SECOND PHYSICAL MACHINE ยท WORKFLOW 1.5.6
Workflow 1.5.6 reproduced on an M4 Mac mini
document-reproduction-v156-m4-20261018.1. Measurements completed 2026-10-19T03:23:21Z.
The unchanged workflow passed the frozen gates on a dedicated rented Mac mini with Apple M4, 10 CPU cores and 16 GiB unified memory, running macOS 27.0.1 build 26A434.
We ran 40 public question clusters three times in fresh model-server sessions: 120 original attempts. The bank contains five English selectable-text PDFs, 30 pages, 30 answerable questions and 10 absent-information questions per repetition. This measures reproduction on a second physical machine using a known public bank.
Observed results
| Session | Attempts | Supported answers | Supported citations | Clean abstentions | Frozen gate |
|---|---|---|---|---|---|
| 1 | 40/40 | 30/30 | 35/35 | 10/10 | Pass |
| 2 | 40/40 | 30/30 | 35/35 | 10/10 | Pass |
| 3 | 40/40 | 30/30 | 35/35 | 10/10 | Pass |
All 120 original rows completed with the frozen pass result. No errors, timeouts or incomplete rows were observed.
Observed median complete answer time: 8.70 seconds; p95: 18.45 seconds; maximum: 18.91 seconds across 120 attempts. These include retrieval and the model selection/sufficiency phases. PDF import/indexing and process/model startup are separate.
Fresh-process startup to authenticated health took 1.55, 1.54, 1.54 seconds. Sampled server peak RSS was 4.66, 4.66, 4.66 GiB by session. RSS does not measure total unified-memory use. OS file caches were not claimed flushed; wall power was not measured.
The original runtime logs did not expose accelerator placement. A one-second read-only process sample during session three recorded the model graph-computation call chain through Metal synchronization and Metal command-buffer completion. The exact offloaded layer count remains unmeasured. That observer's overhead stayed inside the original clocks; no question was repeated for it.
What this adds
This adds a measured reproduction profile for this M4/16 GiB/macOS 27.0.1 configuration, with the pinned Qwen3-4B-Q4_K_M weights, Python 3.12.14, pypdf 6.10.0 and llama.cpp b11429/d81235049. All 25 published package members remained unchanged. The actual official-release runtime binary differs from the original Mac's Homebrew binary; its hash and observed backend are recorded.
The fresh installation passed all four public self-tests. The server listened only on 127.0.0.1:18482. The model API returned 401 for missing and incorrect keys and 200 for its private key. The declared external IPv4, external IPv6 and wrong-loopback-port connection probes returned policy-denial errno 1 under the supplied macOS profile. These checks describe that declared boundary.
What we can prove against LM Studio
The preserved v24 native LM Studio comparison supports a narrow citation-support advantage for Kingy on Q11 and Q29, two unique cross-page tasks repeated three times on the original M4 Pro Mac. LM Studio supplied correct factual values but claimed chunks that omitted the requested facts. Its answers were shorter and complete on five other cases: Q09, Q13, Q15, Q19 and Q23. Both workflows had correct requested factual values on all 30 answerable cases in that study's independent source review.
This second-machine run tests Kingy alone. It adds no new LM Studio measurements and establishes no overall competitor, speed, privacy or cost winner.
Scope and setup
The rental is dedicated physical cloud-hosted hardware operated by the same Kingy operator. The public synthetic corpus was transferred to it. This run does not establish an independent external lab result or retention of private documents on a user-owned machine. Kingy paid for the rental; there was no provider sponsorship.
All 30 answerable questions draw their facts from one PDF each; five use adjacent pages. OCR, scanned pages, private business documents, cross-document synthesis, long corpora, other chips and other software versions remain unmeasured. Forty question clusters repeated three times are not 120 independent questions. The frozen rubric and source review do not establish novice-user acceptance.
The original Mac used M4 Pro/24 GiB and macOS 26.3. Hardware, OS, runtime binary and background environment differ, so cross-machine timing does not isolate a hardware effect. The immutable ZIP README describes the original declared machine and includes a generic multi-document sentence; the measured v24 bank has no cross-document synthesis. The evaluator also retains a historical scoring-label string; the actual graded rows identify kingy-common-document-grade/1.5.0, with the unchanged scorer hash below.
Setup dependencies and weights were prepared before inference. One prerequisite observation failed because the ZIP was not adjacent to the checker; the layout was corrected and a fresh observation passed. The original Mac's private identity record stayed local; the two host hashes were compared locally before inference. No original document answers were retried or tuned. The prepared protocol predates the later approved paid rental; the workflow, corpus, model and scoring pins stayed fixed.
Reproduce the measured workflow
Use the exact versions and settings in the configuration receipt, then extract the unchanged tested ZIP into a fresh folder. Run the four self-tests and the packaged evaluator with a new output directory. Keep the local API key private and review citations on your own documents. Other configurations need their own measurements.
ZIP SHA-256: 3be3b7da503894be99829496b6aa8ce38bb635a5a5c86916f694ec7e4eeea396 Runner SHA-256: 3bcc03ffef818fa7fb6062ae17f20aa09d2fcc25fc9d57572c77ec35bc1ef02b Scorer SHA-256: e9b95ce6b5a821c0ab703c60de388f7604d919d62cfe13ac544cfbcb95c14c5f
Inspect every answer
Every distinct actual answer is shown with its original repetition labels and claimed source pages. Longer source passages are a deliberate part of the extractive workflow; shorter answers remain a usability tradeoff.
V24-Q01: On what date does the current Larkspur packing roster take effect?
Repetitions [1, 2, 3]. Frozen result: pass.
- roster_effective_date: The current packing roster takes effect on 7 April 2028. larkspur-seed-dispatch, page 1
V24-Q02: Who is Larkspur's dispatch coordinator?
Repetitions [1, 2, 3]. Frozen result: pass.
- dispatch_coordinator: Dispatch coordinator: Mira Dellen. larkspur-seed-dispatch, page 1
V24-Q03: What release thresholds does Larkspur specify for ordinary-tray moisture and cooled-tray core temperature?
Repetitions [1, 2, 3]. Frozen result: pass.
- release_thresholds: The release thresholds are 38 percent moisture for ordinary trays and 7 degrees Celsius core temperature for cooled trays. Moisture is read from the centre plug of an ordinary tray. larkspur-seed-dispatch, page 2
V24-Q04: What is the interval between ordinary Larkspur dispatches?
Repetitions [1, 2, 3]. Frozen result: pass.
- dispatch_interval: The ordinary dispatch interval is every 35 minutes. This interval describes the planned spacing between ordinary dispatches. larkspur-seed-dispatch, page 3
V24-Q05: For Larkspur's weekend consolidation run, what is the handoff point, and what is the manifest deadline measured from attachment of the green seal?
Repetitions [1, 2, 3]. Frozen result: pass.
- handoff_point: For the weekend consolidation run, the handoff point is Marten Gate. larkspur-seed-dispatch, page 3
- manifest_deadline: Weekend manifest continuation Packing and dispatch handbook The manifest reaches the desk within 14 minutes after the green seal is attached. The seal is attached only after the weekend trolley is closed. A spare green seal remains in the dispatch drawer until it is needed; taking a spare seal from the drawer does not start the manifest window. larkspur-seed-dispatch, page 4
V24-Q06: What is Larkspur's return handling charge, including the unit charged and its tax treatment?
Repetitions [1, 2, 3]. Frozen result: pass.
- return_handling_charge: The return handling charge is 18.50 Canadian dollars per insulated crate, excluding tax. This handling charge is for an insulated crate returned through the cooperative's service. larkspur-seed-dispatch, page 5
V24-Q07: How far before a visit does Briar Quay's current advance-booking window extend?
Repetitions [1, 2, 3]. Frozen result: pass.
- advance_booking_window: The current advance-booking window is up to 12 calendar days before the visit. briar-quay-listening, page 1
V24-Q08: Who is the booking lead at Briar Quay Sound Archive?
Repetitions [1, 2, 3]. Frozen result: pass.
- booking_lead: Booking lead: Tessa Venn. briar-quay-listening, page 1
V24-Q09: What durations does Briar Quay use for the leader segment and the calibration segment when preparing an access reel?
Repetitions [1, 2, 3]. Frozen result: pass.
- preparation_durations: A reel with an uncertain label stays on the review shelf. The preparation sequence uses 90 seconds for the leader segment and 24 seconds for the calibration segment. The leader segment gives the worker a quiet lead-in before the recorded material begins. The calibration segment is the short reference used to set the listening chain. briar-quay-listening, page 2
V24-Q10: What is Briar Quay's listening ceiling, and where is it measured?
Repetitions [1, 2, 3]. Frozen result: pass.
- listening_ceiling: The listening ceiling is at most 62 decibels at the listening seat. briar-quay-listening, page 3
V24-Q11: For a reserved Briar Quay listening session, which return sleeve is used, and what is the seat-reset deadline measured from the return of the headphones?
Repetitions [1, 2, 3]. Frozen result: pass.
- return_sleeve: For a reserved listening session, the return sleeve is the amber sleeve. briar-quay-listening, page 3
- seat_reset_deadline: Session reset continuation Listening-room booking guide The seat is reset within 9 minutes after the headphones are returned. Returning the case alone does not start this reset window. The host first receives the headphones, then checks the seat, control surface and cable. briar-quay-listening, page 4
V24-Q12: What does Briar Quay charge for its listening-copy service, including the unit charged and tax treatment?
Repetitions [1, 2, 3]. Frozen result: pass.
- listening_copy_charge: The listening-copy service charge is 6.25 Canadian dollars per analogue reel, including tax. briar-quay-listening, page 5
V24-Q13: What is the current standard loan period at Kestrel Reach Tool Library?
Repetitions [1, 2, 3]. Frozen result: pass.
- standard_loan_period: The current standard loan period is 11 calendar days. Repair contact: Nalin Orr. The archived spring leaflet described a standard loan period of 16 calendar days. kestrel-reach-loans, page 1
V24-Q14: Who is the repair contact at Kestrel Reach Tool Library?
Repetitions [1, 2, 3]. Frozen result: pass.
- repair_contact: Repair contact: Nalin Orr. kestrel-reach-loans, page 1
V24-Q15: What jaw gaps does Kestrel Reach use to inspect bench clamps and floor clamps?
Repetitions [1, 2, 3]. Frozen result: pass.
- clamp_jaw_gaps: The inspection settings are 4 millimeters jaw gap for bench clamps and 9 millimeters jaw gap for floor clamps. These settings allow the worker to check the two clamp styles consistently. A floor clamp does not use the bench-clamp setting just because it fits on the inspection bench. The jaw gap is read at the pad faces. kestrel-reach-loans, page 2
V24-Q16: What finishing-drum cap does Kestrel Reach apply, including its booking basis?
Repetitions [1, 2, 3]. Frozen result: pass.
- finishing_drum_cap: The finishing-drum cap is 3 drums per booking. kestrel-reach-loans, page 3
V24-Q17: For Kestrel Reach's closed-counter returns, what is the receiving location, and what is the inspection-start deadline measured from removal of the red tag?
Repetitions [1, 2, 3]. Frozen result: pass.
- receiving_location: For closed-counter tool returns, the receiving location is Bay Fenn. kestrel-reach-loans, page 3
- inspection_start_deadline: The return inspection begins within 27 minutes after the red tag is removed. The tag stays on an item while it waits for a receiving worker. Moving an item between trolleys is not the start of the inspection window. kestrel-reach-loans, page 4
V24-Q18: What is Kestrel Reach's replacement charge for missing fasteners, including the unit charged and tax treatment?
Repetitions [1, 2, 3]. Frozen result: pass.
- fastener_replacement_charge: The replacement charge is 4.80 Canadian dollars per missing fastener, including tax. kestrel-reach-loans, page 5
V24-Q19: On what date does Morrow Vale's current sample workflow take effect?
Repetitions [1, 2, 3]. Frozen result: pass.
- workflow_effective_date: The current sample workflow takes effect on 19 August 2028. Sample-room owner: Riva Pell. The archived workflow took effect on 3 May 2026. morrow-vale-textiles, page 1
V24-Q20: Who owns the sample room at Morrow Vale Textile Lab?
Repetitions [1, 2, 3]. Frozen result: pass.
- sample_room_owner: Sample-room owner: Riva Pell. morrow-vale-textiles, page 1
V24-Q21: What working temperatures does Morrow Vale specify for the rinse bath and the drying cabinet?
Repetitions [1, 2, 3]. Frozen result: pass.
- working_temperatures: The working settings are 18 degrees Celsius for the rinse bath and 43 degrees Celsius for the drying cabinet. The bath setting belongs to the rinse stage and the cabinet setting belongs to the drying stage. They are not interchangeable target temperatures. morrow-vale-textiles, page 2
V24-Q22: What is the batch observation interval during an active Morrow Vale comparison run?
Repetitions [1, 2, 3]. Frozen result: pass.
- batch_observation_interval: The batch observation interval is every 8 hours. The interval is the planned spacing between observations during an active comparison run. morrow-vale-textiles, page 3
V24-Q23: For Morrow Vale's freshly blotted comparison swatches, what is the holding location, and what is the reweighing deadline measured from the end of blotting?
Repetitions [1, 2, 3]. Frozen result: pass.
- holding_location: For freshly blotted comparison swatches, the holding location is Rack Juniper. morrow-vale-textiles, page 3
- reweighing_deadline: Reweighing continuation Washfastness sample workflow The swatches are reweighed within 22 minutes after blotting ends. The blotting step ends when the technician lifts the final absorbent sheet. Placing unused sheets on the bench does not begin the reweighing window. The technician notes the end of blotting on the batch card and carries the card to the balance with the swatches. morrow-vale-textiles, page 4
V24-Q24: What is Morrow Vale's comparison service charge, including the unit charged and tax treatment?
Repetitions [1, 2, 3]. Frozen result: pass.
- comparison_service_charge: The comparison service charge is 12.75 Canadian dollars per fabric swatch, excluding tax. The charge belongs to a fabric swatch accepted for the comparison service. morrow-vale-textiles, page 5
V24-Q25: What is the current public-group cap at Tamarind Ridge Observatory?
Repetitions [1, 2, 3]. Frozen result: pass.
- public_visitor_cap: The current public-group cap is 14 visitors. tamarind-ridge-visitors, page 1
V24-Q26: Who is Tamarind Ridge's night-host coordinator?
Repetitions [1, 2, 3]. Frozen result: pass.
- night_host_coordinator: Night-host coordinator: Oren Sable. tamarind-ridge-visitors, page 1
V24-Q27: What durations does Tamarind Ridge use for the ordinary exposure and the dark reference in the visitor camera sequence?
Repetitions [1, 2, 3]. Frozen result: pass.
- camera_durations: The camera sequence uses 75 seconds for the ordinary exposure and 18 seconds for the dark reference. The ordinary exposure records the demonstration view. tamarind-ridge-visitors, page 2
V24-Q28: What is Tamarind Ridge's interval between visibility records during a visitor night?
Repetitions [1, 2, 3]. Frozen result: pass.
- visibility_record_interval: The visibility-record interval is every 26 minutes. The interval is the planned spacing between visibility notes during a visitor night. tamarind-ridge-visitors, page 3
V24-Q29: For Tamarind Ridge's final guided group, what is the muster location, and what is the shutter-log deadline measured from the closing of the shutter?
Repetitions [1, 2, 3]. Frozen result: pass.
- muster_location: For the final guided group, the muster location is North Cairn. tamarind-ridge-visitors, page 3
- shutter_log_deadline: Shutter log continuation Visitor night field memo The shutter log is completed within 16 minutes after the shutter closes. The host starts the log window from the actual closing of the shutter. Calling the group together does not start that window. tamarind-ridge-visitors, page 4
V24-Q30: What is Tamarind Ridge's lamp service charge, including the unit charged and tax treatment?
Repetitions [1, 2, 3]. Frozen result: pass.
- lamp_service_charge: The lamp service charge is 9.40 Canadian dollars per borrowed red lamp, including tax. The charge applies to each borrowed red lamp. It is not a group-admission charge and does not describe the sale price of a lamp. tamarind-ridge-visitors, page 5
V24-Q31: What insurance deductible applies to Larkspur Seed Cooperative's delivery vehicles?
Repetitions [1, 2, 3]. Frozen result: pass.
Clean decline with no answer fields or citations.
V24-Q32: What is the birth date of the founder of Larkspur Seed Cooperative?
Repetitions [1, 2, 3]. Frozen result: pass.
Clean decline with no answer fields or citations.
V24-Q33: What is Briar Quay Sound Archive's annual recording-licensing budget?
Repetitions [1, 2, 3]. Frozen result: pass.
Clean decline with no answer fields or citations.
V24-Q34: What is the part number of the ventilation filter in Briar Quay's listening room?
Repetitions [1, 2, 3]. Frozen result: pass.
Clean decline with no answer fields or citations.
V24-Q35: What is the postal address of Kestrel Reach Tool Library's battery supplier?
Repetitions [1, 2, 3]. Frozen result: pass.
Clean decline with no answer fields or citations.
V24-Q36: Which company provides overnight security for Kestrel Reach Tool Library?
Repetitions [1, 2, 3]. Frozen result: pass.
Clean decline with no answer fields or citations.
V24-Q37: What is Morrow Vale Textile Lab's wastewater permit number?
Repetitions [1, 2, 3]. Frozen result: pass.
Clean decline with no answer fields or citations.
V24-Q38: What is the serial number of Morrow Vale's balance?
Repetitions [1, 2, 3]. Frozen result: pass.
Clean decline with no answer fields or citations.
V24-Q39: What is the insured value of Tamarind Ridge Observatory's meteorite collection?
Repetitions [1, 2, 3]. Frozen result: pass.
Clean decline with no answer fields or citations.
V24-Q40: What coating thickness is specified for Tamarind Ridge Observatory's telescope mirror?
Repetitions [1, 2, 3]. Frozen result: pass.
Clean decline with no answer fields or citations.
Evidence downloads
Public derivatives retain every original question attempt and model-request record while removing rented test-root paths. Original hashes identify the retained private originals. No authentication keys, SSH connection details, machine fingerprints, billing details or private documents are published.
- Qualification decision
- Exact observed configuration and backend
- Three-session results and timings
- All 120 original attempts, public derivative
- Every-attempt source checks
- Operator review of every unique answer
- Fresh authentication, network and self-test checks
- All four self-test originals, public derivative
- Unchanged evaluator freeze
- All 25 package-member hashes
- Prepared reproduction protocol
- All 40 questions
- Frozen references
- Public file manifest
- Byte-identical tested workflow ZIP