Dated findings.

• Weekly payroll prepared and sent through the platform every period since June 10, 2026, each with a period receipt and duplicate-send protection.
• Nightly POS ticket export at 10 PM with automatic gap healing; the daily shift COGS estimate runs after it.
• Weekly review collection every Sunday night.
• Payroll processor reports reconciled to the sent file at gross-total and employee-detail level each week.

Audit dated October 7, 2026, covering September 1, 2025 to August 31, 2026.

• 36 of 36 monthly and category sales checks agreed with the annual live POS report to the cent.
• A $578.81 discrepancy in the older order-detail archive was found and kept visible.
• Internal order-sum reconciliation passes on 1,332 days.

Graph snapshot September 9, verified October 16, 2026.

• 127 sellable items, 206 recorded recipe components, 74 pack relationships examined.
• 37 partially costed dishes carried $458,323.46 of $646,754.04 historical item subtotal, 70.9 percent, so the first job was to complete those recipe costs before touching prices.
• 7 dishes fully costed; 83 with no priced component; $75,402.16 (11.7 percent) of POS revenue with no sellable mapping.
• The audit produced a ranked six-step action list: confirm recipes and packs for the 37, resolve unmapped POS items, refresh and reconcile, establish a count window, complete bank tie-outs, then run a measured reversible experiment.

• Training record dated October 20, 2026: 14 configurations compared on the same eight seven-day origins, three seeds per learned neural family; 56 test days, July 7 to August 31, 2026.
• Every Chronos-2 variant beat the four-week baseline on point error with tighter bands; 80 percent coverage between 80.4 and 83.9 percent. The full table is on the platform page.
• Later rounds, October 2026: the knowledge-graph model 31.51 percent WAPE over 130 observed days in 19 weekly blocks; 728-day-context Chronos-2 30.92 percent; TimesFM 30.95 percent; retrained 728-day sales LoRA 30.94 percent, a 2.29-point improvement over its earlier run with a paired interval excluding zero; daily-issued one-day-ahead TimesFM 30.18 percent; complete weekly totals 13.73 percent WAPE over 17 weeks.
• Gradient boosting, October 16, 2026: 30.89 percent WAPE against 39.30 percent seasonal naive on 56 days; the nominal 80 percent band covered 85.71 percent.
• Live assistant run, July 2026: 25.56 percent against 36.95 percent on seven days; the 80 percent band covered six of seven, the 90 percent band seven of seven; two fresh-process refits produced byte-identical artifacts.

Report dated September 11, 2026.

• Two stable states, 205 and 448 guests a week, separated by a break-even line at 242 guests a week ($10,912 a week).
• The 2024 to 2025 fall from the high state hindcast within 12 percent, including the absence of recovery after the shock.
• A $1,037 weekly fixed-cost cut, or $192 a week of acquisition spend at $50 per household, moves the line below the current base.
• Six hours a week where congestion costs referrals; the knee at 8.08 guests per server-hour.
• Capital and timing to cross: $20,242 holding full staffing over 40 weeks; commit before the season.

October 21 to 22, 2026.

• 1,000 simulated restaurant and operator pairs: 200 restaurant families with five operator styles each. 700 pairs trained the policy and the world model; 300 were held out by whole family, 150 for calibration and selection, 150 for the final test. Source, protocol and checkpoints hash-pinned before the test was opened.
• 26-week episodes over 48 controls from 233 public state and history features; three collection rounds, five policy initializations, 47,215 weekly decisions in training; a second corpus of 15,921 decisions with 175 episodes each from frozen RL, a one-week planner, a four-week planner and uniform exploration.
• Red team: all 1,200 sealed episodes and 36,220 weekly transitions replayed exactly; every action, weekly profit, utility, cash value and terminal flag matched; cumulative profit reconciled to the change in equity. 1,522 artifact hashes, 300 checkpoint roundtrips and 2,002 tests passed.
• World model on held-out families: served-covers error 14.51 to 12.70 percent against the inherited mechanism, interval excluding zero; weekly profit prediction error $1,705 to $1,463; local adaptation 16.64 to 12.88 percent; 80 percent four-week band coverage 65 to 99 percent.
• Policy: the four-week planner is the executable reference and the learned policy earns each departure on held-out families. The learned hybrid finished within $3 per restaurant of the planner over 26 weeks on the 150 held-out pairs: 1,050 episodes, 26,275 operating weeks, a 450-checkpoint audit. Six policy versions, V7 through V12, each sealed before its test.
• Assistant evaluation: zero critical failures across 39 recorded episodes on independently authored owner tasks.

• Ticket-level history across two POS systems from December 2021.
• 538 observed sales dates in the current graph; 14,176 source objects and 5,517,377 observations preserved; 14,527 nodes and 103,783 edges; 88.3 percent of sales mapped to sellables.

October 20, 2026.

• 2,120 passing tests in the assembled repository.
• Import contracts: 4 kept across 289 files and 1,865 dependencies.
• Deterministic result hashes in fresh processes; a synthetic owner workflow produces an identical manifest on replay.