Step 2 Results: Quote-Backed Hotel Evidence Extraction
Step 1 ended with four frozen, hash-verified source texts for Omni Mount Washington Resort & Spa. Step 2 tested whether a local model could turn those files into a usable evidence ledger without browsing, changing sources, or writing review prose.
The result is a sealed ledger containing 26 accepted quotation-backed claims. Every accepted quotation was independently found as an exact contiguous substring in its assigned local source file. The validator found zero package, source, quotation, quote-hash, subject, or artifact-hash errors in the sealed result.
Result summary
| Status | complete |
|---|---|
| Local model | qwen3.8:latest, digest 22130167c4c20e20c7b71454612966ca8e8171e9b3cc8ab6ce8aa6cbfec79643 |
| Route | LAN Ollama on 10.7.6.39; no cloud model calls |
| Accepted claims | 26 |
| Sources completed | 4 of 4 |
| Model calls | 6 |
| Input / output tokens | 25,950 / 4,609 |
| Observed model time | 210.26 seconds |
| Failed first attempts / retries | 2 / 2 |
| Accepted salvage attempts | 2 |
| Rows discarded without repair | 4 |
| Unresolved conflicts | 2 |
| Freshness qualification required | 7 claims |
The exact Step 2 input package and packet assignments are documented in What Comes After Hotel Source Acquisition: The Next Kanban Cards.
What ran
Packet 001 contained the official property page, Tripadvisor, and U.S. News Travel. Packet 002 contained Hotels.com. The local extractor processed one source at a time inside those fixed packets and requested atomic claims with literal quotations. It had no web access.
Packet 001 required five calls. The first attempt for the official page and the first attempt for U.S. News each contained one or more quotations that were not literal source substrings. Both attempts were preserved as failures. On the bounded second attempts, deterministic salvage retained only valid untouched rows: five official claims and five U.S. News claims. Four malformed rows from those successful salvage attempts were discarded rather than repaired. Tripadvisor passed on its first call with eight claims.
Packet 002 passed on its first call with eight Hotels.com claims.
Claims accepted by source
| Source | Accepted claims |
|---|---|
| Official Omni property page | 5 |
| Tripadvisor property corpus | 8 |
| U.S. News Travel property review | 5 |
| Hotels.com property corpus | 8 |
The complete normalized claim output
These are the local model’s normalized claims, not review prose. The sealed local ledger also retains each exact quotation, source ID, packet ID, quote hash, source hash, and validation fields.
| Claim ID | Source | Normalized claim | Freshness policy |
|---|---|---|---|
OFFICIAL-1-001 |
Official Omni property page | The resort features a 25,000 square-foot spa and wellness center. | Stable |
OFFICIAL-1-002 |
Official Omni property page | The resort offers 30,000 square feet of space for hosting events. | Stable |
OFFICIAL-1-003 |
Official Omni property page | The resort's golf course was designed by Donald Ross. | Stable |
OFFICIAL-1-004 |
Official Omni property page | The resort features Spanish Renaissance architecture. | Stable |
OFFICIAL-1-005 |
Official Omni property page | Presidential rooms provide exclusive access to a lounge with daily continental breakfast. | Qualification required |
GUEST-1-001 |
Tripadvisor property corpus | The hotel has 284 rooms. | Stable |
GUEST-1-002 |
Tripadvisor property corpus | The hotel is located at 310 Mount Washington Hotel Rd, Bretton Woods, NH 03575. | Stable |
GUEST-1-003 |
Tripadvisor property corpus | The hotel features two golf courses. | Stable |
GUEST-1-004 |
Tripadvisor property corpus | The hotel offers a heated indoor pool. | Stable |
GUEST-1-005 |
Tripadvisor property corpus | The hotel provides free high-speed internet (WiFi). | Stable |
GUEST-1-006 |
Tripadvisor property corpus | The hotel has an electric vehicle charging station. | Stable |
GUEST-1-007 |
Tripadvisor property corpus | The hotel is classified as a 4-star property. | Stable |
GUEST-1-008 |
Tripadvisor property corpus | The hotel offers allergy-free rooms. | Stable |
GUEST-2-001 |
U.S. News Travel property review | The hotel has 269 rooms. | Stable |
GUEST-2-002 |
U.S. News Travel property review | The property features a 25,000-square-foot spa. | Stable |
GUEST-2-003 |
U.S. News Travel property review | The resort includes an 18-hole golf course. | Stable |
GUEST-2-004 |
U.S. News Travel property review | The resort has an indoor heated pool. | Stable |
GUEST-2-005 |
U.S. News Travel property review | Pets weighing 60 pounds or under are allowed for a $150 fee. | Qualification required |
GUEST-3-001 |
Hotels.com property corpus | The hotel has 269 units arranged over 6 floors. | Stable |
GUEST-3-002 |
Hotels.com property corpus | The resort features 9 restaurants and 3 bars/lounges. | Qualification required |
GUEST-3-003 |
Hotels.com property corpus | The spa includes 13 treatment rooms, a sauna, a hot tub, and a steam room. | Stable |
GUEST-3-004 |
Hotels.com property corpus | A mandatory resort fee of USD 48.83 per accommodation per night is charged. | Qualification required |
GUEST-3-005 |
Hotels.com property corpus | Uncovered self parking costs USD 20 per night. | Qualification required |
GUEST-3-006 |
Hotels.com property corpus | Pets are allowed for an extra charge of USD 217 per accommodation per stay. | Qualification required |
GUEST-3-007 |
Hotels.com property corpus | The pool is accessible from 6:00 AM to 10:00 PM. | Qualification required |
GUEST-3-008 |
Hotels.com property corpus | The hotel is located at 310 Mt Washington Hotel Rd, Bretton Woods, NH, 03575. | Stable |
Coverage produced by the extraction
| Topic | State | Accepted claims |
|---|---|---|
amenities_recreation |
present | 10 |
dining |
present | 1 |
events |
present | 1 |
operations_policies |
present | 3 |
rooms_identity |
present | 6 |
Zero claims would have produced unknown, never an unsupported assertion of absence. Six additional accepted claims fell into the conservative other_property_fact bucket rather than being forced into an ill-fitting topic.
Two conflicts were preserved, not resolved
Room count
Tripadvisor reported 284 rooms. U.S. News and Hotels.com each reported 269 rooms or units. The ledger marks the values 269 and 284 as an unresolved identity/room-count conflict and preserves all three source rows.
Pet fee
U.S. News reported a $150-plus-tax pet fee for pets weighing 60 pounds or less. Hotels.com reported a $217 charge per accommodation per stay. The ledger marks $150 and $217 as an unresolved, time-sensitive policy conflict. Neither value is eligible for unqualified drafting.
Freshness policy caught more than the model did
The model marked four claims as time-sensitive. A deterministic policy pass expanded that to seven claims requiring date-and-source qualification by detecting currency amounts, operating schedules, and current-frequency language. That additional gate caught, among other things, pool-access hours and a daily lounge offering that the model had marked as stable.
A controller bug was exposed and fixed
The first packet originally emitted a blocked receipt even though it had 18 valid accepted claims and no terminal source failure. The controller was counting quotation errors from superseded first attempts against final completion. That contradicted its own bounded-retry and deterministic-salvage contract.
The failed receipt remains preserved. The extraction skill was patched so completion considers quotation errors in the accepted attempt while separately reporting errors from earlier attempts. A fresh deterministic sealing pass then revalidated every package hash, source hash, quotation, quote hash, subject ID, packet assignment, and artifact hash without another model call.
Sealed artifacts
- Evidence ledger SHA-256:
17f95706d738fe78c0bb4ada260884b431969d8e3600c2cb893bcaf145d3171e - Validation SHA-256:
ab8ecb4c0e7a74baeaaab3088701068c6967a4cc965aeac399c8e30126a7b79b - Coverage SHA-256:
fb3683f2827066fd992890c0facd23440cd3ecc9a983e0617bb5e4d252a5b63b
What this establishes—and what it does not
This trial establishes that the local extraction boundary can produce a nonempty, quotation-backed, deterministically verified ledger from the frozen Step 1 package. It also demonstrates why extraction cannot be treated as fact adjudication: the model faithfully surfaced incompatible room counts and pet fees, but a later resolution step must decide which current source is authoritative.
The next work is therefore narrow: resolve the room-count discrepancy, refresh current fee and policy evidence, and qualify the same extraction contract across additional hotel classes. Drafting should not begin from card completion alone; it should begin only from the sealed ledger after conflicts, freshness, and required coverage gates are satisfied.
