Draft v0.1 (assembled 2026-08-19) — living system paper of the PRDT / P-Research program. The corpus numbers in this draft are claims about the P-Research corpus snapshot (503 papers, 2023-2026; 92 full-text verified) as of 2026-08-18. The live corpus grows weekly — every number here re-derives from
data/papers.jsonlin this repository. The markdown is the living source; venue-targeted PDFs are compiled per submission. License: CC BY 4.0 (text) — reuse it, cite us.A companion paper, “Agent-Native Trading Systems” (
paper3-quant-ai.md), applies the framework’s corpus to the Quant×AI frontier.
Quantitative finance treats markets as data and converts raw price streams into factors, portfolios, backtests, and risk reports. This paper proposes that the research frontier admits the same treatment, with a materially better starting position: the market data — papers, citations, code, and institutional affiliations — is public by default.
Under this lens, the mapping is direct. A paper is a data point with high-dimensional features (methods, benchmarks, models, citations). A method family is a factor whose momentum can be measured week over week. A survey is a portfolio: claims are positions, each backed by evidence and sized by confidence. A citation audit is risk control. And the researcher’s alpha is the same as a quant’s: seeing structure earlier and more reproducibly than the market does.
The thesis is most tractable in CS/AI. Papers ship code, so “backtesting” a claim means running the released repository. The feedback loop is weekly — arXiv updates every day. And unlike trading, the evidence chain is free: every claim can be traced from survey text to database record to original manuscript, with no proprietary data in between.
LLM research agents are increasingly capable, but they inherit the stochasticity of their substrate: two runs produce two answers, and “reasoning” is not auditable in the sense risk control requires. PRDT takes the opposite design position: the pipeline itself contains no LLM calls. Ingestion, structuring, verification, clustering, and auditing are deterministic programs with schema-validated outputs. LLMs are — by design — kept outside the evidence chain, reserved for the human-readable synthesis layers where judgment, not reproducibility, is the goal.
This buys three properties that agents cannot yet claim:
PRDT ships as two coordinated systems rather than one monolith:
The echo between them mirrors the quant pattern of private funds and public indices: the public layer thickens the data, attracts collaborators, and earns reputation; the private layer consumes that corpus to do deeper, opinionated research. Neither competes with the other; they compound.
This is a systems paper about infrastructure, not a claim that research is solved. The system’s limitations are enumerated in Section 7: the corpus sees only the open literature; full-text coverage is a fraction of the database; recent papers suffer citation lag; and the final publishing gates — human review and endorsement — remain, by design, outside the machine.
PRDT is a systems position paper as much as a system description. Four principles separate it from both “paper database + LLM wrapper” tools and from pure agent frameworks.
Every transformation in the evidence chain — arXiv ingestion, structuring, verification, clustering, auditing — is a deterministic program. The structurer is a keyword/regex engine with word-boundary rules for short acronyms (e.g., “cot” must not match inside “scotland”); the verifier re-runs the same taggers over full text; the auditor re-derives every quoted number from the database at run time. No LLM call sits anywhere inside the chain, because an LLM in the chain would make two properties unattainable:
The corpus (data/papers.jsonl) is append-only: ingestion never rewrites
history. Verification metadata merges into records via atomic file
replacement (temp file + os.replace), so a concurrent sweep can never
read a half-written database. The invariant: the database grows, but any
past state is reconstructible from the sweep logs and the raw arXiv cache.
The survey is not a frozen PDF: it is assembled (presearch assemble),
audited (presearch audit), and compiled (presearch latex) from the
current corpus state on every weekly run. Quoted numbers are therefore
claims about the corpus at time t, and the audit fails loudly when the
corpus moves out from under a claim. Static surveys rot; this one is
designed to contradict itself loudly rather than silently.
Evidence is labeled by its verification state, and only the strongest state backs survey claims:
abstract-level — query-of-origin pillar + keyword tags (weak signal).full-text-verified — ar5iv/arXiv-HTML full text re-tagged, method
overlap recorded (strong signal).audited — the claim-level check (record ids exist, numbers match).The README and every digest carry the disclaimer that abstract-level evidence is a research signal, not a fact. This is not modesty; it is risk control — the research-market equivalent of marking a position’s liquidity class.
PRDT’s internal layer is organized as five cooperating agents over a shared, schema-validated evidence format:
| Agent | Input class | Output |
|---|---|---|
| Open-Source Project Agent | repositories | reproducibility/integration triage |
| Frontier Paper Agent | arXiv papers | thesis, method, replication plan, targets |
| Professional Report Agent | institutional reports | fact/opinion/forecast separation |
| Real-Time Information Agent | events | impact ranking, affected conclusions |
| Analyze & Synthesis Agent | fused evidence | associations, contradictions, hypotheses, experiments |
Each agent’s output is validated against a required-field schema before
it may enter the pipeline (REQUIRED_FIELDS), and every agent is
deterministic — the paper agent, for example, converts an arXiv entry
into evidence with pure keyword/regex rules rather than an LLM call.
The synthesis agent builds a research graph whose nodes are evidence,
associations, contradictions, and hypotheses, and whose edges are typed
(SUPPORTS, CONTRADICTS, TESTS, …). Three engines produce the
graph’s intellectual content:
ai domain, cross-pillar convergence (a method family spanning ≥2
pillars with ≥3 papers) is the primary association rule.pficc domain instead
consumes explicit affected_existing_conclusions metadata.Domain-awareness is the design point: the same engines produce
fixed-income factor hypotheses for pficc and frontier-research claims
for ai, from domain adapter definitions rather than hardcoded text.
A domain adapter is a declarative bundle: taxonomy, source priorities,
evaluation metrics, integration targets, and (for ai) per-pillar arXiv
query plans. Two adapters ship:
pficc — China fixed-income multi-factor research (rates/credit/
portfolio/trading taxonomy).ai — the six-pillar frontier taxonomy (Reasoning / Agents /
Efficiency / RL-Alignment / Multimodal / Quant×AI).Adding a domain is a configuration act, not a code change; the registry
(get_domain) resolves adapters by name.
PRDT is deliberately two tracks that compound:
The echo is mechanical, not metaphorical: PRDT ingests the public corpus
verbatim (prdt ingest presearch, 503 records) and appends a
“P-Research corpus echo” section to its own weekly digest; P-Research’s
methodology (structurer, audit, connectors) originates in PRDT. This is
the research-market version of the private-fund/public-index pattern:
the public layer thickens the data and earns reputation, the private
layer spends that reputation on opinionated research.
The public scholarly data layer is friendly in principle and hostile in practice: rate limits, truncated responses, moving API fields, and coverage gaps. This section documents the engineering that keeps the loop running — each item below is a failure we actually hit, and the mechanism that survived it.
Flaky networks truncate large arXiv Atom responses mid-stream
(http.client.IncompleteRead, e.g. 31KB of 41KB delivered). Retries
alone are insufficient when truncation is bursty. The connector salvages
the partial payload: the IncompleteRead.partial bytes are parsed for
complete <entry>...</entry> blocks (each block is wrapped in a minimal
namespace-bearing feed and parsed independently), so a truncated response
yields most of its papers instead of nothing. On the first full sweep,
5 of 6 pillar queries were salvaged this way.
Full-text verification prefers ar5iv (LaTeXML HTML) and falls back to arXiv’s native HTML rendering for papers ar5iv has not converted. Two pitfalls required explicit handling:
ltx_ markers; abs pages do not.
False positives are rejected, not ingested.The verification gate re-runs the deterministic taggers over full text and records the method-tag overlap against the abstract-level record. On the first 45 verified papers the match rate was 99% (81/82), empirically validating the abstract-level structurer.
Semantic Scholar’s anonymous pool is frequently saturated (we observed sustained HTTP 429). The citation layer therefore treats OpenAlex — fully keyless, with citation counts and institutional metadata — as the default source, and S2 as an optional keyed upgrade. Two data-engineering lessons surfaced:
raw_affiliation_strings is not a top-level work
field and is not in the select whitelist; it lives inside each
authorships entry. Department-level group tracking (e.g., “HKU
Computer Science” vs “HKU” at large) is only possible after extracting
it from the right place.Citation lag is handled at the scoring layer: the novelty-citation blend falls back to method-overlap novelty alone when citation data is absent, so recent papers are never penalized for being too new to be cited.
Sustained 429s are not retried into the ground: the connector breaks after two consecutive rate-limited attempts and a group-sweep circuit breaker stops remaining watchlist entries, preserving the polite pool for the next scheduled run. Merge semantics guarantee that a failed refresh keeps the previous data — the system degrades to stale-but-true rather than empty.
presearch audit closes the loop with six mechanical checks: record-id
existence, verification status of cited records, arXiv id format and
membership, headline corpus numbers, the §3 trend table recomputed from
the verified subset, and hard claim checks (three quoted numbers
re-derived from the database). It has caught 24 mechanical errors in
this project’s own draft — 21 truncated record ids and 3 wrong external
arXiv ids — before a single human reviewed the paper. The lesson we
draw: in a pipeline that distills thousands of records into prose,
mechanical error is the base rate, and the only defense is a machine
that checks the machine.
Evaluation of a research-intelligence framework is unusual: the “test set” is the live scholarly frontier, and the “labels” are the paper records themselves. We evaluate along four axes: corpus scale, signal quality, mechanical error detection, and the usefulness of downstream artifacts. All numbers below are re-derivable from the public repository.
The P-Research corpus holds 503 papers (2023-2026), built from weekly sweeps (21-day windows) plus non-overlapping historical backfills (0-360, 360-720, 720-1080, 1080-2160 days). Year coverage: 2023 (140), 2024 (112), 2025 (131), 2026 (120). Six pillars are balanced at the abstract level (77/75/70/64/60/59 at first gate; quant pillar later grew to 95+ via backfill).
Full-text verification spans 92 records across all six pillars (45/10/10/9/9/9). On the first 45 verified lead-pillar records, the method-tag match rate between abstract-level and full-text tagging was 99% (81/82) — empirical validation that the deterministic structurer’s abstract-level signal is a reliable proxy where full text has not yet been pulled. This number is the bridge that licenses §5.8 of the survey: the abstract-level convergence table stands in where verification has not reached.
The automated citation audit runs six checks on every survey regeneration (record-id existence, cited-record verification status, arXiv id format and membership, headline corpus numbers, trend-table recomputation, and three hard claim checks). In production it has caught:
Both failures predated any human review. The base-rate claim we make is specific: in a pipeline that distills hundreds of records into prose, mechanical citation error is the base rate, and a deterministic auditor is the cheapest defense against it.
Numbers are corpus statements, not field-wide claims; recent papers are over-represented by weekly windows; pillar assignment follows the query of origin; and the closed frontier (commercial systems) is invisible to the open-literature corpus by construction.
A framework is evaluated by what it produces. PRDT’s flagship output is the living survey (“The Compute-Allocation Frontier”), and the survey is worth examining as an artifact of the system rather than as a standalone document.
The survey is not written once; it is assembled. Seven section files are
concatenated by presearch assemble, converted to LaTeX by a
deterministic markdown-to-TeX converter, and compiled to a 13-page PDF
with zero errors. Every quoted number in the draft is a claim about the
corpus at generation time, and the audit re-derives each one.
The survey’s intellectual content flows through three gates:
evg-* records.This pipeline is what turns “a survey” into “a survey with an evidence chain” — the property that makes the document updateable without rotting.
We believe the pattern generalizes beyond this survey: any research program that maintains a corpus can produce living documents — surveys, landscape reports, group radars — assembled from the same evidence layer, audited by the same machine. The survey is the proof of concept; the framework is the claim.
presearch_pillars.json ships; watchlist and connector manifests are
next) so external researchers extend the taxonomy without forks.We built PRDT on a specific wager: that in research, as in markets, the durable edge is infrastructure that sees structure earlier and can prove it. The survey is the first portfolio; the pipeline is the fund; the audit is the risk desk. Whether the wager pays is now an empirical question — and the machine will keep measuring it every Monday.