promiseledger

THE CUSTOMER REALITY CHECK

Promises made. Truth checked.

Engineering closed the ticket. Can the customer actually use it? Promise Ledger checks every customer promise against customer-specific evidence, shows the gap, and drafts an honest update for you to approve.

No sign-up. Four fictional sample accounts. Live runs are rate-limited; the reference replay is always free.

Powered by NVIDIA Nemotron on Nebius Token Factory

PL-101 EVIDENCE BRIEFDelivery gap

Audit log export

Northstar · promised by Maya Chen · due Sep 14

Maya Chen: I will make audit log export available to Northstar by 2026-09-14.Meeting · SRC-01
  • Built
  • Enabled
  • Verified
Publicly GA ≠ usable by this customerVendor changelog: “Audit log export is generally available (2026-09-12).”

THE REALITYThe customer-specific audit_export entitlement is disabled. Don’t send the delivery email yet.

Draft for your review. Never sent automatically.

THE PROBLEM

“Done” in engineering is not delivered to the customer.

A customer-success manager has to answer one question honestly: can I tell this customer we delivered what we promised? The systems they read each tell a different part of the story.

Engineering ticket
“ENG-428 audit log export: implementation complete; built=true.”

The ticket closed. Its definition of done stops at the merge, not at this customer’s workspace.

Public changelog
“Audit log export is generally available.”

General availability says nothing about one customer’s plan, entitlement or feature flag.

The customer
“The audit export button is still missing in our Northstar workspace.”

Send the “it’s live” email now and you spend trust you can’t easily win back.

THE APPROACH

Models read. Rules decide. People send.

Models read

NVIDIA Nemotron models triage the sources, extract promises and explain disagreements. Every claim cites source text, checked character for character.

Rules decide

Built, enabled and customer-verified are separate facts. A deterministic policy sets every verdict, and no model can change one.

People send

Drafts are editable and approved locally. Export exists; sending does not. A closed ticket never earns a green badge on its own.

HOW IT WORKS

One run. Four steps. Three Nemotron models.

The live agent view streams every step as it happens: the model, latency, tokens, estimated cost, and each cited quote with its exact-match check.

  1. 01

    Triage

    Nemotron 3 Nano

    Labels every source and routes conversations to extraction. Delivery records go straight to the rules.

  2. 02

    Extract

    Nemotron 3 Super

    Pulls out commitments, owners and dates, each with an exact quote. Anything it cannot ground is dropped; a failed extraction stops the run.

  3. 03

    Decide

    Deterministic rules · no model

    An ordered policy turns customer-specific built, enabled and verified facts into one of seven conservative verdicts.

  4. 04

    Explain

    Nemotron 3 Ultra

    Only after the verdict is fixed: why the evidence disagrees, a customer update and a nudge for the owner, each claim cited.

Evidence the rules can weigh

Sample packs or your own evidence

Meeting notes, tickets and availability snapshots. Paste text or drop .txt, .md, .csv or .eml files; you confirm every extracted fact before the rules read it.

Public claims · Tavily Extract

Fetches the vendor’s public changelog at runtime and shows it beside the verdict: “Publicly GA ≠ usable by this customer”. It never changes the verdict.

Runtime errors · error monitoring

Errors for this customer and feature after acceptance lower “verified” to “needs verification”. A quiet error feed never proves delivery.

GUARDRAILS

Checked before you ever see a draft.

  • Exact quotes onlyEvery cited quote is matched against its source as it streams. “Not in source” is shown, not hidden.
  • No new dates or promisesCustomer updates cannot add a date, a timeframe or “will be enabled” that the evidence doesn’t contain.
  • No verdict changesUltra writes after the rules decide, with no verdict field to fill in. Contradicting the facts is rejected.
  • Visible fallbacksA brief that fails a check becomes a labelled template draft, with the exact reason on screen.
  • Bounded costEach live run reserves its worst case before any model call, under a $0.05 per-run ceiling and shared rate limits.
  • Nothing sendsNo email or CRM write exists. Approval is local, and export is the only way out.

POWERED BY

NVIDIA Nemotron on Nebius Token Factory.

Three open NVIDIA Nemotron models, served by Nebius Token Factory’s inference API, each do one job. The browser never sees a key; every call runs server-side with a strict schema and a budget.

Full live run
16–20 s
Estimated cost per run
≈ $0.01
Frozen held-out cases
32 / 32

Measured in September 2026 on synthetic data: a four-run pipeline smoke test and a small assistant-authored evaluation of the Super extraction step. Not a real-world accuracy claim.

  • TriageNemotron 3 Nanonvidia/NVIDIA-Nemotron-3-Nano-30B-A3B
  • ExtractNemotron 3 Supernvidia/nemotron-3-super-120b-a12b
  • ExplainNemotron 3 Ultranvidia/Nemotron-3-Ultra-550b-a55b

TRY IT

Four fictional customers. One reality check.

Honest about what this is

Every customer, person and document in the samples is fictional, and snapshots are fixed at September 13, 2026. No real CRM, support desk or customer system is connected. Reference mode replays a hand-labelled trace with no AI calls and says so on screen. Evidence you bring stays in your browser until you run a check; a live run sends it to Nebius for inference.