Market Research Intelligence White Paper

00

ABSTRACT

Evidence organization—not autonomous market judgment.

Grandice provides two constrained market-research workflows. The first analyzes caller-provided dated evidence for exact metric changes, recency, declared source concentration, contradictions, and research gaps. The second compares declared entities only across fully aligned metric dimensions.

The current implementation contains no model, browser, retrieval, forecasting, ranking, or recommendation step. Source text remains inert data. Every displayed evidence reference carries raw-statement and canonical-record digests, and an analysis manifest binds the methodology, scope, source registry, and evidence set.

Sources remain caller attributed and unverified. Coverage measures the shape of submitted evidence; it does not measure source independence, factual truth, research quality, accuracy, or confidence.

01

EXECUTIVE PERSPECTIVE

Separate evidence processing from source truth and strategic interpretation.

Research teams routinely combine analyst reports, filings, surveys, announcements, internal observations, and structured metrics. Similar labels may hide different currencies, scales, periods, scopes, or accounting bases. Fresh-looking evidence can repeat one underlying source, while unresolved conflicts disappear inside narrative synthesis.

01

False comparability

Metrics with matching names are compared despite incompatible units, periods, currencies, or bases.

02

Source inflation

Multiple caller IDs can describe one publisher or syndicated underlying record.

03

Conflict suppression

A selected “latest” value can hide same-period disagreement or directional ambiguity.

04

Advice drift

Research summaries can become rankings, forecasts, or recommendations without an approval boundary.

Grandice addresses the bounded evidence-processing layer. The adopting organization remains responsible for retrieval, source licensing, authenticity, independence, domain interpretation, methodology approval, decisions, and externally distributed conclusions.

02

SCOPE AND RESPONSIBILITY BOUNDARY

Useful structure without claiming market truth.

Implemented

Typed caller-provided evidence and research scopes
Exact Decimal interval arithmetic
Observation and unique-source recency summaries
Declared source-ID concentration and HHI
Disagreement, contradiction, reversal, and gap counts
Strict cross-entity comparability dimensions
Canonical evidence, source, and manifest digests
Killable capacity-limited analysis workers
Privacy Shield and metadata-only domain audit

Outside current authority

Live browsing, URL retrieval, or source verification
Currency, unit, accounting, or period conversion
Forecasts, causal conclusions, and market predictions
Objective source quality or independence scoring
Vendor, company, product, or opportunity rankings
Investment, market, commercial, or strategic advice
Price targets, expected returns, or buy/sell/hold language

03

DESIGN PRINCIPLES

Make uncertainty inspectable before making evidence convenient.

01

Caller attribution is not verification

Source labels, dates, publishers, and identities remain caller assertions. Digests bind records; they do not prove truth.

02

Comparability is a schema decision

Name, unit, kind, scale, currency, period, fiscal calendar, and accounting basis must align explicitly.

03

Conflicts remain visible

Disputed endpoints are not selected as authoritative and cannot silently drive trend arithmetic.

04

Coverage is not confidence

A deterministic coverage score reports supplied cells, source IDs, and recency—not accuracy or truth probability.

05

Every output path is bounded

Evidence, rows, bindings, excerpts, response bytes, worker slots, and execution time have enforced limits.

06

Research cannot authorize action

Outputs require professional review and cannot rank, recommend, forecast, trade, purchase, or publish.

04

SYSTEM ARCHITECTURE

A constrained path from protected evidence to inspectable market outputs.

The gateway validates authentication, quotas, JSON bounds, and configured privacy patterns before strict schema validation. Validated analysis then runs inside a capacity-limited child process. A deadline terminates the worker instead of leaving expensive work running in a thread.

Figure 1. Market evidence never gains control-plane authority. Embedded instructions, HTML, URLs, and recommendation language remain inert input text.

05

MANAGED WORKFLOWS

Two deterministic agents with no retrieval or recommendation tools.

EVIDENCE

Market Evidence Analysis

market.evidence-analyst.v1

Calculates exact trends, recency, source concentration, coverage, disagreements, contradictions, transitions, and research gaps.

No browsing, source verification, forecasts, ranking, or strategic recommendation.

LANDSCAPE

Competitive Landscape

market.competitive-landscape.v1

Builds non-ranking entity profiles and compares only complete, aligned metric dimensions and period boundaries.

No conversion, imputation, “best” vendor, investment attractiveness, or decision.

Registry invariantdeterministic = truesource verification = unverifiedprofessional review = required

06

EVIDENCE CONTRACT

Structured observations replace unconstrained research prompts.

06.1

Research scope

A scope declares subject IDs, date boundaries, and expected metric dimensions. Metric evidence outside those dimensions is rejected rather than allowed to inflate coverage or suppress declared gaps.

06.2

Evidence record

Each item carries an opaque evidence ID, caller source ID, subject, type, observation/publication date, bounded title and statement, evidence kind, signal, tags, and optional complete metric record.

06.3

Metric semantics

Metric values declare kind, unit, scale, optional ISO currency, period type, start, end, fiscal calendar, and accounting basis. Floats and incomplete metric tuples are rejected.

06.4

Date separation

observed_date represents the caller-declared evidence observation/publication date. Metric periods have independent start and end fields. Future-period forecasts are outside this v1 contract.

07

EXACT METRICS AND TRENDS

Exact interval values; explicit quantization where division is required.

Accepted metric numbers contain at most 34 digits, twelve decimal places, and magnitude below 1e25. Interval subtraction uses an 80-digit local Decimal context so accepted values do not inherit ambient 28-digit rounding.

INTERVAL CHANGEΔ = V_end − V_start

Exact for accepted values.

RELATIVE CHANGE100 × Δ / V_start

Half-even quantized to six places; unavailable at zero or disputed baselines.

JSON FLOATS REJECTEDCAGR NOT INFERREDDISPUTED ENDPOINTS BLOCKEDPERCENT UNITS EXPLICIT

A percentage value of "12.5" means 12.5%, not the ratio 0.125. Percentage-point deltas use a distinct value kind and unit.

08

RECENCY, CONCENTRATION, AND COVERAGE

Describe the submitted evidence without calling it trustworthy.

Observation recency uses caller-declared evidence dates. Source recency uses each referenced source ID once, based on its caller-declared as_of date. Repeating rows from one source cannot improve the recency component.

EVIDENCE COVERAGE SCORE 25% subject cells + 25% expected metric cells + 25% declared source-ID diversity + 25% unique-source recency

Each weighted component contribution is exposed. The result is not a confidence score or truth probability.

Declared source concentration

Counts, largest share, and HHI operate on caller source IDs. Duplicate or syndicated sources can still game this measure.

Research gaps

No evidence, missing dimensions, single-source support, and stale evidence are emitted under reserved quotas.

Completeness

Any row or evidence-binding omission sets complete=false and reports section-level totals.

09

CONTRADICTION SEMANTICS

Disagreement, material conflict, and temporal change are different events.

Figure 2. A disputed endpoint prevents authoritative latest-value selection and trend arithmetic. A temporal reversal requires one unambiguous opposing signal at each endpoint.

10

COMPETITIVE COMPARABILITY

Matching names are necessary—and nowhere near sufficient.

01Metric identityname + unit + value kind

02Numeric basisscale + ISO currency

03Time basisperiod type + start + end

04Reporting basisfiscal calendar + accounting basis

05Evidence statepresent + undisputed + known basis

A row becomes comparable only when every required dimension aligns for at least two entities and no entity is missing or disputed. Unknown accounting bases are incomparable. Grandice performs no conversion or imputation.

COMPARABLEall dimensions aligned

INCOMPLETEmissing entity evidence

INCOMPARABLEmismatch or dispute

11

EVIDENCE AND MANIFEST INTEGRITY

Bind the analysis record without pretending to authenticate the source.

Figure 3. The manifest detects record substitution across a completed analysis. It does not verify publishers, dates, source independence, licenses, or factual claims.

12

PRIVACY AND CONFIDENTIAL RESEARCH

Protect locally and treat privacy headers as assertions.

Market evidence can contain contacts, account details, confidential strategy, product roadmaps, pricing, acquisition targets, and trade secrets. Local tokenization is the preferred path for confidential workloads. Pattern detection cannot identify every contextual secret.

Figure 4. X-Grandice-Privacy-Mode is reported as a caller assertion with mode_attested=false. The gateway does not infer that local protection actually occurred.

13

BOUNDED EXECUTION

Resource controls are part of research integrity.

2 MiBrequest / result ceiling

2,000evidence records

100sources / subjects / dimensions

25competitive entities

300total section rows

8bindings per displayed row

160excerpt characters

2worker slots per gateway process

8s / 10sanalysis / landscape deadline

Latest metrics, trends, contradictions, and gaps receive independent quotas so high-volume metric rows cannot hide critical conflict or gap summaries. Every omission contributes to explicit section counts and sets complete=false.

14

AUDIT AND RETENTION

Operational receipts exclude research content.

Domain audits retain request ID, timestamp, authenticated key and available organization/member references, endpoint, agent, requested privacy mode, status, duration, schema names, source/warning counts, and methodology version. Evidence statements, citations, metrics, excerpts, digests, manifests, and responses are excluded.

domain.runCONTENT RETAINED: FALSE

request_iddom_...agentmarket.evidence-analyst.v1 status200privacy_modesidecar · asserted input_schemaMarketResearchInputoutput_schemaMarketResearchAnalysis methodologygrandice-market-deterministic-1.3.0contentnot stored

Primary domain audit metadata is purged at startup and by hourly opportunistic cleanup after the configured period—90 days by default. Infrastructure logs, crash reporting, and encrypted backups have separate deployment-specific lifecycles.

15

DEVELOPER INTEGRATION

Direct market APIs or the common managed-agent envelope.

POST/v1/market/research/analyze

POST/v1/market/competitive-landscape/compare

GET/v1/agents

POST/v1/agents/runs

GET/v1/domain/runs/{request_id}

{
  "metadata": {
    "as_of": "2026-09-15",
    "sources": [{
      "source_id": "filing-a",
      "citation": "Caller repository record",
      "as_of": "2026-09-01"
    }]
  },
  "scope": {
    "scope_id": "software-market",
    "start_date": "2024-01-01",
    "end_date": "2026-09-15",
    "subject_ids": ["entity-a"],
    "metric_dimensions": [{
      "metric_name": "Revenue",
      "metric_unit": "USD",
      "value_kind": "currency",
      "scale": "millions",
      "currency": "USD",
      "period_type": "fiscal_year",
      "accounting_basis": "gaap",
      "fiscal_calendar_id": "dec31"
    }]
  },
  "evidence": ["..."]
}

16

EVALUATION ROADMAP

Evaluate arithmetic, coverage, comparability, and reviewer utility separately.

TRACK 01

Data contracts

Decimal precision and boundary values
Period-duration and fiscal-calendar validity
Currency, scale, unit, and basis mismatches
Canonical digest reproducibility

TRACK 02

Conflict behavior

Threshold boundary fixtures
Disputed latest observations
Clean reversals and ambiguous transitions
Contradiction quota preservation

TRACK 03

Research coverage

Declared-dimension cell accounting
Duplicate source-ID gaming
Missing entity and one-sided evidence
Stale and sparse evidence corpora

TRACK 04

Operational safety

Near-limit request and response bytes
Worker termination and queue saturation
Privacy and trade-secret limitations
Unsafe HTML and URL rendering tests

17

CURRENT LIMITATIONS

What this initial evidence engine does not know.

Source limits

No publisher, author, domain, or license verification
No retrieval timestamp or independent fetch record
Source IDs can duplicate or disguise one underlying source
Citations can be false, stale, incomplete, or biased
Digests prove record consistency—not factual truth

Analytical limits

No semantic extraction from unstructured reports
No causal inference, forecast, or market-size model
No currency, scope, or accounting conversion
No source quality or confidence calibration
No ranking, recommendation, or decision optimization

Privacy limits

Trade secrets and strategy may evade pattern detection
Direct hosted requests expose content to process memory
Privacy-mode headers are caller assertions
Audit metadata and backups have separate lifecycles
Downstream consumers must escape untrusted text

18

CONCLUSION

Reliable market intelligence begins by refusing to hide incompatible or missing evidence.

Grandice combines exact deterministic arithmetic, strict metric comparability, contradiction-aware evidence handling, caller-attributed manifests, bounded process isolation, Privacy Shield integration, metadata-only audit, and mandatory professional review.

The result is an evidence-processing foundation—not autonomous research, verified market truth, or strategic advice. Production adoption still requires approved retrieval, licensing, source validation, representative evaluation, governance, and accountable human interpretation.

EXACT METRICSEXPLICIT CONFLICTSHUMAN INTERPRETATION

A

APPENDIX

Response invariants and implementation references

{
  "privacy": {
    "mode": "sidecar",
    "mode_attested": false,
    "mode_source": "caller_asserted_header",
    "content_retained": false
  },
  "data": {
    "methodology_version": "grandice-market-deterministic-1.3.0",
    "evidence_provenance": "caller_provided",
    "verification_status": "unverified",
    "coverage": {
      "interpretation": "evidence_coverage_not_truth_probability"
    },
    "complete": true,
    "market_research_advice": false,
    "professional_review_required": true
  }
}

grandice_domain/market.pyEvidence schemas, metrics, recency, coverage, conflicts, manifests, and comparison

grandice_domain/agents.pyVersioned market workflow registry and declared capabilities

gateway/routes/market.pyAuthenticated APIs, process isolation, deadlines, output ceiling, audit, and headers

gateway/domain_runtime.pyRaw-JSON limits, privacy-mode disclosure, residual checks, and body-free audit

MARKET-APIS.mdInput contracts, limits, privacy guidance, and responsibility boundary

tests/test_market_domain.pyPrecision, comparability, conflicts, provenance, bounds, and adversarial tests

Important notice

This document describes the September 2026 initial implementation. It is technical product documentation, not investment, market, commercial, strategic, legal, privacy, security, or compliance advice. Sources are caller provided and unverified. Adopting institutions must perform their own validation and professional review.