OpenAI Build Week · Apps for Your Life
See what the deal says—and what it doesn’t.
Dealer pricing is often scattered across advertisements, messages, worksheets, fees, rebates, and financing conditions. TermsTrail turns those fragments into an evidence-linked trail so the buyer can see what is documented, what is missing, and what needs to be clarified.
GPT-5.6 integration verified through bounded synthetic qualification. The public demo replays sanitized evidence without making a new API request.
Deal trailReview required
82%Evidence Coverage
$2,220Unexplained
3Unresolved conditions
Supported total$32,490
Stated total$34,710
Missing3 conditions
Unexplained$2,220
StatusReview required
ridgeway-buyer-worksheet.pdf · Page 1Stated $34,710Calculated $32,490Difference · $2,220
◆ Synthetic demonstration data
30-second judge path
The problem
A price is often spread across several versions of the conversation.
A dealer may provide an advertised price, a monthly payment, a worksheet, and a later explanation without ever presenting all assumptions in one place. TermsTrail preserves each version and distinguishes a missing term from a documented term.
A summary tells you what a document says.An evidence trail shows what every source actually supports.
Interactive synthetic demonstration
Follow the trail as the deal changes.
Four fictional sources. One deterministic recalculation. Every finding stays connected to the evidence that supports it.
DealerRidgeway Motors
Vehicle2026 Meridian LX
ReferenceDEMO-VEHICLE-001
BuyerJordan Sample
Fictional caseSource material2 sources
2026 Meridian LX
$28,480Advertised cash selling price
$599 per month · $2,500 due at signing“The $28,480 cash price includes the $1,500 Meridian Customer Bonus. Ridgeway financing, a trade-in, loyalty, military status, and residency are not required.”
TermsTrail resultNOT COMPARABLE YET
More terms neededKnown
- A $28,480 cash selling price is documented.Advertisement + dealer message
- A $599 payment with $2,500 due at signing is documented.Advertisement · payment line
Missing
Written out-the-door totalItemized dealer chargesAPRNumber of payments
Stage 1 of 4
Three-minute walkthrough
Watch the complete evidence loop.
The narrated recording follows the real no-login product from the fictional Ridgeway sources through the revised trail and measured evaluation. The screen capture uses committed synthetic results and makes no live product-route request. Its OpenAI-generated marin voice used one separately authorized Speech API request.
Product evidence gallery
The exact screens behind the story.
These production-build captures are hash-linked to the same synthetic release candidate shown in the walkthrough.
How it works
From fragmented claims to a traceable deal record.
Collect the sources
PDFs, screenshots, worksheets, advertisements, and pasted messages.
Evidence is interpreted
Committed mock extraction powers this sample; a bounded, opt-in GPT-5.6 adapter handles authorized synthetic sources.
Deterministic code verifies
TypeScript recalculates totals, measures Evidence Coverage, enforces hard gates, validates evidence references, and compares revisions.
The buyer decides
TermsTrail prepares clarification questions but never sends messages or makes financial, legal, or contract-safety decisions.
Technology
AI interprets. Code calculates. The buyer decides.
AI
Bounded interpretation layer
GPT-5.6 · opt-in
- Reads PDFs, screenshots, and natural-language messages
- Associates differently worded amounts with their meaning
- Detects conditional or ambiguous language
- Preserves page and excerpt references
- Produces schema-constrained evidence
- Returns source-linked clarification points for deterministic assembly
TS
Verification layer
Deterministic TypeScript
- Normalizes currency
- Recalculates totals
- Computes unexplained differences
- Measures Evidence Coverage
- Enforces critical-field gates
- Compares revisions
- Rejects unsupported statuses
- Validates evidence references
Private synthetic intake lab
Mixed-source verification without silent assumptions.
The product can assemble one evidence trail from pasted text, a PDF, a JPEG or PNG, and one later response. The no-login sample stays fully offline; the optional interpreter remains disabled unless a local operator explicitly enables it.
- One to four sources, counted by the server.
- Exact synthetic-only confirmation before processing.
- Every source is preflighted before any provider call.
- Cached extraction; no paid calls from CI or page load.
Built with Codex
One standalone project, built source-first.
Codex was used to build the standalone repository, structured schemas, extraction flow, deterministic verification modules, synthetic test fixtures, interface, and evaluation framework.
Vision and mission
Clarity before commitment.
No guesses. Every material finding points back to its source.
Every person can verify the terms of a major purchase before committing to it.
Turn fragmented seller claims and purchase documents into an evidence-linked trail that shows what is documented, missing, conditional, unexplained, or changed—then help the buyer ask the next precise question without acting on the buyer’s behalf.
Measured evaluation
Synthetic ground truth, published with the misses.
Twenty committed fictional fixtures compare hand-authored expected results with mock-provider outputs and deterministic verification. This offline suite made zero live calls and met every stated target.
Structured-output validity
100% · 20/20
Deterministic arithmetic correctness
100% · 20/20
Critical-field extraction accuracy
99.7% · 331/332
Finding-category precision
97.44% · 38/39
Finding-category recall
100% · 38/38
Source-reference resolvability
100% · 20/20
Clarification-to-finding linkage
100% · 20/20
End-to-end fixture agreement
95% · 19/20
Known limitation
These scores measure committed synthetic sources and mock-provider outputs—not live-model or OCR accuracy. One fixture deliberately preserves a known APR extraction miss so failure propagation remains visible.
Offline suite
20 synthetic fixtures
20/20Schema valid
20/20Arithmetic correct
19/20Outcome agreement
One mock extraction deliberately omits a source-stated APR. The downstream trail correctly raises a missing-APR finding, which keeps the suite from presenting a falsely perfect outcome story.
Scope: committed synthetic sources and mock outputs. Not a live-model or visual-OCR accuracy claim.
Recorded GPT-5.6 qualification
The corrected hero passed every gate.
17Source-linked claims
9/9Hero gates
$2,220Verified gap
$0.08149875Hero estimate
The request used the exact gpt-5.6 alias and returned gpt-5.6-sol. Every claim resolved to supplied fictional evidence. Deterministic TypeScript—not the model—reconciled $32,490 supported against $34,710 stated.
This is a saved synthetic result. Replaying it makes no API request and is not a completed broad live evaluation.
Narrow correction passed
Four requests used; one remained unspent.
The original three-request mission stopped on the sparse dealer message after a cent-normalization miss. One separately authorized v3 request then rechecked only that fictional message, returned gpt-5.6-sol, and passed all 6 gates for an estimated $0.01952000. The broader live suite was not resumed, and there were zero unsupported or hallucinated claims. The one unused historical cap slot is not authorization; zero further provider requests are authorized.
- $0.20196750Cumulative estimated cost
- 4 of 5Project GPT requests used
- v3 correction passedOne fixture verified; broader suite not resumed
Safety and product boundaries
Designed to clarify evidence—not make the decision.
Synthetic demonstration
No real customer, dealer, credit, or identity information is used on this Site.
No automatic communication
TermsTrail prepares copyable questions but does not contact a dealership.
No legal or financial advice
TermsTrail does not determine whether a charge, contract, price, or financing term is legal, fair, advisable, fraudulent, or safe.
No payment or credit processing
The project does not process payments, financing applications, credit checks, or identity verification.
No hidden assumptions
Unknown information remains labeled as missing or unresolved.
Build Week provenance
A new standalone Build Week project.
TermsTrail was created as a standalone project during OpenAI Build Week. OTDZEN is a pre-existing private vehicle-buying business and application. No OTDZEN source files, prompts, database migrations, interface components, routes, copy, branding, customer data, or production integrations were copied into TermsTrail.
Prior experience informed several general engineering principles used in the new project: structured model outputs, deterministic arithmetic, evidence-linked findings, explicit uncertainty, and buyer-controlled drafts.
All TermsTrail implementation, synthetic fixtures, tests, interface work, and evaluation materials were created in the new Build Week repository.
Follow every number to its source.
TermsTrail does not tell the buyer whether a deal is good. It shows what the seller’s own material actually supports, so the buyer knows what to clarify before committing.