chemlit-kinetics-miner
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@chemlit-kinetics-miner extract O3 rate constants from this table and verify units, SMILES, and evidence spans"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
ChemLit-Kinetics-Miner
Evidence-Traceable Scientific Literature Extraction & Cheminformatics/Physics Verification Pipeline for Aqueous Oxidation Kinetics (with MCP Server & CLI)
ChemLit-Kinetics-Miner is a closed-loop AI4Science literature curation and domain-physics verification engine designed to mine bimolecular aqueous oxidation rate constants ($k_{\text{O}_3}$, $k_{\bullet\text{OH}}$, $k_{\text{SO}_4^{\bullet-}}$, $k_{\text{ClO}^\bullet}$), experimental water-chemistry conditions ($\text{pH}$, $T$, $\text{p}K_a$, buffer matrix, radical scavengers), and canonical molecular structures (SMILES, InChIKey) from noisy environmental chemistry papers.
Bridging Three Research & Systems Pillars
ChemLit-Kinetics-Miner unifies three complementary lines of work into a single verifiable scientific extraction stack:
DECAT(Environmental Aquatic Chemistry & Reaction Kinetics Curation): Domain-grounded modeling of ozone and radical oxidation kinetics across dissociating organic micropollutants (phenols, sulfonamides, $\beta$-blockers, $s$-triazines, olefins, and amines), enforcing physical rate bounds and Henderson–Hasselbalch speciation laws.shiye/ 识页 (Local-First Scientific Document & Formula/Table OCR Parsing): Tolerant parsing of Markdown pipe tables, LaTeX\begin{tabular}environments, table footnotes, and OCR-degraded scientific notations ((3.2 ± 0.4) × 10^5 M^-1 s^-1,1.8 \cdot 10^{-2} L mol^{-1} s^{-1},k_O3 < 0.1,10⁵,l0^5,1O^3).OmniHarness&local-chat-insights(Dual-Agent Cross-Verification & Provenance): AGenerator × Domain-Physics Auditorarchitecture that cross-checks extracted table rows against body prose, deterministically repairs recoverable OCR/column-shift errors, rejects chemical/physical hallucinations, and binds every accepted record to an exact[char_start:char_end]verbatim document span.
flowchart LR
A["Scientific Literature Excerpt<br/>(Markdown / LaTeX Table + OCR Prose)"] --> B["Agent 1: Scientific Kinetics Parser<br/>(Table, Footnote & Prose Extraction)"]
B --> C["RDKit Cheminformatics Gate<br/>(Sanitization, Valence, Stereochem InChIKey)"]
C --> D["Physical Chemistry & Unit Auditor<br/>(Unit Norm, log10 k Parity, Smoluchowski, Speciation)"]
D --> E["Agent 2: Table x Prose Reconciler<br/>(Auto-Repair OCR Shifts & pH↔pKa Swaps)"]
E --> F["Verified ReactionKineticsRecord<br/>+ Exact EvidenceSpan + ChemAuditReport"]
F --> G["MCP JSON-RPC 2.0 Server & CLI<br/>(Claude Code / Codex / OmniHarness)"]Related MCP server: genpark-smart-citation-credibility-scorer-skill
Cheminformatics & Domain-Physics Audit Rules
Single-pass LLM and naive regex extractors routinely contaminate kinetic databases with broken OCSR SMILES, $1000\times$ millimolar unit inversions, OCR exponent misreads, and transposed $\text{pH}/\text{p}K_a$ columns. ChemLit-Kinetics-Miner enforces four hard scientific guardrails:
1. RDKit Valence, Ring & Skeleton Verification (chemlit_miner/chem_validator.py)
Sanitizes raw SMILES via
RDKit(DetectChemistryProblems+SanitizeMol+AssignStereochemistry), rejecting pentavalent carbons (e.g.,CC(=O)(=O)C), pentavalent neutral nitrogens, and unclosed aromatic rings (c1ccccc).Computes canonical isomeric
SMILES, 27-characterInChIKey, Hill molecular formula, monoisotopic mass, and net formal charge.Cross-checks the 14-character
InChIKeyconnectivity layer against named micropollutant skeletons to catch LLM name–structure mismatches.
2. Second-Order Unit Normalization & Logarithmic Parity Check
All second-order rate constants are converted to canonical $\text{M}^{-1}\text{s}^{-1}$ ($\equiv \text{L}\cdot\text{mol}^{-1}\cdot\text{s}^{-1}$):
1\;\text{mM}^{-1}\text{s}^{-1} = 10^3\;\text{M}^{-1}\text{s}^{-1}, \qquad 1\;\text{M}^{-1}\text{min}^{-1} = \frac{1}{60}\;\text{M}^{-1}\text{s}^{-1}, \qquad 1\;\text{cm}^3\,\text{molecule}^{-1}\text{s}^{-1} = 10^{-3} N_{\text{A}}\;\text{M}^{-1}\text{s}^{-1}Pseudo-first-order decay constants ($\text{min}^{-1}$, $\text{s}^{-1}$) lacking oxidant concentration dimensions are flagged (PSEUDO_FIRST_ORDER_UNIT). Whenever $\log_{10} k$ is co-reported, the auditor enforces:
\left|\log_{10}\!\left(k_{\text{norm}}\right) - \text{reported }\log_{10} k\right| \le \tau \quad (\tau = 0.08)catching $10\times\text{--}1000\times$ OCR exponent misreads and omitted footnote unit multipliers.
3. Smoluchowski Aqueous Diffusion Bound & Thermodynamic Bounds
Bimolecular aqueous oxidation reactions cannot exceed the Smoluchowski diffusion-controlled encounter limit at ambient conditions:
0 < k_{\text{norm}} \le k_{\text{diff, max}} = 2.5 \times 10^{10}\;\text{M}^{-1}\text{s}^{-1}, \qquad T_{\text{K}} = T_{^\circ\text{C}} + 273.15 > 0\;\text{K}4. Henderson–Hasselbalch Acid–Base Speciation Physics
Because ozone ($E^\circ = 2.07\;\text{V}$) is a highly selective electrophile, deprotonated phenolate ($\text{A}^-$) and neutral amine ($\text{B}$) species react $10^4\text{--}10^7\times$ faster than their conjugate acids ($\text{HA}$ / $\text{BH}^+$). When intrinsic species constants $(k_{\text{HA}}, k_{\text{A}^-})$ and $\text{p}K_a$ are available, the auditor verifies:
k_{\text{app}}(\text{pH}) = \alpha_0(\text{pH})\,k_{\text{HA}} + \alpha_1(\text{pH})\,k_{\text{A}^-}, \qquad \alpha_1(\text{pH}) = \frac{1}{1 + 10^{\text{p}K_a - \text{pH}}}, \quad \alpha_0(\text{pH}) = 1 - \alpha_1(\text{pH})(with full diprotic extension $\alpha_0 k_0 + \alpha_1 k_1 + \alpha_2 k_2$ for ampholytes). If $k_{\text{app}}$ fails under reported $(\text{pH}, \text{p}K_a)$ but satisfies the equation upon transposing $\text{pH} \leftrightarrow \text{p}K_a$ (corroborated by prose/reference $\text{p}K_a$), the auditor deterministically unswaps the transposed table columns (AUTO_REPAIRED).
Evaluation on constructed excerpts
The deterministic parser and audit rules were evaluated on 48 project-authored test excerpts, with 80 target records and 18 injected traps. The builder specifies the tables, prose, corruptions and expected answers. Some source notes cite related literature. The fixtures exercise six chemical families and known failure modes. Performance on independently sampled papers and scans remains unmeasured.
Pipeline Configuration | Precision | Recall | F1 Score | FP Contamination ↓ | Injected traps absent from accepted output | Unit / exponent endpoint met | Speciation / pH endpoint met | Exact text offsets + InChIKey present |
1. Baseline Single-Pass Extractor | 58.1% | 67.5% | 62.4% | 41.9% | 0.0% (0/18) | 0.0% | 0.0% | 0.0% |
2. Extractor + RDKit Structure Gate | 68.5% | 76.2% | 72.2% | 31.5% | 38.9% (7/18)* | 33.3% | 0.0% | 100.0% |
3. | 97.5% | 97.5% | 97.5% | 2.5% | 100.0% (18/18) | 100.0% | 100.0% | 100.0% |
Within the constructed fixtures, a target counts when an accepted record matches its answer key. An injected trap meets its endpoint when absent from accepted records, including when never proposed. The unit/exponent and speciation columns combine the relevant target and trap cases. The final column checks quote offsets and the presence of a 27-character InChIKey.
Paired replay of saved model responses
The trace file contains 48 DeepSeek responses and 41 Qwen responses. Offline scoring uses the same 41 excerpts, 66 target records and 16 injected traps for both. The model names below are the identifiers stored in the cache. The full pipeline can recover source records using the deterministic parser even when the model returns no usable JSON.
Stored model | Responses with a JSON records array | Raw F1 | After RDKit F1 | Full pipeline F1 | Full TP / FP / FN |
| 0/41 | 0.00% | 0.00% | 96.97% | 64 / 2 / 2 |
| 41/41 | 83.44% | 87.50% | 97.78% | 66 / 3 / 0 |
The DeepSeek raw score reflects failure to produce the required JSON in these cached calls. Its full-pipeline records were recovered entirely from the source text. Trap absence counts include cases where the extractor never proposed a trap. Exact character offsets verify that a quote occurs in an excerpt; factual support for an individual field requires inspecting the quote and its context.
Reproduce the paired results with python scripts/run_live_llm_evaluation.py --cached-only. This mode performs no network calls and includes malformed cached answers in the denominator. The result file lists the scored and missing document IDs.
Offline scores and document coverage · Constructed corpus
1. Progressive Pipeline Ablation

2. Unit and logarithmic consistency in constructed excerpts

The parity plot uses the 87 of 98 candidate traces with a positive standardized rate and a reported log value. Both coordinates come from the constructed excerpt and its parse.
3. Henderson–Hasselbalch Speciation Physics Verification

The plotted curves use the parameters in the figure script. Markers illustrate constructed examples, including contradictions and swaps.
4. Annotated cases and expected outputs across chemical families

Model Context Protocol (MCP) Integration
chemlit_miner/mcp_server.py implements a zero-external-framework MCP JSON-RPC 2.0 stdio server exposing three structured tools:
extract_kinetics_records: End-to-end dual-agent extraction, RDKit canonicalization, physics verification, and[char_start:char_end]provenance binding.validate_smiles_and_units: Standalone RDKit sanitization + rate unit conversion + $\log_{10} k$ & Smoluchowski diffusion limit audit.verify_speciation_kinetics: Standalone monoprotic/diprotic Henderson–Hasselbalch $k_{\text{app}}(\text{pH})$ consistency & $\text{pH}\leftrightarrow\text{p}K_a$ swap checker.
Mount directly into Claude Code, Codex, Cursor, or OmniHarness (mcpServers configuration):
{
"mcpServers": {
"chemlit-kinetics-miner": {
"command": "python",
"args": ["-m", "chemlit_miner", "mcp"],
"cwd": "/path/to/ChemLit-Kinetics-Miner"
}
}
}Installation & Quickstart
Installation
pip install -e ".[dev]"Command-Line Interface (chemlit-miner)
# 1. Extract & cross-verify kinetics records from a Markdown/LaTeX paper excerpt
python -m chemlit_miner extract --input paper_excerpt.md --doc-id "EST_2007_TCS" --mode full
# 2. Validate a SMILES structure, millimolar unit conversion, and log10(k) parity
python -m chemlit_miner validate --compound "Sulfamethazine" --k-value 620 --units "mM^-1 s^-1" --log-k 5.79
# 3. Verify Henderson-Hasselbalch speciation kinetics for Triclosan at pH 7.0
python -m chemlit_miner speciation --ph 7.0 --pka 7.90 --k-ha 13.0 --k-a-minus 3.8e7 --reported-k-app 4.25e6Python API
from chemlit_miner import CrossVerificationEngine
excerpt = """
Ozonation in 10 mM phosphate buffer at pH = 7.0 and T = 298 K.
| Compound | SMILES | Oxidant | k_app (M^-1 s^-1) | log k | pH | pKa | k_HA (M^-1 s^-1) | k_A- (M^-1 s^-1) |
| --- | --- | --- | --- | --- | --- | --- | --- | --- |
| Triclosan | Oc1cc(Cl)ccc1Oc1ccc(Cl)cc1Cl | O3 | 4.25 × 10^6 | 6.63 | 7.90 | 7.00 | 13 | 3.8 × 10^7 |
At pH 7.0, Triclosan (pKa = 7.90) reacted rapidly with ozone with k_O3 = 4.25 × 10^6 M^-1 s^-1.
"""
engine = CrossVerificationEngine(mode="full")
result = engine.run(excerpt, doc_id="SUAREZ_2007")
rec = result.accepted_records[0]
print(rec.compound_name, rec.inchikey, rec.standardized_k, rec.ph, rec.audit_report.status)
# -> Triclosan XEFQLINVKFYRCS-UHFFFAOYSA-N 4250000.0 7.0 AUTO_REPAIREDReproducing Benchmark Metrics, Figures & Tests
# Run full unit & integration test suite
pytest -v
# Rebuild gold benchmark corpus, run 3-stage pipeline ablation, and render 300 DPI figures
python scripts/build_benchmark_corpus.py
python scripts/run_evaluation.py
python scripts/generate_figures.pyRepository Structure
ChemLit-Kinetics-Miner/
├── chemlit_miner/
│ ├── __init__.py # Public API exports
│ ├── __main__.py # `python -m chemlit_miner` entrypoint
│ ├── schema.py # Typed Pydantic v2 models (ReactionKineticsRecord, EvidenceSpan, ChemAuditReport)
│ ├── parser.py # Markdown/LaTeX/OCR scientific table, footnote & prose parser
│ ├── chem_validator.py # RDKit structure gate, unit normalizer, Smoluchowski & speciation physics
│ ├── cross_verifier.py # Dual-Agent (Extractor x Domain-Physics Auditor) engine + optional LLM client
│ ├── mcp_server.py # Standard MCP JSON-RPC 2.0 stdio server & tool registry
│ └── cli.py # Command-line interface (`extract`, `validate`, `speciation`, `mcp`)
├── data/
│ └── benchmark_corpus.json # 48 constructed literature-style excerpts (80 gold records + 18 failure traps)
├── scripts/
│ ├── build_benchmark_corpus.py
│ ├── run_evaluation.py # Computes real ablation metrics -> results/
│ └── generate_figures.py # Generates 4 publication-grade 300 DPI figures -> assets/
├── results/
│ ├── miner_benchmark.json # Full pipeline metrics & candidate parity audit trace
│ └── error_breakdown.csv # Caught failure taxonomy across 6 chemical families
├── assets/ # 300 DPI benchmark & domain-physics diagnostic figures
├── tests/
│ └── test_miner.py # Unit and integration tests
├── pyproject.toml
├── requirements.txt
└── LICENSELicense
Released under the MIT License.
Available Tools
3 toolsextract_kinetics_recordsA
Extract aqueous oxidation rate constants (O3, •OH, SO4•-, ClO•), experimental conditions (pH, T, pKa, water matrix), canonical RDKit SMILES/InChIKey, and exact character-offset EvidenceSpan provenance from Markdown/LaTeX/OCR paper excerpts.
Inputs: text (required), doc_id, page_num, pipeline_mode.
validate_smiles_and_unitsB
Validate a compound name / raw SMILES via RDKit (catching valence & broken-ring hallucinations, computing canonical SMILES, InChIKey, formula, exact MW, charge) and audit rate constant units, log10(k) parity, and Smoluchowski diffusion bounds.
Inputs: k_value (required), compound_name (required), ph, pka, +4 more.
verify_speciation_kineticsB
Verify acid-base speciation kinetics consistency using the Henderson-Hasselbalch law k_app(pH) = alpha_0k_HA + alpha_1k_A- (and diprotic extension), catching pH <-> pKa column swaps and unphysical apparent rate constants.
Inputs: ph (required), pka (required), k_ha (required), k_a_minus (required), +4 more.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 3 tools
The tools target mostly distinct pipeline stages: extraction, chemical/unit validation, and speciation consistency verification. There is some conceptual overlap between validate_smiles_and_units and verify_speciation_kinetics because both audit correctness, but their scopes are sufficiently separated by descriptions.
All tool names use snake_case and begin with a clear action verb (extract, validate, verify). The pattern is predictable and readable throughout.
Three tools are well-scoped for a focused extraction-and-validation pipeline. Each tool has a clear role, and there is no redundant surface area.
The core workflow is covered: extract kinetics records, validate structures/units, and verify speciation consistency. Minor gaps exist around literature retrieval, record export, deduplication, or broader kinetics validation, but agents can work around these.
Maintenance
Related MCP Connectors
Auditable MCP server for PubMed, Europe PMC, ClinicalTrials.gov, and bioRxiv/medRxiv queries
Read-only MCP over an agentic SLR workspace with per-claim citation verification
Read-only MCP over an agentic SLR workspace with per-claim citation verification
Entity-resolved NIH grant corpus (FY1985-2026) via 14 read-only, provenance-carrying MCP tools.
Related MCP Servers
- AlicenseBqualityBmaintenanceEnables provenance-first scholarly retrieval, paper ingestion, and reproducible research workflows by searching academic and developer sources, extracting source-located facts and claims, and preserving evidence and provider uncertainty for MCP clients.371Apache 2.0
- FlicenseNot gradedqualityBmaintenanceEnables automated scientific paper analysis, citation credibility verification, consensus ratio calculation, and multi-hop research queries through the Model Context Protocol, integrating with MCP-compliant clients.8-
- AlicenseBqualityCmaintenanceEnables biomedical research workflows via MCP, including experiment and document search, PDF text extraction, evidence retrieval, and validation with provenance tracking.10MIT

ScholarMuse MCPofficial
FlicenseNot gradedqualityCmaintenanceConnects AI agents to a hosted remote MCP endpoint that searches academic papers, verifies references and retraction status, retrieves research datasets, grants, and biological or chemical data, and generates scientific figures across 50+ upstream sources. Supports OAuth sign-in with per-tool credit accounting, structured error reporting, and source attribution for results.-