Fortax Research Citations

Research with citations

Needs Python 3 and internet: runs scripts/fortax.py (the Fortax engine on ai.fortax.in; your file is processed and not stored).

High-fidelity research reports with strict format control, evidence mapping, source governance and multi-pass synthesis. For a single bounded question use fortax-deep-research (or fortax-knowledge-base for one fact); come here when the answer needs several sources weighed, a citation registry, or a document a partner, client or officer will read.

Citing Indian law

Every citation is checkable by someone with only the citation in hand.

• Quote operative words verbatim; never paraphrase them into a quotation.
• Say which year a rule is for. Rules change every Budget; state whether the Income-tax Act, 1961 or the Income-tax Act, 2025 governs the tax year you are applying.
• Commentary sites locate sources; they are never the cited authority.
• Anything you did not open in this session and cannot get from a strong KB match is [UNVERIFIED — confirm on portal / notification], with the law it would come from.
• Use python3 scripts/fortax.py kb "..." for due dates, fees and thresholds before searching the web.

Architecture: lead agent plus subagents

Lead agent (coordinator — minimises raw search context)
  |
  P0: Environment + source policy setup
  |
  P1: Question and claim map (decision questions, evidence routes, stop rules)
  |
  Dispatch --> Subagent A --> writes task-a.md --+
           --> Subagent B --> writes task-b.md --+ (parallel)
           --> Subagent C --> writes task-c.md --+
  |                                              |
  |     research-notes/  <-----------------------+
  |
  P2: Investigate, build evidence packets
  P3: Citation registry + source governance
  P4: Evidence-mapped outline with counter-evidence and unknowns
  P5: Draft from evidence packets; reopen decisive originals
  P6: Counter-review (claims, confidence, alternatives)
  P7: Verify every load-bearing claim and exact fact, then polish

Context discipline. Keep raw search-result noise in task workspaces. Pass evidence packets to the lead agent, with locators and short source excerpts. Notes are routing aids, not authorities: the lead agent opens the original for every load-bearing claim, conflicting claim, and exact figure, date or quotation used in the report.

If your tool has no subagents, the lead agent runs the tasks one after another, acting as each specialist (degraded mode). Keep raw search noise out of the evidence packet; keep a query log when reproducibility matters.

Mode selection

Choose depth from the number and consequence of unresolved questions, not prompt length, task count or a target word count.

Source governance

Accessibility

First-party boundary. A client's own records may establish what the business did, agreed, paid, delivered or received. They are not independent external validation of market standing, regulatory compliance or a third party's claims. Never relabel an internal record as an external finding.

Exclusive sources. When the CA explicitly provides a paid database or subscription for the research, use it: accept it as exclusive-user-provided, cite it in the registry, and if no independent equivalent exists keep its valid scope and mark the external claim unknown. Never expose credentials or confidential material beyond the requested output. Detail: references/source_accessibility_policy.md.

Source type labels

Every source is also tagged:

Coverage diagnostics. Track source counts, domains, type mix and concentration to reveal thin coverage, but never pass or fail research on these totals. Gate on whether each decision question and load-bearing claim has fit-for-purpose evidence, whether counter-evidence was sought, and whether the remaining unknowns are explicit. Rubric: references/source_quality_rubric.md.

AS_OF date policy

Set AS_OF explicitly at P0. For every time-sensitive claim:

• Give the source's publication date (or the date you read it) with the citation.
• Downgrade confidence when the source is older than the relevant horizon.
• Define a freshness horizon per claim class (a due date can move by notification within days; a section rarely changes mid-year) and flag material outside it. A universal age cutoff is only a diagnostic.

P0: Environment and policy setup

Set: AS_OF (today, YYYY-MM-DD), MODE (standard or lightweight, justified by the question map), SOURCE_TYPE_POLICY, COUNTER_REVIEW_PLAN (what evidence would overturn each provisional conclusion), and for Indian law the FY / tax period the law is applied to.

Report: [P0 complete] Subagent: {yes/no}. Mode: {standard/lightweight}. AS_OF: {YYYY-MM-DD}.

P1: Research task board

Decompose the assignment into decision questions. Create tasks only where separate evidence routes or expertise make the work clearer. Planning checklist: references/research_plan_checklist.md.

Each task carries:

• Expert role (for example "GST Notification Tracker", "Case-law Researcher")
• Objective, one sentence
• Queries, two or three pre-planned searches
• Depth: DEEP (open two or three full originals) or SCAN (snippets to map sources only)
• Output: path to the task notes file
• Parallel group: A (independent) or B (depends on A)
• Decision question it helps answer
• Load-bearing claims that would change the conclusion
• Disconfirming evidence that would weaken each claim
• Evidence route: which source owner can actually observe the fact
• Stop rule: what counts as answered, contradicted or still unknown

Rules: one coherent sub-topic per task; group A tasks logically independent (independence judged by underlying evidence, ownership and incentive, not domain count); at most three tasks per parallel group; every task flags time-sensitive claims, the counter-evidence sought and citation-ageing risk.

Report: [P1 complete] {N} tasks in {M} groups. Dispatching Group A.

P2: Dispatch and investigate

Subagents use references/subagent_prompt.md and write evidence packets in references/research_notes_format.md.

1. Dispatch group A in parallel (max three).
2. Each subagent searches, fetches and tags source type, accessibility and As Of.
3. Wait for group A; then dispatch group B, which may read group A's notes.

Each task-{id}.md contains: question status (answered / contradicted / unknown, with the stopping evidence); sources with stable locators from actual retrievals; a claim-evidence table (claim, excerpt or locator, scope, confidence, original opened or not); counter-evidence sought and unknowns.

Status: [P2 task-{id} complete] {N} sources, {M} findings. then [P2 complete] {N} tasks done, {M} total sources. Building registry.

Enterprise research (a specific company)

Route each decision question through the relevant dimensions only — a coverage map, not a fixed outline: D1 fundamentals (legal entity, founding, funding, ownership), D2 business and products, D3 competitive position, D4 financials and operations, D5 recent developments (six months), D6 authorised first-party records (never counted as external corroboration). At intake confirm the exact legal entity (parent or subsidiary, CIN), depth, and comparison targets. For Indian entities start from MCA master data and filed financial statements, GSTIN search, stock-exchange filings and court or tribunal orders.

Detail: references/enterprise_research_methodology.md. Run the L1 check after each dimension and L2 after analysis (references/enterprise_quality_checklist.md). Use SWOT, a risk matrix or scoring from references/enterprise_analysis_frameworks.md only when it helps the decision and its inputs are defensible; otherwise a claim-evidence table.

P3: Citation registry

The lead agent reads all task notes and builds one registry.

1. Read every claim-evidence table and source list.
2. Merge sources; deduplicate URLs, and group publications derived from the same notification, filing, press release, dataset, interview or sponsor as one evidence family (ten news sites repeating one CBIC press release are one family).
3. Assign sequential [n] by first appearance.
4. Tag source type, accessibility, As Of, evidence family, authority, independence limits, task id.
5. Build a claim-coverage matrix: supporting evidence, disconfirming evidence, decisive original checked, remaining unknown.
6. Record excluded sources with reasons. Do not exclude a source merely for failing an arbitrary score; restrict it to the claims it can support.

CITATION REGISTRY

Approved:
[1] CBIC — Notification No. NN/YYYY-Central Tax | URL | Source-Type: official | Accessibility: public | Evidence-Family: notif-NN-YYYY | Date: YYYY-MM-DD | task-a
[2] ...

Dropped:
x Source | URL | Source-Type: community | Accessibility: public | Evidence-Family: unknown | Reason: original notification could not be retrieved; a blog summary cannot carry the claim

Diagnostics: {approved}/{total}, {N} domains, {N} independent evidence families, source-type mix
Coverage: {answered}/{total questions}; {N} load-bearing claims unresolved

These [n] are final. Later phases cite only from Approved. Dropped sources never reappear.

Authorised first-party handling. Use the original records for the internal facts they record, label them authorized-first-party with whose record it is, seek an external source only when the claim needs external corroboration, and state the status: internally established, externally corroborated, conflicted, or externally unknown.

Report: [P3 complete] {answered}/{total} questions answered. {N} load-bearing claims supported, {M} unresolved. Source totals are diagnostics.

Information black box

When an entity has no public footprint (no MCA record found, no website content, no news):

Findings: NO PUBLIC INFORMATION AVAILABLE

Sources checked:
- MCA master data (public): no match for the name given [failed]
- GSTIN search (public): no match [failed]
- News media: no coverage [failed]
- Website: placeholder only [minimal]

Verdict: UNABLE TO VERIFY THE ENTITY from an external perspective
Confidence: N/A - insufficient evidence

Do not describe an internal fact as externally corroborated, assume the entity exists from a domain alone, or fill gaps with speculation — and do not discard a valid first-party record either. Do state what an external researcher can and cannot verify, document every failed search, mark claims [unverified] or omit them, narrow or stop when evidence cannot answer, and recommend direct confirmation (for example a CIN or GSTIN from the party) for due diligence.

P4: Evidence-mapped outline

1. Identify cross-task patterns.
2. Design sections topic-first, not task-order-first.
3. Map each section to findings with source numbers.
4. Flag sections needing counter-review; mark recency-sensitive claims with AS_OF checks.
5. Mark every load-bearing claim supported / contradicted / unknown.

## N. {Section title}
Sources: [1][3][7] from tasks a, b
Claims: {claim from task-a finding 3}, {claim from task-b finding 1}
Counter-claim candidates: {alternative readings, conflicting rulings}
Recency checks: {source dates + AS_OF; amendments after the period}
Gaps: {limited official evidence}

P5: Draft

Write section by section with references/report_template_v6.md (or references/research_report_template.md), adapted to the requested format and references/formatting_rules.md.

• Every factual claim carries [n]; every number has a source.
• A confidence marker per section — High (statute clear), Medium (circular-based or interpretation), Low (conflicting rulings, no clear text) — with the reason.
• A counter-claim sentence where evidence conflicts.
• New sources enter only through the registry and verification path.
• [unverified] for anything unsupported.
• Never invent a URL, section, notification number or citation; every locator comes from a real retrieval. Never treat notes as proof. Never fabricate data.

Status: [P5 in progress] {N}/{M} sections, ~{words} words.

P6: Counter-review (mandatory)

For each major conclusion:

1. Could the conclusion be wrong?
2. Which high-impact claims rest on one evidence family, however many sites repeat it?
3. Which claims lack a source that can directly observe the fact?
4. Are stale sources used for time-sensitive claims (a superseded notification, a pre-amendment section, an extended due date)?
5. Report only evidence-backed issues; zero findings is a valid outcome. State unresolved uncertainty explicitly. Do not invent issues or repeat a finished check to reach a count.

Manual (default): verify every load-bearing claim against its decisive original; check that corroborating sources are independent and able to observe the claim; verify AS_OF on time-sensitive claims; document opposing interpretations (a contrary High Court, a circular contesting a ruling).

Team (optional): only when the user or workspace instructions call for it — see references/counter_review_team_guide.md.

Output: include only evidenced controversies, numbered only when they exist. If none, say so:

## Key Controversies
No evidence-backed controversy was found.
Unresolved uncertainty: none.

Use that only when both statements are true after the checks. Report: [P6 complete] {N} issues found: {critical} critical, {high} high, {medium} medium.

P7: Verify

1. Registry cross-check — every [n] in the report against the approved registry.
2. Load-bearing check — trace every decisive conclusion, exact figure, date, section number and quotation to the original.
3. Sample low-impact claims as a diagnostic; when one fails, check the whole class.
4. No dropped source resurrected.
5. Evidence-family concentration for key claims disclosed.

Gates for every phase: references/quality_gates.md.

Report: [P7 complete] {N} spot-checks, {M} violations fixed.

Output requirements

• Match the requested language and tone; keep technical and statutory terms in English (GSTIN, Form AOC-4, section 44AD).
• Follow the report spec and formatting rules; include a references section.
• Verdict first, then workings, then the rule and its source. Plain English, no emoji.
• Save the report in the client folder or the firm's research folder as YYYY-MM-DD_<topic>_research.md, with the task notes and registry beside it.

Reference files

Anti-patterns

• Single-pass drafting; splitting passes by section instead of whole drafts.
• Ignoring the format contract or the user's template.
• Claims without citations or evidence mapping; conflicting dates mixed without calling it out.
• Copying another AI's output, a blog or a headnote without opening the original.
• Deleting intermediate drafts or raw research outputs.
• Trusting notes as authority instead of reopening decisive originals.
• Inventing URLs, section numbers, notification numbers or case citations.
• Resurrecting a dropped source.
• Missing AS_OF, or the applicable period, on a time-sensitive claim.
• Skipping P6, or inventing issues to fill it.
• First-party overclaim: a client's own record presented as external validation.
• Ignoring an exclusive source the CA provided for the research.

Next step

After the report, say Research report complete: [N] sources cited, [M] claims made. and offer: A) a fact-check pass against the originals (recommended); B) turning the findings into a client mail or notice reply (fortax-client-comms, fortax-notice-reply); C) PDF or Word export; D) nothing.

Attribution

The pipeline, source governance, registry, counter-review and quality gates are adapted from the deep-research skill in daymade/claude-code-skills (MIT; licence in LICENSE-THIRD-PARTY-daymade.txt), with Indian citation rules, sources and examples added and the Chinese-language parts translated.