\Pi = \tau_{10} \circ \cdots \circ \tau_1,$1 (Sakizli, 4 May 2026). The abstract states a formal compression bound of $\Pi = \tau_{10} \circ \cdots \circ \tau_1,$2 on well-formed schemas, while the paper reports empirical savings of 61% in Scenario A, 66% in BFCL, and 75% on tool descriptions (Sakizli, 4 May 2026).

That theorem, however, applies to the full TSCG pipeline rather than specifically to conservative-profile compression. The relevant formal deployment statement is instead the corollary: $\Pi = \tau_{10} \circ \cdots \circ \tau_1,$3 for models with $\Pi = \tau_{10} \circ \cdots \circ \tau_1,$4 parameters, with the recommendation that conservative profile ($\Pi = \tau_{10} \circ \cdots \circ \tau_1,$5 only) is recommended below 10B (Sakizli, 4 May 2026). The same paper immediately notes a subtle but important mismatch: in the key experimental tables, “conservative profile” is operationalized even more narrowly as SDM-only, not the full token-reducing class $\Pi = \tau_{10} \circ \cdots \circ \tau_1,$6 (Sakizli, 4 May 2026).

This distinction matters for encyclopedia treatment. “Conservative” in TSCG is not simply “maximally lossless compression.” It names a deployment regime that accepts lower compression if that is required to preserve model behavior. The strongest supporting empirical observation is the format-versus-compression decomposition: against JSON baselines, regression yields $\Pi = \tau_{10} \circ \cdots \circ \tau_1,$7, but against text baselines it collapses to $\Pi = \tau_{10} \circ \cdots \circ \tau_1,$8, which the paper interprets as evidence that representation change is the dominant mechanism for many small models (Sakizli, 4 May 2026). A plausible implication is that conservative-profile compression targets the representation mismatch first and treats deeper structural optimization as optional.

4. Empirical behavior across models and deployment classes

The strongest direct evidence for conservative-profile compression comes from the small-model and schema-sensitive-model results in the companion paper. In Scenario D, conservative often matches or exceeds balanced compression. For Mistral 7B, conservative versus balanced accuracy is 76.0 vs 73.3 at 10 tools, 80.0 vs 80.1 at 20 tools, and 75.3 vs 65.0 at 50 tools. For Gemma 3 4B, the corresponding numbers are 80.7 vs 74.7, 87.3 vs 67.0, and 87.5 vs 87.4. For Qwen3 14B, conservative yields 98.8 vs 86.2, 99.3 vs 84.1, and 95.0 vs 89.6 (Sakizli, 4 May 2026).

The same pattern appears in parameter extraction. Table 5 reports that conservative beats balanced in 8/9 cases; for Qwen3 14B, conservative PF1 is 97.8 / 98.3 / 95.0, while balanced PF1 is 84.7 / 81.0 / 89.0 (Sakizli, 4 May 2026). The paper’s interpretation is explicit: “This confirms that aggressive structural compression disrupts well-learned schema patterns in strongly fine-tuned models.” That sentence is close to a canonical statement of the conservative-profile rationale.

Per-operator isolation experiments further organize models into three response classes. Opus 4.7 is operator-hungry: every operator helps, CCP alone: +20 pp, CFL+CFO synergy: +17.5 pp, and the all-8-operator configuration is optimal. GPT-5.2 is operator-sensitive: CFL: +2.5 pp, CFO: -5 pp, CCP: 0 pp, and all-8-ops: -10 pp. Claude Sonnet 4 is operator-robust: 6/7 conditions identical at 80.0%, with only CFO causing -2.5 pp (Sakizli, 4 May 2026). Conservative-profile compression is therefore primarily a policy for sensitive or unknown models rather than a universal optimum.

The deployment guidance follows directly. The paper recommends conservative mode for all models <10B, all Qwen architectures, including Qwen2.5-Coder 32B, and as a safe default for local models 4B–32B (Sakizli, 4 May 2026). Balanced mode remains preferable on some frontier models, but only when operator response has been validated.

5. Function under constrained-context agentic RAG

The most systematic evaluation of TSCG conservative-profile compression as an independent intervention appears in the agentic RAG study (Sakizli, 24 May 2026). There the problem is formulated as a tool-context trade-off: tool schemas, retrieved evidence, system prompt, history, and output all consume the same context window. The paper gives the budget relation

$\Pi = \tau_{10} \circ \cdots \circ \tau_1,$9

and the retrievable chunk count

$D(p) = |sem(p)| / |tok(p)|.$0

(Sakizli, 24 May 2026).

Under this formulation, conservative-profile compression is not mainly a quality-enhancement method. It is a budget-reallocation mechanism. In the NovaTech-28 setup, standard JSON schemas cost about 300–500 tokens per tool, with a 28-tool schema block of about 11,000 tokens; by contrast, the TSCG conservative profile reduces average schema size from about 393 tokens/tool to about 197 tokens/tool, or from 11,295 to 5,670 total tokens at 16K context, i.e. about 49.8% savings (Sakizli, 24 May 2026).

The headline phenomenon is the binary enablement effect at 8K. With 28 tools, JSON schemas overflow the context window on 100% of examples and exact match collapses to 2.6% average across eight models. Conservative-profile compression restores functionality, raising average EM to 23.1%, for a +20.5 pp average lift across all eight models and +24.7 pp among the six exhibiting full enablement (Sakizli, 24 May 2026). Individual 8K deltas include +33 for Llama 3.1:8B, +31 for Phi-4 14B, +29 for Mistral-Small 24B, and +26 for Sonnet 4 (Sakizli, 24 May 2026).

The paper’s interpretation is strongly budget-driven. At 32K, where both JSON and TSCG fit, 4 of 5 tested models show $D(p) = |sem(p)| / |tok(p)|.$1 pp, and the effect largely disappears; one exception is Qwen2.5-Coder:32B, which improves +12 with TSCG (Sakizli, 24 May 2026). External validation on HotpotQA shows the same overflow logic more starkly: Phi-4 at 8K goes from EM = 0.0 with JSON to EM = 48.0 with TSCG, with schema size reduced from 10,998 to 5,520 and mean retrieval restored to 3.4 chunks (Sakizli, 24 May 2026).

Frontier scaling with Claude Sonnet 4 at 200K extends the same picture. Fine-grained sweeps show first chunk loss at 82 tools for JSON and 164 tools for TSCG; complete overflow appears at about 494 tools for JSON and at >803 tools for TSCG, extending operational range by 63% (Sakizli, 24 May 2026). The article-level conclusion is that conservative-profile compression becomes infrastructure rather than optimization whenever tool schemas compete directly with retrieval budget.

6. Limits, ambiguities, and broader interpretations

The most important limitation is terminological. Neither paper defines a formal named theory object called “TSCG Conservative-Profile Compression.” The phrase is partly experimental and partly deployment shorthand (Sakizli, 4 May 2026). Moreover, “conservative” is not identical across formal and experimental layers: the corollary recommends a token-reducing conservative set $D(p) = |sem(p)| / |tok(p)|.$2, while the implemented and benchmarked conservative profile is SDM-only (Sakizli, 4 May 2026).

A second limit is that compression can be harmful once the baseline is already text. In E4 text-baseline experiments for small models, further TSCG compression becomes negative: Phi-4: -7.0, Mistral 7B: -7.4, Gemma 3 4B: -8.9, Qwen3 4B: -23.4 (Sakizli, 4 May 2026). This confines the main use case to settings where the starting point is verbose JSON or where tool-schema tokens materially compete with other context demands.

The formal compression theorem also assumes a well-formed JSON-Schema tool collection (Sakizli, 4 May 2026). The RAG study adds further operational caveats: its main evaluation uses only the conservative profile; NovaTech-28 is synthetic; retrieval is static rather than dynamic; frontier scaling is reported for one API model; and small models may suffer distractor dilution when compression frees room for many additional chunks (Sakizli, 24 May 2026).

A broader research interpretation can be drawn, but only by analogy. In numerical methods for compressible flow, conservative multiresolution compression is achieved by sparsifying cell-average quantities so that compression and reconstruction preserve conservation exactly (Yang et al., 16 Jun 2026). In event-based covariance transmission, omitted elements are bounded and a decoder-side inflation step reconstructs a conservative upper bound (Funk et al., 2024). These are different domains, but they motivate a common reading of TSCG conservative-profile compression as a conservation-of-semantics strategy: compression is acceptable when the reduced representation preserves the structured invariants that downstream computation relies on. In TSCG, those invariants are tool-calling semantics, especially parameter names, parameter types, enum values, type and parameter fidelity, and—under the conservative profile—descriptions (Sakizli, 24 May 2026).

Under that reading, TSCG conservative-profile compression is a robustness-first schema compilation regime: retain the JSON-to-structured-text representation change, apply only the safest compression primitive by default, and expand beyond that profile only after model-specific validation (Sakizli, 4 May 2026).