FFidelog
中文English日本語
Open prototypes

Fidelog · The evidence system

Hand it the documents. The problems surface themselves.

Give it the piles of contracts, ledgers, receipts and conversations: it reads all of them, reads them right, finds the problems, and answers any question about them. Every conclusion traces back to the source line.

Sign in to prototypesHow a conclusion is built

This page loads no external resources and calls no model. Anything holding real data sits behind a login.

Synthetic demo data

  • Business management1 / 3

    Sourceafter-sales-ledger-2026-08.csv
    1. 012026-08-03,East,SF,received intact
    2. 022026-08-05,East,YD,carton damaged · refunded
    3. 032026-08-07,South,SF,received intact
    4. 042026-08-09,East,YD,contents damaged · reshipped
    5. 052026-08-11,North,JD,received intact
    6. 062026-08-12,East,YD,carton damaged · refundedline 217

    Conclusion

    Damage-in-transit rate this month is 3.4%, concentrated on one carrier’s east-coast lane.

    Recalculateddamaged 41 ÷ shipped 1,206 = 3.4%

    Raised as a ticketOwner: Logistics · Due in 10 business days

    1. Located in the source
    2. Recalculated on the spot
    3. Matches

  • Government compliance review2 / 3

    SourceRoad upgrade project · payment approval
    1. 01Project: Road upgrade, stage 1
    2. 02Amount: $6.2m (progress payment 3)
    3. 03Approved by: Zhou Ming · Finance lead
    4. 04Approval date: 4 Mar 2026page 3 · approval date
    5. 05Contract no.: HT-2026-0031
    6. 06Note: paid under clause 5.2

    Conclusion

    Road project: payment approved 12 days before the contract was signed.

    Checked word for wordcontract signed 16 Mar 2026 − approved 4 Mar 2026 = 12 days

    Added to evidence packTimeline built automatically · remediation with an owner

    1. Located in two documents
    2. Checked word for word
    3. Sequence breach confirmed

  • Online-store customer service3 / 3

    SourceProduct specs · SKU 4471 · store wording
    1. 01Item: long-sleeve shirt · blue check
    2. 02Fabric: 100% cottonspec sheet · fabric
    3. 03Weight: 135 g/m²
    4. 04Care: cold machine wash · no bleach
    5. 05Wording #12: fabric per spec sheet, no verbal promises
    6. 06Wording #19: sizing — ask height and weight first

    Conclusion

    Customer: Is this one 100% cotton?

    Basisspec sheet “100% cotton” → candidate reply cites line 2

    OK to sendNo source, no suggestion — it says so

    1. Store wording
    2. Product specification
    3. Candidate with a source
  • Reads all of itHundreds of thousands of documents in one night, no sampling
  • Reads it rightEvery conclusion goes back to the source line, recalculated on the spot; wrong means unpublished
  • Finds the problemsPaid twice, approval before contract, one supplier for everything, including what nobody thought to look for

Who it’s for

Three roles, one problem: too many documents for anyone to read

They don’t lack documents. They lack the time to read all of them, read them right, and see the problems.

Group finance · head of audit

Hundreds of contracts, receipts and invoices, and a report due at year end.

Before

  • Three people, three days, and under a tenth of it sampled
  • Cross-department documents don’t reconcile, so each side checks its own

Now

  • The whole set checked overnight, issues ranked by strength of evidence
  • Every issue carries the source position and the recalculation, straight into the working papers

Unexpected finding

The same valve model bought from one supplier for three quarters running, never quoted; two invoices dated before the goods shipped.

Review team lead · government projects

Dozens of documents per project; the order of payments, approvals and contracts has to be paged through by hand.

Before

  • Timelines built by hand; miss one document and a chain of errors follows
  • Citations in the notice have to be traced back one by one

Now

  • The evidence pack lays itself out on a timeline; a sequence breach is obvious at a glance
  • The notice is previewed as A4 inside the system, every citation clicking back to the source

Unexpected finding

The minutes say “start payments after the contract is signed”; the approval was signed 12 days before the contract. Each document is fine on its own; the problem only appears when they meet.

Online-store owner

Thousands of conversations a day; no idea which need a person, who needs coaching, or why bad reviews cluster.

Before

  • A supervisor spot-listens and picks the coaching list from memory
  • Bad reviews are a total; nobody sees which product or which issue

Now

  • Risky conversations surface first; one of four decisions in a click, all reversible
  • Coaching looks only at the person’s attributable behaviour; a bad review opens to the customer’s own words

Unexpected finding

Bad reviews cluster on one issue of one product while agents answer exactly by the script; the fault is not the agents but the product page’s specifications.

What it does

Documents in; conclusions, evidence and next steps out

Eight things, one system. Every output traces back to the source.

  • File and organise

    Contracts, invoices, receipts, ledgers, minutes, scans, images and archives go straight in; recognised, classified, matched to an entity, de-duplicated, fingerprinted.

    Scanned contract → 97% of key fields read · filed under “Hongxin · procurement”

  • Check automatically

    Thirty-plus general rules plus domain rules: paid twice, amounts don’t match, approval before contract, purchase without contract, tax miscalculated, one model concentrated with one supplier…

    Payment approved 12 days before contract signed · needs action

  • Cite and recalculate

    Every conclusion points to the document and the line; the page re-runs the sum from the source numbers and an independent replay verifies it; a mismatch is flagged or abstained.

    (2,318,400 − 1,720,600) ÷ 1,720,600 = 34.7% ✓

  • Ask the documents

    Questions about the system, conclusions, data or common knowledge: rules answer first, then passages from the source; the model only phrases, and every sentence is traced back.

    “Payments over 50k last quarter?” → 7, each one opens

  • Discover relationships

    Documents connect themselves through shared entities, businesses, contract and receipt numbers; click a company to see all its documents, issues and linked IDs.

    Hongxin → 12 documents · 3 need action · 6 linked IDs

  • Close the loop

    A conclusion becomes a ticket with an owner, a plan and a due date; closing requires an outcome; outcomes feed the knowledge base.

    Owner Procurement · quote review · 10 business days · closed → knowledge

  • Notices and replies

    Compliance: the evidence pack is laid on a timeline and the notice is previewed, reviewed and archived inside the system. Service: a reply is only suggested with a source; coaching looks only at the person’s own behaviour.

    Draft notice · all 3 citations land word for word in the source

  • Keep data in

    The whole chain runs on your machine or network; documents are rated L0–L4 and sensitive ones never leave; the model proposes, never decides; payments, HR and approvals are out of its reach.

    0 internet calls · every document rated before anything leaves

How it works

From a pile of documents to an accountable conclusion in seven steps

Each step’s output carries a position and a fingerprint, and the next step can only use it; the model appears only in steps four and six, and only proposes.

  1. 1

    Drop the documents in

    Any format, any volume; originals are immutable

  2. 2

    Recognise and file

    Two OCR engines cross-checked; one fingerprint per document

  3. 3

    Extract facts

    IDs, amounts, dates, entities, each with its exact position

  4. 4

    Rules and candidates

    Deterministic operators conclude; the model only proposes

  5. 5

    Check and recalculate

    Back to the source line, re-run in front of you

  6. 6

    Multi-model review

    Re-checked in idle time; confidence changes, publication doesn’t

  7. 7

    Conclusion · ticket · notice · answer

    Each with a source, an owner, a due date and a review

The evidence system

How a conclusion is built

Follow it all the way down: to the source, to the recalculation, to the ticket, to the review. If any step fails to line up, the system publishes nothing.

  1. 01

    The conclusion goes back to the source sentence

    Not “these documents were consulted” but which line, which sentence, with the arithmetic re-run in front of you.

  2. 02

    Recalculated live, never read from a cache

    The system re-runs the sums from the numbers in the source; a mismatch with the stated conclusion is flagged in red.

  3. 03

    The model proposes; it never decides

    If the data fails the local check, no conclusion is published. If the system cannot decide, it abstains and hands over to a person.

  4. 04

    An answer is only the start

    Raising a ticket attaches a plan and a due date; closing it requires an outcome, and outcomes go into the knowledge base.

Conclusion

Spend on the same valve model is 34.7% higher this quarter than last, concentrated with a single supplier.

The judgement first, then how it was reached.

Located in the sourceProcurement ledger 2026-Q2 · line 143
Recalculated on the spot(2,318,400 − 1,720,600) ÷ 1,720,600 = 34.7%
CheckMatches
Raised as a ticketOwner: Procurement · Plan: competitive quote review · Due in 10 business days · Outcome required to close

Try it

Pick a question and watch it find the evidence

Three synthetic documents. Pick a question or type one; the chain is built in front of you. If there is no source, it abstains rather than guessing.

Documents (synthetic)

procurement-ledger-2026-Q2.csv

  1. 012026-04-02,Valve DN50,Hongxin,86,412,800
  2. 022026-05-11,Valve DN50,Hongxin,120,576,000
  3. 032026-06-18,Valve DN50,Hongxin,278,1,329,600
  4. 04Subtotal Q2,Valve DN50,Hongxin,484,2,318,400

procurement-ledger-2026-Q1.csv

  1. 012026-01-15,Valve DN50,Hongxin,110,528,000
  2. 022026-02-20,Valve DN50,Hongxin,96,460,800
  3. 032026-03-19,Valve DN50,Yongtai,152,731,800
  4. 04Subtotal Q1,Valve DN50,all,358,1,720,600

supplier-register.xlsx

  1. 01Hongxin Electromechanical · since 2023-06 · contact Wang
  2. 02Yongtai Valves · since 2021-02 · contact Li
  3. 03Note: same model needs quotes from 2+ suppliers (policy 4.3)

Click a line to see which conclusions cite it

or pick a question

Run it on your own documents →

Documents (synthetic)

Payment approval · progress payment 3

  1. 01Project: Road upgrade, stage 1
  2. 02Amount: 6.2m
  3. 03Approval date: 2026-03-04
  4. 04Approved by: Zhou Ming

Construction contract HT-2026-0031

  1. 01Principal: City Investment Group · Contractor: XX Construction
  2. 02Contract signed: 2026-03-16
  3. 03Contract price: 31.0m
  4. 04Clause 5.2: progress payments approved monthly

Meeting minutes 2026-02-27

  1. 01Topic: stage 1 mobilisation
  2. 02Decision: start payments after the contract is signed
  3. 03Present: Zhou Ming, Tang Hua, Zhang Wei

Click a line to see which conclusions cite it

or pick a question

Run it on your own documents →

Documents (synthetic)

Product specs · SKU 4471

  1. 01Item: long-sleeve shirt · blue check
  2. 02Fabric: 100% cotton
  3. 03Weight: 135 g/m²
  4. 04Care: cold machine wash · no bleach

Store wording

  1. 01#12 Fabric per spec sheet; no verbal promises
  2. 02#19 Sizing: ask height and weight first
  3. 03#23 Address change after dispatch: contact carrier, not guaranteed

Conversation · today 10:32

  1. 01Customer: Is this 100% cotton?
  2. 02Customer: Can I machine wash it
  3. 03Customer: 170 cm / 60 kg, which size

Click a line to see which conclusions cite it

or pick a question

Run it on your own documents →

What the product looks like

Four screens, one evidence chain

Documents in, conclusions out, ask anything, see the relationships. A miniature of the product screens; all data synthetic.

Library
DocumentTypeEntityStatus
procurement-ledger-2026-Q2.csvLedgerHongxinChecked
contract HT-2026-0031.pdfContractCity Investment GroupChecked
bank-receipt-0409.jpgReceiptYongtai ValvesRecognising
minutes-2026-02-27.docxMinutesCity Investment GroupChecked
invoice INV-20260518.pdfInvoiceHongxinUnfiled

Scans are recognised before filing; every document has a fingerprint, and a single changed character stops matching.

Conclusions
ConclusionScopeStatus
Same-model valve spend 34.7% higher than last quarterProcurement · HongxinNeeds action
Payment approved 12 days before contract signedRoad upgrade projectNeeds action
Progress payment 3 is 20.0% of contract priceRoad upgrade projectConfirmed
Receipt 0409 matches the ledgerProcurement · YongtaiConfirmed

Hover a conclusion to see its evidence

Hover a row and the evidence sits beside it; what can’t be decided is marked as such, not dressed up.

Ask

Which supplier received the most this quarter?

Hongxin, 2,318,400 in Q2 (3 orders). Basis: ledger 2026-Q2 lines 1–4. Route: totals by entity, no model call.

Are they reliable?

There is no verifiable “reliability” criterion in the documents. What can be said: supplier since 2023-06; three Q2 orders with no competing quotes (conflicts with policy 4.3).

Questions about the system, conclusions, data or common knowledge are answered by deterministic routes first; only then does it retrieve source text, and the model only phrases.

Relationships

Click an entity to see its documents and conclusions; only exact-name matches count.

Why us

Others hand you file links. We hand you the line in the source.

How far can one conclusion be questioned? To the source, to the recalculation, to the ticket, to the review.

The questionThe usual approachFidelog
What is this conclusion based on?A list of “documents consulted”The line and sentence, recalculated in front of you
Is the number right?Check it yourselfThe page re-runs the sum from the source numbers and flags any mismatch in red
Does it work offline?Needs the internet, or “supports on-prem” with gaps in the chainThe whole chain is installed on your hardware and never touches the internet
What if the model is wrong?Sent as-is; someone has to noticeFails the local check, doesn’t go out; can’t decide, abstains and hands to a person
What happens after the answer?Nothing; the answer is the endA ticket with a plan and due date; an outcome is required to close
Does my data have to leave?Upload the lotClassified first; sensitive material never leaves, and what does is the minimum necessary, de-identified
Who is accountable?Nobody owns itEvery conclusion has a source, an owner, a due date and a review
What about another workplace?Buy another product, learn another toolOne foundation, three workplaces; documents come in once

Three workplaces

One evidence chain, three workplaces

Deployed independently, sharing one foundation. Moving to another workplace doesn’t mean learning another tool.

Workplace one

Business management

One screen for the people running the business: see the whole picture, find the issues, drive the actions.

  • The most important changes in the business first
  • The evidence behind each issue and the person who owns it
  • Every next step pushed through to a review

Prototype after sign-inOpen prototypes →

Workplace two

Government compliance review

From a pile of financial records to a findings notice that is ready to issue.

  • Drop the documents in; the anomalies surface themselves
  • The evidence pack is laid out on a timeline
  • The notice is previewed, reviewed and archived inside the system

Prototype after sign-inOpen prototypes →

Workplace three

Online-store customer service

Deal with today’s conversations first, then decide who needs coaching.

  • Risky conversations surface first; every decision can be undone
  • A reply is only suggested when there is a source for it
  • Coaching is based only on the person’s own behaviour

Runs on real customer data · private

Data security

Data stays in; anything leaving is classified first

Security is a rule, not a promise. Every document gets its level and destination on the way in; any page holding real data sits behind a login.

  1. 01

    The full chain runs offline

    Documents, model, evidence and approvals all sit on your machine or your network. There is no internet egress.

  2. 02

    Classify first, then decide where it may go

    Every document is rated L0–L4. The highest levels never leave; anything permitted out is the minimum necessary, de-identified.

  3. 03

    Encrypted in transit and at rest

    Encrypted to approved standards, with keys managed separately from business data; backups are encrypted and access-restricted.

  4. 04

    Clear on who can see what

    Viewing, analysing, restoring and approving are separate roles. Every action leaves an auditable record; logs never store sensitive text.

  5. 05

    Tamper-evident history

    Every document has a fingerprint and every change is chained. Alter one thing and everything after it stops matching.

  6. 06

    The model proposes; it never decides

    Every conclusion passes a local check first. Payments, HR and approvals are actions the model cannot touch.

  • Fully offline

    Government and highly sensitive industries

    Documents, model, evidence and approvals all inside your network; the full chain runs with the internet off.

  • Hosted in your own environment

    Organisations with their own server room or a cloud tenancy they control

    The same system installed where you choose. You can switch it off, audit it and export from it at any time.

  • Hybrid

    Organisations that want an external model quickly

    Sensitive documents and real values stay local; only ordinary text goes to an approved external model, and only when you turn that on.

One foundation

Three workplaces, one foundation

Documents come in once and serve all three. Change the interface, keep the foundation; connecting a new source is a matter of registering it.

  • One document, one fingerprint

    Change a single character and it stops matching; every change is chained on record.

  • Scans are readable too

    Paper contracts and photographed invoices are recognised before they are filed.

  • 790k

    documents ingested and individually checked (synthetic and public corpora).

Frequently asked

The questions you’ll probably ask first

Including the ones that don’t flatter us. Where we can, we say so; where we can’t, we say that too.

Will the model make things up?

Yes, left to itself. That is why every conclusion has to pass a local check: if it can’t be matched to the source it isn’t published; if the system can’t decide, it abstains and hands over to a person. Abstentions and errors are counted separately, and an error is an incident.

Does it replace our ERP or accounting software?

No. It reads your documents, finds the issues and drives the actions. It doesn’t replace your accounting or workflow systems, and it doesn’t do staff performance ratings.

Can we see a demo of the customer service system?

It runs on a client’s real store data, so it isn’t public. You can take an export of your own and run it once in your own environment.

Do we have to hand our data to you?

No. The system is installed on your machine or inside your network. Whether an external model is used at all, and up to which sensitivity level, is your switch.

What hardware do we need?

One machine on your own network is enough to start; add more as the volume grows. The specifics depend on your document volume and workplace; we don’t open with a shopping list.

How soon do we see a first result?

Pick one issue that matters, load a batch of existing documents, and look at the results together. See one real result first, then decide whether to widen the scope.

See one real result first

After sign-in you can open the business-management and compliance-review prototypes. To run it once on your own documents, leave us a contact.

Sign in to prototypesContact us