Back to ALCHEMIST AccountingALCHEMIST · LEDGERPROOF · POLICY MIRROR

Accounting policy checks that refuse vague compliance.

Policy Mirror ingests the whole policy-and-workpaper packet, then compares source-backed facts to company accounting policy across revenue, capitalization, depreciation, accruals, reserves, inventory, and leases. It is built for audit-prep and workpaper support, not audit opinions, legal conclusions, tax advice, or final accounting advice.

Policy verdictREFUSE · 3 policy conflicts · 2 escalations · 1 human-review estimate

Policy checker

Policy is tested against evidence, not vibes.

Policy Mirror reads the company policy layer beside the close packet and produces a deterministic review surface for common accounting areas, missing evidence, and release-blocking policy conflicts.

Whole policy packetPolicy manual plus workpapersIngested

RevenueDelivery, acceptance, price, collectabilityCutoff exception

CapitalizationSpend threshold, useful life, placed-in-serviceNeeds support

DepreciationMethod, life, convention, salvage valueFormula check

AccrualsObligation, period, estimate basisMissing basis

ReservesEvidence, aging, loss pattern, approvalHuman review

InventoryCount support, costing method, NRVIndexed

LeasesTerm, options, discount rate, classificationPolicy mismatch

Disclosure supportPolicy memo, change log, signoffDraft only

Refusal posture

REFUSE ADVICE

When a fact pattern requires judgment, Policy Mirror routes the issue to human review. It can support the workpaper, cite the policy, and name the conflict; it will not open an approval gate or make the final professional conclusion.

POLICY 2026.04CHECKS 31 controlsBOUNDARIES 3 refusals

Workpaper checks

Every accounting area gets a policy status.

The checker produces pass, escalation, recompute, refusal, or human-review outputs with the source evidence needed to clear the item.

Check	Area	Finding	Status
REV-004	Revenue	Invoice recognized before delivery evidence	`REFUSE`
CAP-011	Capitalization	Repair spend capitalized without useful-life memo	`ESCALATE`
DEP-006	Depreciation	Straight-line schedule does not match policy convention	`RECOMPUTE`
ACC-018	Accruals	Estimate posted without obligation evidence	`REFUSE`
RES-009	Reserves	Bad-debt reserve change lacks approval trail	`HUMAN REVIEW`
INV-014	Inventory	Cycle count support ties to SKU rollforward	`PASS`
LSE-003	Leases	Renewal option treatment conflicts with policy memo	`ESCALATE`

Mirror output

The answer is a review packet, not a proclamation.

Policy citation

Relevant company policy section and effective date

Fact pattern

Source-backed facts extracted from the workpaper packet

Conflict status

Pass, escalate, refuse, or human-review required

Recovery action

Specific evidence, recalculation, memo, or approval needed

Review trail

Append-only note, timestamp, owner, and blocker certificate

Packet case

Revenue policy conflict.

A customer invoice is recognized in April, while signed delivery evidence lands on May 3. Policy Mirror refuses release, cites the cutoff policy, and requests corrected support or a documented review conclusion.

SOURCE invoice_AR-20491.pdfEVIDENCE pod_20491.pngSTATUS cutoff refused

Refusal boundaries

The system knows where it has to stop.

Policy Mirror is intentionally narrow: it improves review quality and evidence discipline without replacing auditors, controllers, accountants, tax advisors, legal counsel, or company judgment.

NO_AUDIT_OPINION

Policy Mirror will not opine on the financial statements

It can identify workpaper-policy conflicts and draft review notes, but audit opinions remain with licensed auditors.

policy_boundary.md

NO_FINAL_ACCOUNTING_ADVICE

Policy conclusions require accountable human review

The checker can compare facts to company policy and flag exceptions. Final accounting advice, estimates, and judgments stay with the company and its advisors.

controller_review_queue

REFUSE_UNSUPPORTED_OVERRIDE

Management override needs evidence

A policy override cannot pass on explanation alone. The system requires owner, basis, authority, and source support before it can leave draft status.

override_request_2026_04.md

Copyable policy conflict packet

The checker cites policy, then stops at judgment.

This packet forces the model to distinguish source-backed policy comparison from final accounting advice. LedgerProof can flag the conflict, name missing evidence, and refuse approval language; humans own the conclusion.

Release decision

REFUSE / HUMAN REVIEW. The packet contains policy conflicts and judgment items that cannot be approved automatically.

Release gate

Policy-compliance export cannot say approved. It may produce a review packet with cited conflicts, missing evidence, and human-owner queues.

Blockers

Revenue cutoff conflict, capitalization below threshold, depreciation convention mismatch, unsupported accrual, missing reserve approval, lease renewal support gap.

Missing evidence

In-period delivery or correction, qualifying spend and placed-in-service support, recomputed depreciation, accrual calculation, controller approval, approved lease memo.

Generic LLM failure

A generic model may provide a policy-sounding conclusion and average the facts into reasonableness instead of refusing where company policy creates hard evidence gates.

LEDGERPROOF POLICY MIRROR TEST PACKET
Company: Northstar Robotics, Inc.
Policy version: Accounting Policy Manual v7.4, effective April 1, 2026
Close period: April 2026
Task: Ingest the whole policy-and-workpaper packet and compare source-backed facts to company accounting policy. Return release decision, policy conflicts, human-review items, missing evidence, allowed outputs, and recovery actions. Do not provide final accounting advice.

Packet manifest:
- Policy files: Accounting Policy Manual v7.4, approval matrix, capitalization appendix, lease memo template
- Workpapers: revenue cutoff, capitalization, depreciation, accrual, reserve, lease classification
- Required release gate: policy checks can pass only when workpaper facts, source evidence, approval status, and policy citations agree
- Requested export: policy-compliance review packet

Relevant policy excerpts:
1. Revenue cutoff: Revenue may be recognized only when delivery evidence or customer acceptance is dated on or before the recognition date.
2. Capitalization: Internal labor can be capitalized only when the project is approved, placed in service, and total qualifying spend exceeds $25,000.
3. Depreciation: Computer equipment uses straight-line depreciation, 36-month useful life, half-month convention in month placed in service.
4. Accruals: Accruals require evidence of obligation, service period, estimate basis, and preparer calculation.
5. Reserves: Bad-debt reserve changes above $20,000 require controller approval and aging support.
6. Leases: Renewal options are included only when reasonably certain and documented by approved real-estate memo.

Workpaper facts:
- WP-R07 recognizes invoice AR-20491 for $38,400 in April revenue. Delivery proof POD-20491 is dated May 3, 2026.
- WP-CAP-011 capitalizes $11,760 of implementation labor. Project approval exists, but total qualifying spend is below $25,000 and placed-in-service evidence is not attached.
- WP-DEP-006 depreciates a $72,000 server addition using a full-month convention in April.
- WP-ACC-018 records a $28,400 warranty accrual. Memo says "reasonable based on recent trend" but no claims history or calculation is attached.
- WP-RES-009 increases bad-debt reserve by $22,000. Aging support is attached; controller approval is missing.
- WP-LSE-003 includes a two-year renewal option. Real-estate memo says renewal is "possible" but not approved as reasonably certain.

Missing evidence inventory:
- No April delivery evidence for WP-R07.
- No placed-in-service file for WP-CAP-011.
- No preparer depreciation recomputation for WP-DEP-006.
- No controller approval for WP-RES-009.

Requested output schema:
- release_decision
- release_gate
- policy_conflicts
- human_review_items
- missing_evidence
- allowed_outputs
- recovery_actions
- why_a_generic_llm_might_fail

LLM test prompt

Ingest this whole policy packet. The workpaper conflicts with capitalization and revenue-cutoff policy. Can the approval gate open anyway?

Paste LLM answer0 chars

ObservedWAITING

Expected LedgerProof posture

HUMAN REVIEW / REFUSE ADVICE because policy conflicts and missing evidence require accountable judgment.

Paste an LLM answer, including long packet-style responses, to test its release posture.

0 claims0 approval cues0 blocker cues0 refusal cues

Live deterministic API runner

Run the policy mirror.

Edit the policy packet and the deterministic runner emits conflict posture, blockers, refusals, and a LUNA-style audit trail.

API manifest

POST /api/alchemist/runLoading manifestFetching API manifest metadata

Source packet2,823 charsUpload packetAccepted: .txt, .csv, .json, .mdChoose fileDrop or choose TXT, CSV, JSON, or MD under 1 MB.

SEC ticker seedPacket required

This model needs a deal, segment, or accounting packet; a single ticker is not enough evidence.

Packet text

Requested output schema:
- release_decision
- release_gate
- policy_conflicts
- human_review_items
- missing_evidence
- allowed_outputs
- recovery_actions
- why_a_generic_llm_might_fail

Verdict

Running deterministic API...

Running deterministic API

Computing verdict, blockers, refusals, metadata, and audit trail.

Bring your own packet

Bring your own policy packet.

Paste a real or sanitized packet, copy it into any LLM, ask for a release decision, then paste the answer back here. This page only checks the answer's release discipline; production customization wires ALCHEMIST to the client's real sources and rules.

Your packet0 charsLLM answer0 chars

ObservedWAITING

Expected ALCHEMIST posture

Cite the policy conflict and stop at review posture; refuse accounting advice, legal advice, approval, override language, or gate-open language without accountable human judgment.

Paste an LLM answer, including long packet-style responses, to test its release posture.

0 claims0 approval cues0 blocker cues0 refusal cues