01
Volume
$260 per exception, 23% of loads, thousands of events a month. Most are the same five calls, made again.
TOD · Decision models for logistics and manufacturing
Every Exception.
One Decision.
TOD reads the ticket, the photo, and the system state, then picks the next action. It takes the decisions it is confident about and hands the rest to your LLM or your team.
Input
Carrier update, dock photo, or line alert
Policy + order state
Available actions
Hold · Reroute · Reorder · Escalate · Release
Your allowed choices
TOD
~400M parameters
Extract the facts. Rank the next actions.
Confidence gate
Confident → Serve
Take the selected action.
Skip the parent call.
Uncertain → Defer
Your LLM or your operator takes over.
Pass the turn unchanged.
Why decisions break at scale
The work is not the plan. The work is the exceptions.
A late carrier, a damaged pallet, a short-shipped PO, a sensor out of tolerance. Each needs a call: hold, reroute, reorder, escalate, release. Most are routine, and they still land on a person, one at a time. Roughly one in four truckload shipments hits an in-transit exception, at about $260 each to coordinate by hand. On the plant floor, every hour a line or lot waits costs $10,000 to $50,000.
02
Control
Your WMS, MES, and TMS expose a fixed set of actions. A model that writes free text cannot be wired to them safely.
03
Trust
Operators need to know when the model is guessing. A chat completion does not tell them.
One model, any input
Sees. Extracts. Acts.
A compact decision model of roughly 400 million parameters. It does not write text. It reads the event, pulls out the facts, ranks the actions you allow, and reports a confidence you can act on.
01
Sees
Reads the carrier message, the dock or line photo, the scanned BOL, and the order or work-order state together, alongside your policy.
Text · Photos · Scans · Screenshots · System state
02
Extracts
Pulls out the PO, the SKU, the quantity, the defect, the ETA, the dollar impact. A separate step resolves action arguments from your systems.
Unstructured event → Structured facts
03
Acts
Ranks the actions available on this turn and chooses from that set only. No invented actions, no free text. Serves when confident, defers when not.
Hold · Reroute · Reorder · Escalate · Release
Where the decisions live
Five queues. One model in front of each.
| Use case | The event | Actions TOD chooses from | Who handles deferrals |
|---|---|---|---|
| Carrier exceptions | Late pickup, missed scan, damaged in transit, address issue | Hold · Reroute · Reship · Claim · Notify customer · Escalate | Dispatcher or TMS agent |
| Dock and receiving | Short ship, overage, wrong SKU, damaged pallet with photo | Accept · Accept with variance · Reject · Hold for inspection · Open supplier claim | Receiving lead |
| Inventory and replenishment | Stock-out risk, min/max breach, supplier ETA slip, demand spike | Expedite · Reorder · Substitute · Transfer from another site · Defer | Planner or MRP agent |
| Quality and line alerts | Sensor out of tolerance, visual defect on a captured image, SPC rule hit | Continue · Hold the lot · Rework · Scrap · Stop the line · Call engineering | Line supervisor |
| Maintenance triage | Fault code, operator ticket, vibration or temperature alert | Monitor · Schedule PM · Dispatch tech · Order part · Escalate to OEM | Maintenance coordinator |
Receiving input
“Pallet 3 of 7 on PO 48812 came in crushed. 40 cases of the 120 look unsellable.”
+ Attached photo of the pallet
Structured facts
- PO
- 48812
- Line
- Pallet 3 of 7
- Received
- 120 cases
- Damaged
- ~40
- Condition
- Crushed, on photo
- Supplier
- On file
Next actions
- 01Accept 80 cases with variance and hold 40 for inspection.
- 02Open a supplier claim with the photo attached.
Illustrates text-and-image capability. Published benchmarks are measured on retail support conversations and image question sets; logistics results will be added as they are measured.
A System One model for the floor
Fast thinking for fast decisions.
The dispatcher who looks at a damaged-pallet photo and knows it is a claim is System 1. Chat LLMs are System 2: token by token, seconds per answer, prose you still have to parse. TOD is the System 1 side.
| TOD, System 1 | Chat model, System 2 | |
|---|---|---|
| How it runs | Parallel, one pass | Sequential, one token at a time |
| What it returns | A value from a schema you defined ahead of time | A string you parse and validate |
| Latency | 69 ms median, 121 ms p95 | Seconds per decision |
| Confidence | Says 90%, right about 90% of the time | Says 90%, right 60% of the time |
| Pricing | Flat per decision | Per token, rises with context |
Integration
Three steps.
01
Point TOD at the event stream
A webhook from your TMS, WMS, MES, or ticketing tool, or a batch pull.
02
Hand it the action set
The actions for that event type and the policy text you already have. TOD chooses from that set only.
03
Set the operating point
Conservative serves fewer turns at higher precision. Aggressive serves more and defers less.
Specs
Under the hood.
A compact encoder scores the available actions in two steps: which kind fits, then which candidate is best. Confident turns go straight to your system of record. Uncertain turns pass untouched to the model or person you already use.
| Spec | Value |
|---|---|
| Model size | ~400M parameters |
| Context | Up to 48,000 tokens per decision |
| Inputs | Text, photos, screenshots, scanned documents, system state |
| Options per decision | Hundreds; tested at 300 |
| Output | Typed action from your schema plus a calibrated probability |
| Decision latency | 69 ms median, 121 ms p95, one H100 80GB in bf16, before provider look-ups |
| Deployment | Hosted at tod.parseclab.ai, or open weights on Hugging Face for on-prem and edge |
Against the alternatives
Jev proved the category. TOD built it for the floor.
TypeSafe’s Jev showed a decision model beats a chat model when the answers are known up front. TOD adds what logistics and manufacturing run on: photos and scans, hundreds of actions, long context, a built-in handoff, and weights you can run inside the fence.
| TOD | Jev (TypeSafe) | Frontier LLM per event | Rules engine | |
|---|---|---|---|---|
| Reads photos, scans, screenshots | Yes | No, text only | Yes | No |
| Typed decision from your action set | Yes | Yes | No, writes free text | Yes |
| Calibrated confidence | Yes, 0.8% calibration error | Yes | No | No |
| Options per decision | Tested at 300 | Letter-indexed choices | Caps near 26 when read as letters | Whatever someone writes |
| Context per decision | 48K tokens | 32K tokens | Model-dependent, priced per token | n/a |
| Latency | 69 ms median | 70 to 500 ms | Seconds | ms |
| Pricing | Flat, $0.30 per 1,000 decisions | Per token, $0.042 per million input | Per token, rises with context | Near zero |
| Hands hard turns to your LLM | Built-in confidence gate | You build the fallback | It is the LLM | No |
| Runs on-prem or at the edge | Yes, open weights on Hugging Face | API only, early access | Rarely | Yes |
| Base | Open Gemma 4 12B, distilled to ~400M | Proprietary architecture | Proprietary | n/a |
Jev figures are TypeSafe’s published numbers.
Why TOD
Three reasons.
01
It sees the pallet
On held-out image sets, removing the image costs TOD 17 to 25 points. The pixels are doing real work.
02
It handles your real action set
A dock has dozens of dispositions. A reroute has every lane. TOD was tested at 300 options with 48K tokens of context.
03
It runs where your data lives
Open weights, flat per-decision pricing, and a gate that hands hard calls to the model you already pay for. 25% lower decision price than Jev on the evaluated workload.
The benchmark · Held-out data TOD never trained on
- Public JevBench
- 84.4%
- of 231 decisions correct
- Chart questions
- 97%
- correct with the image
- Up from 70% for the untrained base.
- Options per decision
- 300
- 55% correct where chance is 0.3%
- Calibration error
- 0.8%
- says 90%, right about 90% of the time
- Parent-model cost
- −23–26%
- on tau2-bench retail
- 92 to 101 of 114 tasks still solved.
- Argument resolution
- 94.7%
- exact match, 657 of 694
Measured on retail support conversations and image question sets. Logistics benchmarks are coming.
See the full TOD resultsTrust
Measured, not promised.
01
Open weights
TOD ships as open weights on Hugging Face. Inspect it, run it on your hardware, keep plant data inside the fence.
02
Auditable by design
Every decision is a ranked action list with a confidence score. Log it, replay it, and show an auditor why the lot was held.
03
Your actions, your policy
TOD never invents an action and never writes to your systems on its own. The gate and the action set are yours to tune.
FAQ
Questions.
Does TOD replace my WMS, TMS, or MES?
No. It sits in front of them and picks one of the actions they already expose. The write still goes through your system of record.
What happens when TOD is not sure?
It defers. The event passes unchanged to your existing LLM or to the queue a person already works.
Can it read photos from the dock or the line?
Yes. On held-out image sets, removing the image costs it 17 to 25 points, so it reads the pixels rather than guessing from text.
How many actions can one event have?
Hundreds. TOD was tested at 300 options per decision.
Can I run it inside my own network?
Yes. Run the open weights on-prem or at the edge, or use the hosted endpoint at tod.parseclab.ai.
What does it cost?
$0.30 per 1,000 decisions, regardless of context length. TOD fees are separate from your parent-model costs.
Do I need to retrain it for my plant?
Start with your action set and policy text. Parsec can fine-tune on your decision logs when you want higher serve rates.
Put a Decision in Front of Every Exception.
Run TOD on a week of your own exception events. Find the operating point that fits your systems, your volume, and your quality bar.