LogisticsFuture of AI & Visionary Thinking
Supply fusion in logistics: merging demand, capacity, transit, inventory, finance and risk into one estimate
Supply fusion is the practice of combining supply-chain signals that differ in latency, reliability and coverage — demand, supply and capacity, in-transit visibility, inventory position, finance, and risk — into a single estimate carrying a stated confidence, so one commitment can be made against all six rather than six commitments made against one each.
Key takeaways
- Supply fusion is combining signals of different latency, reliability and coverage into one estimate with a stated confidence. Putting six dashboards on one screen is not fusion — it renders the disagreement at higher resolution and leaves the resolution to whoever is looking.
- A fused estimate only beats the best single source when the sources' errors are not copies of each other. Two feeds derived from the same upstream reduce nothing and inflate confidence, which is why a correlation register belongs in the design before any weighting does.
- Most apparent disagreement between supply-chain sources is frame error — different event definitions, identifiers, time zones or units — not signal error. Resolve the frame first; weighting incompatible claims is arithmetic on nonsense.
- The moment a fused number drives a commitment it needs provenance: the contribution vector — which source, what value, what weight, how stale — stored alongside the result and retained for as long as the commitment can be disputed.
- The characteristic failure of fusion is one confident wrong number replacing three honest uncertain ones. The cheapest defence is to make confidence a first-class field the system of record refuses to write without.
Abbreviations used on this page
- TMS
- Transport management system
- WMS
- Warehouse management system
- YMS
- Yard management system
- ERP
- Enterprise resource planning — the finance and order ledger
- S&OP
- Sales and operations planning
- EDI
- Electronic data interchange — e.g. the 214 shipment status and 315 ocean status messages
- ETA
- Estimated time of arrival
- AIS
- Automatic identification system — broadcast vessel position and heading
- EPCIS
- The GS1 event-sharing standard: the what, when, where, why and how of a supply-chain event
- SSCC
- Serial shipping container code — the GS1 key that identifies a logistics unit
- POS
- Point of sale — retail sell-through, the fastest demand signal
- OTIF
- On-time in-full delivery rate
Free · 8 questions · ~3 minutes
Score your operation on the fusion ladder
Eight questions, one at a time, about three minutes. Answer them and we build your personalised fusion report — your rung on the ladder, your score on each of the four dimensions, and the specific thing standing between your signals and a single defensible estimate — and send it to your inbox.
0 of 8 answered
Pick an option to continue
Report ready
Your personalised fusion report is ready
Tell us where to send it. Your rung appears on screen straight away, and the full report — dimension scores, the signals you are most likely missing, and a first-pass weighting plan for your highest-value commitment — arrives in your inbox.
Your result
Your full report is on its way to your inbox.
Stage 1 · Siloed
Siloed is when each decision domain keeps its own number and nobody is required to reconcile them, so disagreement is invisible rather than absent.
Your next movePick one recurring commitment and write the event dictionary behind it: what each source claims, in what units, keyed to what identifier, refreshed how often.
Stage 2 · Aligned
Aligned is when the domain numbers are put side by side on a calendar and one of them is chosen, so reconciliation happens — but as an event, by seniority, and without a record of the reasoning.
Your next moveMove the tie-breaks out of the meeting and into written rules: for each pair of sources, which one wins under which condition, and what evidence would change that.
Stage 3 · Integrated
Integrated is when all six signals land in one place against shared identifiers and a deterministic rule picks the winner — but the output is still a chosen source, not a combined estimate.
Your next moveScore every source against realised outcomes by horizon, lane and counterparty. That error table is the raw material of every weight you will ever set.
Stage 4 · Fused
Fused is when the estimate is a weighted combination of sources with a stated confidence, and the commitment it drives carries that confidence and its provenance with it.
Your next moveClose the loop: score every source against realised outcomes continuously, re-fit the weights on a cadence, and review the change like code.
Stage 5 · Self-reconciling
Self-reconciling is when the fusion measures its own sources: weights are re-estimated from realised outcomes, weak feeds are demoted automatically, and cross-party feeds carry quality terms in the contract.
Your next movePut the score into the contract — agreed definitions, a published methodology, a revision policy and a review cadence both parties can see.
0 / 24
Domain coverage
— / 6
Fusion quality and confidence
— / 6
Cross-party fusion
— / 6
Fused-decision governance
— / 6
Your score maps to a rung on the fusion ladder. The dimension breakdown matters more than the total: the lowest dimension is what actually caps the trustworthiness of every fused number you produce, and it is where the next investment belongs. Your lowest-scoring dimension is —, and that is where the next investment belongs.
Your score maps to a rung on the fusion ladder. The dimension breakdown matters more than the total: the lowest dimension is what actually caps the trustworthiness of every fused number you produce, and it is where the next investment belongs.Your four dimensions score evenly, so there is no single weak link to attack — follow the stage’s next move above rather than picking a dimension.
Want your sources scored against what actually happened?
We take twelve weeks of your realised outcomes — gate-ins, receipts, sell-through, invoices — and score every source that claimed to predict them, by horizon and by lane. You get the error table, the correlation register and a first weighting proposal, whether or not we build anything.
How the score maps to a stage
- 0–4 — Stage 1, Siloed. Siloed is when each decision domain keeps its own number and nobody is required to reconcile them, so disagreement is invisible rather than absent.
- 5–9 — Stage 2, Aligned. Aligned is when the domain numbers are put side by side on a calendar and one of them is chosen, so reconciliation happens — but as an event, by seniority, and without a record of the reasoning.
- 10–15 — Stage 3, Integrated. Integrated is when all six signals land in one place against shared identifiers and a deterministic rule picks the winner — but the output is still a chosen source, not a combined estimate.
- 16–21 — Stage 4, Fused. Fused is when the estimate is a weighted combination of sources with a stated confidence, and the commitment it drives carries that confidence and its provenance with it.
- 22–24 — Stage 5, Self-reconciling. Self-reconciling is when the fusion measures its own sources: weights are re-estimated from realised outcomes, weak feeds are demoted automatically, and cross-party feeds carry quality terms in the contract.
What supply fusion is — and what it borrows from sensor fusion
A definition, the five disciplines that make fusion rigorous rather than decorative, and the path a signal takes from a feed to a commitment.
Supply fusion is the practice of combining supply-chain signals that differ in latency, reliability and coverage into a single estimate with a stated confidence. The six domains being merged are demand signal, supply and capacity, in-transit visibility, inventory position, finance and working capital, and risk. The product of fusion is not a screen; it is one number, an interval around it, and a record of how both were produced.
The discipline worth borrowing is sensor fusion — the problem of turning radar, lidar and camera readings into one estimate of where an object is. It is worth borrowing because it is rigorous about exactly the thing supply chains are vague about: what to do when two instruments disagree. A sensor-fusion engineer never averages raw readings. They put every reading into a common frame of space and time, attach a measured uncertainty to each instrument, combine so that the result is sharper than any input, gate readings that fall too far outside what the model expected, and re-estimate the weights from the residuals — the gap between what each instrument claimed and what turned out to be true.
Each of those five transfers directly. The common frame is identifiers and event definitions — the work that standards bodies have already done in GS1's identification keys (opens in a new tab) and the EPCIS event standard (opens in a new tab). The uncertainty model is a measured error distribution per source, per horizon, not a reputation. The combination is weighted, so the fused interval is narrower than any single source's. The gate is what stops a large disagreement being quietly averaged into invisibility. And the residual loop is what keeps the weights honest as the network changes. Everything on this page is one of those five, applied to freight.
How six supply signals become one estimate with a confidence
The path from feed to commitment, in three regimes. The top lane is where most operators sit: six views, one human picking, and a commitment that carries no record of the uncertainty it was made under. The middle lane is fusion proper. The bottom lane closes the loop by scoring the sources against what actually happened.
- Data & feeds
- Where value leaks
- System-of-record action
- AI / model
- Human in the loop
The process, in words
- In the top lane the six domain views arrive independently, a planner picks whichever they trust today, and the commitment is made against a single value with no record of the uncertainty or the discarded claims. When the commitment is questioned months later, the evidence that would explain it no longer exists.
- In the middle lane every source first enters a common frame — resolved identifiers, defined events, normalised time and units — then meets a store of its own measured error by horizon, lane and counterparty. The estimator produces a value and an interval, with feeds sharing an upstream weighted as one cluster rather than as independent witnesses.
- The gate is the part that distinguishes fusion from averaging. When two sources disagree by more than their own uncertainty allows, the estimate is not quietly split down the middle: the disagreement escalates to a named owner with both numbers and their history attached, because a large residual usually means the world moved, not that a sensor is noisy.
- In the bottom lane the realised outcome — gate-in, receipt, sell-through, invoice — is scored against what each source claimed, and the residuals re-fit the weights. This is what makes the arrangement self-reconciling, and it is only safe once you can prove the outcome was not influenced by the estimate itself.
Step-by-step insights
- The common frame — where most 'model' problems actually live
- Before any weighting is legitimate, two sources must be proved to describe the same event. In freight that is rarely trivial: a carrier's 'arrival' may be berthing, the terminal's may be first-lift, and the receiving team's may be gate-in at the DC — three different instants, hours or days apart, all called ETA. Add time zones stored inconsistently, units that differ between pallets and cases, and identifiers that only resolve through a reference field, and a large share of what teams experience as disagreement between sources turns out to be frame error. This is why the standards work matters operationally rather than theoretically: shared identification keys and a shared event grammar remove a class of disagreement instead of arbitrating it.
- The source error store — a reputation replaced by a distribution
- Every source needs a measured error distribution, and it must be conditional. The same carrier's ETA is a different instrument at fourteen days than at thirty-six hours; a terminal event is superb inside two days and silent beyond it; a lane's historical dwell is weak in normal weeks and the only thing left standing during a disruption. Storing error as a single number per source throws away the structure that makes fusion work. The practical minimum is error by source × horizon bucket × lane or trade, refreshed as outcomes land, with the sample size stored alongside so a thin cell can be treated as thin rather than as confident.
- Why the fused interval is narrower — and when that is a lie
- Combining independent estimates reduces variance: that is the whole mathematical case for fusion, and it is why a weighted blend beats even the best single source. The condition is independence. If two of your feeds are both derived from the same upstream — a visibility provider reselling the same carrier messages you already receive, or two portals rendering one terminal system — then they are one witness speaking twice. Treating them as two shrinks the computed interval without shrinking the real error, which is the precise recipe for a confident wrong number. A correlation register that caps the combined weight of each cluster is not an optimisation; it is the thing that keeps the confidence honest.
- The gate — a large disagreement is information, not noise
- When a source's claim falls outside what the fused estimate expected, given both uncertainties, the disagreement is itself the most valuable signal in the pipeline. Averaging it away destroys that. Gating means the estimate refuses the outlier, records it, and — if gate firings cluster — escalates, because a rising gate rate almost always means the world has moved outside the model's validity: a new lane, a diverted vessel, a strike, a carrier system migration. Operators who monitor gate rate get a leading indicator of network change for free; operators who average get a lagging one, expressed as complaints.
- The contribution vector — what makes the commitment defensible
- A fused value destroys its inputs unless you deliberately keep them. The contribution vector is the small record stored with every fused number: which sources contributed, what each claimed, what weight each carried, how stale each was at fusion time, and whether the gate fired. It costs a few hundred bytes per estimate and it is the difference between explaining a commitment months later and asserting it. Retention should be set by the commercial dispute window — demurrage, detention, customs and customer service-level claims all run far longer than a typical data-retention default — not by storage convenience.
- Closing the loop safely — the contamination trap
- Re-fitting weights from realised outcomes is only valid when the outcome is independent of the estimate. Supply chains break that assumption constantly: publish your fused ETA to a partner portal and it may come back tomorrow as that partner's own estimate, which your scorer then reads as independent corroboration and rewards with weight. The loop then converges on your own opinion with rising confidence and falling accuracy. Tag the lineage of every inbound value, exclude any source whose value could have been influenced by yours, and audit for round-tripping before automatic re-weighting is switched on.
Fused estimates beat the best single source because errors partially cancel
Combining sources whose errors are not perfectly correlated produces an estimate with lower variance than any input. This is the entire mathematical case, and it holds only under that condition — which is why identifying correlated feeds matters more than choosing an estimator.
…because each source is blind somewhere different
Carrier messaging stops at the gate, terminal events start at the quay, the warehouse system begins at the door, and finance only sees the movement weeks later on an invoice. No single feed covers the journey. The union does, and the seams are exactly where commitments fail.
…and because a fused estimate degrades gracefully when one input dies
A feed that stops updating takes a single-source estimate down silently. In a fused estimate with staleness weighting, the same failure widens the interval and shifts weight to the surviving sources — a visible, proportionate degradation rather than an invisible one.
But only if the confidence travels with the number
A fused point estimate with the interval stripped off is worse than the three honest sources it replaced, because it has destroyed the information that they disagreed. Confidence is not a nice-to-have on a fused number; it is the part that makes fusion defensible.
The five stages in detail: Siloed to Self-reconciling
For each rung: what it looks like on the ground, the diagnostic signals a reviewer can check in an afternoon, the anti-pattern that traps operators there, and what leaving costs.
The ladder runs Siloed → Aligned → Integrated → Fused → Self-reconciling, and each rung is defined by how disagreement between sources is resolved. At Siloed it is not resolved because it is not visible; at Aligned it is resolved in a meeting by authority; at Integrated by a fixed rule; at Fused by measured weights with a stated confidence; and at Self-reconciling by weights the system re-estimates from what actually happened.
The rungs are cumulative in a specific way: each one is mostly the previous one plus a new artefact. Aligned adds a cadence, Integrated adds a frame and a rule, Fused adds an error store and an interval, Self-reconciling adds a residual loop and a contract. Skipping an artefact does not accelerate the climb — it produces a system that computes a confidence from beliefs nobody measured, which is the most expensive failure on this page.
Decision quality released as fusion matures
The value is close to flat across the first two rungs, because a reconciled number produced weekly by a meeting is not usable by any downstream system. It inflects when the frame exists and the estimate starts carrying an interval — the point at which other systems can begin to reason about the number rather than merely display it.
Decision quality released by stage
- Stage 1 · Siloed — 22% of operators. Siloed is when each decision domain keeps its own number and nobody is required to reconcile them, so disagreement is invisible rather than absent.
- Stage 2 · Aligned — 34% of operators. Aligned is when the domain numbers are put side by side on a calendar and one of them is chosen, so reconciliation happens — but as an event, by seniority, and without a record of the reasoning.
- Stage 3 · Integrated — 27% of operators. Integrated is when all six signals land in one place against shared identifiers and a deterministic rule picks the winner — but the output is still a chosen source, not a combined estimate.
- Stage 4 · Fused — 14% of operators. Fused is when the estimate is a weighted combination of sources with a stated confidence, and the commitment it drives carries that confidence and its provenance with it.
- Stage 5 · Self-reconciling — 3% of operators. Self-reconciling is when the fusion measures its own sources: weights are re-estimated from realised outcomes, weak feeds are demoted automatically, and cross-party feeds carry quality terms in the contract.
Curve shape: logistic, plotted from the stage data above. Distribution: Consistent with GS1's published position on shared supply-chain data standards.
Select a rung
Every rung's full detail is in the page source — the selector only changes which panel is visible, so nothing here depends on JavaScript to exist.
Stage 1
Siloed
22% of operators sit here
Siloed is when each decision domain keeps its own number and nobody is required to reconcile them, so disagreement is invisible rather than absent.
Stage 1 is not an absence of data. Most operators at this stage have more supply-chain data than they have ever had: carrier messages, terminal events, telematics, warehouse scans, sell-through, invoices, weather. What is missing is any obligation for those sources to agree, and any mechanism that would notice if they did not. Each domain is internally consistent and externally unreconciled.
The tell is the arrival date. Ask three people in the same building when a specific container lands and you will get three answers — the carrier's published ETA, the terminal's berth plan, and the receiving team's working assumption based on how that lane behaved last month. Every one of them is defensible. None of them is the number, because there is no such thing as the number yet: there are only claims, and the claims have never been asked to stand next to each other.
This is a cheap stage to leave and an expensive stage to remain in, and the expense is invisible on any budget line. It shows up as buffer: extra days of cover, extra labour called in because nobody trusted the arrival window, extra premium freight because the working-capital view and the transport view were reconciled after the decision rather than before it. The buffer is the price of unreconciled signals, and it is paid every week.
In practice
The three arrival dates
A grocery importer's inbound container has an ETA of Tuesday in the carrier portal, a berth window starting Wednesday in the terminal's schedule, and a Thursday assumption in the DC's labour plan, because that terminal has been running two days late all quarter. Each system is correct about what it knows. The receiving manager books labour for Thursday, the planner promises the customer Tuesday, and the finance team accrues on the carrier date. When the box lands on Wednesday, all three are wrong in different directions and nobody can explain why the commitments disagreed.
What it looks like
- Demand, capacity, in-transit, inventory, finance and risk each have their own screen and their own owner
- No shared identifier links a purchase order, a container, a receipt and an invoice
- The same shipment carries three different arrival dates in three systems, and none of them is wrong
- An hour after a number is quoted, nobody can say which system it came from
Diagnostic signals you can check this week
- Ask three teams for the arrival date of one specific container and write down all three answers
- Ask which identifier links the carrier booking to the warehouse receipt — if the answer is 'the reference field, usually', you are here
- Count the screens a planner opens before making one commitment; above four, no fusion is happening
- Ask what happens when the carrier ETA and the terminal event disagree. If the answer is a person's name, the rule is not written down
Anti-pattern · Buying a control tower to fix a definition problem
The instinctive fix is a visibility platform that shows all six domains on one screen. It will, and the disagreement will still be there — now rendered in higher resolution and refreshed more often. A screen cannot resolve a conflict between two claims that were never defined in comparable terms, and every hour spent admiring the new view is an hour not spent writing down what 'arrival' means. Write the event dictionary for one decision first. The platform's real requirements are visible after that, not before it.
What holds you here
There is no common identifier and no shared event definition, so the domain numbers cannot be compared — let alone combined.
Highest-leverage next move
Pick one recurring commitment and write the event dictionary behind it: what each source claims, in what units, keyed to what identifier, refreshed how often.
Cost of leaving
- Effort
- 2–4 months
- Team
- One data engineer and one planner, part-time
- Risk
- Low — the work is documentation and instrumentation; nothing in production depends on it yet
- To next stage
- 2–4 months
If this is you, the next step is
A two-week exercise: list the sources, define the events, find the identifier gaps.
Stage 2
Aligned
34% of operators sit here
Aligned is when the domain numbers are put side by side on a calendar and one of them is chosen, so reconciliation happens — but as an event, by seniority, and without a record of the reasoning.
Stage 2 is where most logistics operators are, and it is a real achievement over stage 1: the claims are now made to stand next to each other on a regular cadence, and someone is accountable for producing a single answer at the end of it. The weekly S&OP call, the daily exception review and the pre-peak alignment session are all this stage in different clothing.
The structural weakness is that reconciliation is an event rather than a function. Between meetings the domains diverge again, quietly and at their own rates: the demand view moves with every order, the transit view moves with every milestone message, the finance view moves once a month. By Thursday the Tuesday consensus is describing a network that no longer exists, and the next meeting rebuilds it from scratch rather than updating it.
The second weakness is that the meeting resolves disagreement by authority. That is not a criticism of the people — in the absence of measured source quality, seniority is a reasonable proxy for judgement. But it means the resolution is unreviewable, unversioned and untransferable. When the person who always called it correctly changes role, the operation loses a capability that was never written down, and nobody can say by how much the estimates got worse, because nothing was ever scored.
In practice
The Tuesday call
A regional 3PL runs a Tuesday alignment call: planning brings the forecast, transport brings the carrier ETAs, warehousing brings the receiving capacity and finance brings the cost view. It takes ninety minutes and it works — a single plan leaves the room. It is also the only ninety minutes in the week when the six views agree. By Thursday, two carriers have revised, an inbound has been rolled, and the plan in the system reflects a Tuesday world. The following week the same ninety minutes is spent rebuilding rather than refining.
What it looks like
- A recurring call compares the demand, supply, inventory and transport views
- Identifiers mostly resolve, via a mapping table one person maintains
- The chosen number is recorded; the numbers not chosen are discarded
- Confidence is expressed verbally — 'we think Thursday' — and never stored anywhere
Diagnostic signals you can check this week
- Ask how long the reconciled plan stays valid; if the honest answer is under 48 hours, the cadence is the constraint
- Ask whether the losing numbers are stored anywhere after the call. Usually they are not
- Ask who resolves a demand-versus-capacity conflict, and whether the rule they use is written down
- Check whether the mapping table between systems lives in one person's spreadsheet
Anti-pattern · Adding a seventh source to settle a six-source argument
When the meeting cannot resolve a disagreement, the reflex is to buy another feed — a second visibility provider, a market-rate index, a weather service — on the theory that more evidence produces more certainty. It usually produces more argument, and often less certainty, because the new source is frequently derived from an upstream you already consume. You have not added an independent opinion; you have added a louder echo. Before buying anything, score the sources you already have against what actually happened.
What holds you here
Reconciliation is an event rather than a function, so the domains diverge between meetings and the reasoning behind each chosen number is lost.
Highest-leverage next move
Move the tie-breaks out of the meeting and into written rules: for each pair of sources, which one wins under which condition, and what evidence would change that.
Cost of leaving
- Effort
- 3–6 months
- Team
- A data engineer, an analyst, and the person who currently runs the reconciliation call
- Risk
- Low to medium — the change is procedural, and the meeting stays in place while the rules are written
- To next stage
- 3–6 months
If this is you, the next step is
We sit in the call, capture the tie-breaks people already make, and write them down as testable rules.
Stage 3
Integrated
27% of operators sit here
Integrated is when all six signals land in one place against shared identifiers and a deterministic rule picks the winner — but the output is still a chosen source, not a combined estimate.
Stage 3 is the first stage where the machine, rather than a meeting, produces the answer. The identifiers resolve, the events are defined, and a precedence ladder decides what wins. Operators arriving here report an immediate and durable relief: the arguments about whose number is right simply stop, because the rule is visible and the same rule ran last night as runs tonight.
What precedence buys is consistency. What it costs is information. When the terminal event overrides the carrier ETA, the carrier's claim is discarded rather than used — and the carrier's claim carried signal, particularly about the future, which the terminal event does not. Precedence is a lossy summary of the evidence, and its loss is largest exactly when the sources disagree most, which is exactly when the decision matters most.
The second cost is that precedence never learns. It is a set of static beliefs about which source is better, usually written by whoever understood the systems best in the year the ladder was designed. Networks change; a carrier's data quality improves after an integration project, a terminal's feed degrades after a system migration, a lane changes mode. Nothing in a precedence ladder notices. The move to stage 4 begins when someone starts measuring how often each source was actually right.
In practice
The precedence ladder that never learned
A retailer's inbound estimate applies a fixed ladder: terminal event first, then carrier ETA, then the lane's historical average. It ran unchanged for three years. In year two one carrier rebuilt its messaging and its ETAs became the most accurate signal in the estate at horizons over five days — better than the terminal, which only exists inside 48 hours. The ladder had no way to express that, so the best available evidence at the horizon where decisions were actually made was systematically discarded in favour of a source that had nothing to say yet.
What it looks like
- One data layer holds all six domains against common keys and a shared event dictionary
- Precedence rules resolve conflicts deterministically — terminal event beats carrier ETA, carrier ETA beats plan
- Every displayed number is traceable to the source it came from
- There are no weights and no confidence: the answer is a pick, not a blend
Diagnostic signals you can check this week
- Ask to see the precedence ladder, then ask when it was last changed and on what evidence
- Ask whether the losing sources' values are retained alongside the winning one — if not, no fusion is possible later
- Ask for the measured error of each source, split by horizon. If it does not exist, the ladder is a belief
- Check whether the estimate is a single value with no interval anywhere in the pipeline
Anti-pattern · Treating precedence as fusion
Because the pipeline is now automated, deterministic and traceable, it is easy to describe it internally as fused. It is not: it is a well-governed choice between claims. The distinction matters commercially, because a picked value cannot carry a defensible confidence, and a commitment made against it inherits an uncertainty nobody has quantified. The honest position at stage 3 is that you have a single number with unknown error. Say that out loud before someone downstream builds an automated decision on it.
What holds you here
Precedence discards information: the losing sources still carried signal, and nobody measures how often the winner was actually right.
Highest-leverage next move
Score every source against realised outcomes by horizon, lane and counterparty. That error table is the raw material of every weight you will ever set.
Cost of leaving
- Effort
- 6–12 months
- Team
- A data engineer, an ML or estimation engineer, and a named operations owner for the decision
- Risk
- Medium — the error store and the shadow estimate can be built without touching any live commitment
- To next stage
- 6–12 months
If this is you, the next step is
Twelve weeks of realised outcomes scored against every source, by horizon and lane.
Stage 4
Fused
14% of operators sit here
Fused is when the estimate is a weighted combination of sources with a stated confidence, and the commitment it drives carries that confidence and its provenance with it.
Stage 4 is where the arithmetic starts paying. A weighted estimate with a measured error distribution is not merely a nicer number: it is a number a downstream system can reason about. A labour planner can book to the 80th percentile of an arrival window; a customer-facing promise can be made at the confidence the contract requires; an automated replenishment can decline to act when the interval is too wide. None of that is possible with a point estimate, however accurate it happens to be on average.
The engineering is unglamorous and mostly bookkeeping. The estimator itself — an inverse-variance weighting, a small state-space model, a gradient-boosted residual correction — is rarely the hard part and is usually a week's work once the error store exists. What takes the quarter is the frame: proving that two sources are describing the same event, normalising time and units, resolving identifiers, and discovering which of your six feeds are secretly two.
The stage's characteristic risk is that the interval gets dropped at the last metre. The estimate is computed with a confidence, and then written into a system-of-record field that takes exactly one value, and the confidence dies at the boundary. Everything downstream then behaves as if the number were certain — which makes the fused estimate strictly worse than the three honest sources it replaced, because the disagreement information has been destroyed rather than summarised. Refusing to write a value without its interval is the single cheapest control on this page.
In practice
The 72-hour labour booking
A distribution centre books agency labour 72 hours ahead. The fused arrival estimate combines the carrier ETA, the terminal berth plan, an AIS-derived vessel position and the lane's realised dwell history, weighted by each source's measured error at the 72-hour horizon. The output is a window and a confidence. The booking rule is explicit: book to the 80th percentile of the window, and when the window is wider than eight hours, book the flexible shift instead. Reversals still happen — but they happen inside the stated interval, which is the definition of the estimate working.
What it looks like
- Weights derive from measured per-source error at the relevant horizon, not from contract tier or seniority
- Correlated sources are registered and weighted as a single cluster
- Every fused value carries an interval and a contribution vector
- Disagreements beyond tolerance fire a gate and escalate to a named owner instead of being averaged away
Diagnostic signals you can check this week
- Open the system of record and check whether the confidence is a field next to the value, or a slide in a deck
- Ask how the weights were set and when they were last re-fitted
- Ask for the correlation register — the list of which feeds share an upstream
- Ask what happens when two sources disagree beyond tolerance: silent average, or gate and escalation?
Anti-pattern · Shipping the point estimate and dropping the interval
The interval is the first casualty of integration, because the destination field takes one value and adding a second column is a change request. Teams promise to add it later and never do, and within a quarter the organisation has forgotten the number was ever uncertain. The failure surfaces as a category of incident nobody can explain: commitments that were reasonable given the evidence, judged afterwards as errors, because the evidence of uncertainty was not retained. Add the interval column in the same change as the value, or accept stage 3 honestly.
What holds you here
Weights are fitted once and reviewed rarely, so the fusion stays correct for the network as it was on the day it was built.
Highest-leverage next move
Close the loop: score every source against realised outcomes continuously, re-fit the weights on a cadence, and review the change like code.
Cost of leaving
- Effort
- 9–18 months
- Team
- An estimation engineer, an integration engineer, an operations owner, and commercial support for partner feeds
- Risk
- Medium to high — the first commitment made automatically on a fused number needs a drilled rollback to the previous source
- To next stage
- 12–24 months
If this is you, the next step is
We audit the weights against your realised outcomes and stress-test the gate on a real disagreement.
Stage 5
Self-reconciling
3% of operators sit here
Self-reconciling is when the fusion measures its own sources: weights are re-estimated from realised outcomes, weak feeds are demoted automatically, and cross-party feeds carry quality terms in the contract.
Stage 5 is narrower than the name suggests. It does not mean the supply chain reconciles itself; it means the estimation layer maintains its own beliefs about its sources without a human re-fitting them. A carrier whose messaging degrades after a system migration is down-weighted within days rather than at the next annual review. A new terminal feed earns weight by being right, not by being procured.
The engineering that makes this safe is the ablation harness: periodically recompute the estimate with each source removed and measure what the estimate loses. It answers two questions no dashboard can. First, which feeds are actually contributing — operators routinely discover they are paying for a source whose removal changes nothing, because it is a repackaging of something they already have. Second, which feed the whole estimate is quietly leaning on, which is the one to hold a contract conversation about before it fails.
The binding constraint at this stage stops being technical. Re-weighting a partner's feed downwards is a commercial act: it changes what you pay them for, what you promise your own customers, and what you will say in a dispute. Operators who reach here find that the work is contract design, published methodology and review cadence — and that the discipline most worth importing is the one used by institutions that publish uncertain estimates for a living: state the method, state the revision policy, and let the counterparty see their own score.
In practice
The demoted feed
An operator's fusion layer scores every in-transit source weekly against realised gate-in times. After a carrier's platform migration, that carrier's ETA error at the five-day horizon roughly doubles, and the weighting shifts automatically towards the terminal feed and lane history. Nothing breaks and no incident is raised — the estimate simply degrades gracefully. What the scorecard actually triggers is a commercial conversation, held with a measured series rather than an anecdote, and the carrier's own engineering team uses the same series to find the regression.
What it looks like
- Source scorecards update continuously from realised outcomes and drive the weights directly
- Ablation runs on a schedule, so each source's marginal contribution is a known number
- Calibration is monitored: of commitments made at a stated confidence, the hit rate is tracked
- Partner feeds carry contractual quality terms and a scorecard both parties can see
Diagnostic signals you can check this week
- Ask when the weights last changed and whether a human changed them
- Ask for the last ablation report and what it said about the least useful feed
- Ask for the calibration curve: of commitments made at 80% stated confidence, how many held?
- Ask whether any partner can see their own score, and whether the contract says what happens when it falls
Anti-pattern · Closing the loop on contaminated data
Automatic re-weighting is only safe if the realised outcomes are independent of the estimate. They frequently are not. If your fused ETA is published to a portal, a partner may ingest it and post it back as their own estimate — which your scorer then reads as an independent source agreeing with you, and rewards with more weight. The loop converges on your own opinion with rising confidence. Tag the lineage of every inbound value, refuse to score a source against an outcome it influenced, and audit for round-tripping before switching automatic re-weighting on.
What holds you here
The remaining constraint is commercial and legal rather than technical: down-weighting a partner's feed is a contract conversation before it is a code change.
Highest-leverage next move
Put the score into the contract — agreed definitions, a published methodology, a revision policy and a review cadence both parties can see.
Cost of leaving
- Effort
- Continuous
- Team
- A standing estimation team, an operations owner, and commercial and legal partners for the feed contracts
- Risk
- Concentrated — low frequency, high consequence, and increasingly contractual rather than technical
If this is you, the next step is
We stress-test the scorer, the ablation harness and the contamination controls against a real feed.
Where logistics operators actually sit on the fusion ladder
The distribution across the five rungs, and why the Integrated → Fused step loses the most operators.
Most logistics operators are at Aligned — reconciling their domain views on a cadence, in a meeting, with the outcome recorded and the reasoning discarded. A substantial minority have reached Integrated, where a common frame and a precedence rule produce the answer automatically. The number producing a genuinely weighted estimate with a stated confidence is small, and the number whose weights re-fit themselves from realised outcomes is very small indeed.
Distribution of logistics operators across the five rungs
Illustrative distribution. Aligned is the mode; the largest single drop is Integrated → Fused, because that step requires an artefact nobody has by accident — a measured error distribution per source, per horizon.
Share of operators
- 22% — 1 · Siloed
- 34% — 2 · Aligned (the mode)
- 27% — 3 · Integrated
- 14% — 4 · Fused
- 3% — 5 · Self-reconciling
The Integrated → Fused step is where the distribution thins, and the reason is that it is the first step requiring an artefact no organisation acquires by accident. A common frame can be built as a side effect of an integration programme; a precedence ladder can be written in an afternoon. A measured error distribution per source, per horizon, per lane can only come from deliberately recording what each source claimed and then scoring it against what happened — which nobody does unless someone decides to. Research centres working on supply-chain analytics, including MIT's Center for Transportation & Logistics (opens in a new tab) and the World Economic Forum's supply-chain centre (opens in a new tab), have made the same point from different directions: the constraint on multi-party supply-chain decision-making has moved from data availability to data comparability.
It is also worth being precise about what the top of the ladder is not. Self-reconciling does not mean a supply chain that runs itself; it means an estimation layer that maintains its own beliefs about its sources. That is a much narrower claim than most futures writing about supply chains makes, and it is the claim this page is prepared to defend. Gartner's supply-chain research (opens in a new tab) tracks the broader adoption picture; what it consistently shows is a gap between organisations that have connected their data and organisations that have made a decision differently as a result.
The signal ledger: six domains, six failure modes, six weight rules
What each fused domain actually claims, how fast it moves, how it fails quietly, where it is blind — and the rule that should set its weight.
Every source in a supply-chain estimate is a different instrument, and the ledger below is the specification sheet for all six. Read it as a sensor datasheet rather than as a data catalogue: the columns that matter are not what a feed contains but how fast it moves, how it degrades, what it cannot see, and therefore how much weight it has earned. A fusion design that skips this table will weight sources by contract value or by which vendor presented most recently, which is how confident wrong numbers get built.
| Domain | What it actually claims | Refresh and lag | How it fails quietly | Where it is blind | Weight rule |
|---|---|---|---|---|---|
| Demand signal | What customers will order, and when | POS daily; orders continuous; forecast weekly or per cycle | Keeps its shape while its level drifts — promotions, channel shifts and price moves all arrive looking like demand | Demand it never saw: substitutions, lost sales, unlisted channels, and the customer's own inventory position | Inside the replenishment lead time, firm orders outweigh any forecast; beyond it, the forecast is the only instrument in the room |
| Supply and capacity | What a supplier, carrier or terminal says it can do | Confirmations per booking cycle; spot capacity hourly; contract capacity per season | Confirmations are commitments, not observations — they stay green while the counterparty quietly falls behind, and correct all at once | The counterparty's own upstream: their raw-material position, their labour, their subcontractors | Weight by the counterparty's realised fulfilment rate, never by contract tier or relationship seniority |
| In-transit visibility | Where the goods are now and when they will arrive | EDI 214/315 at milestone events; AIS in minutes; terminal events per move | A stale ETA is re-served as fresh, and the same upstream is resold by two providers as two opinions | The gaps between milestones — the long silences where nothing is emitted and interpolation is doing the work | Weight by measured error at the specific horizon, and cluster every provider sharing an upstream into one witness |
| Inventory position | What is where, and what can be promised | WMS in near real time; ERP nightly; in-transit stock inferred rather than observed | Counts drift between cycle counts, and 'available' diverges from 'on hand' as reservations, holds and damage accumulate | Quality holds, damaged stock, and inventory sitting in a partner's building under someone else's system | Prefer the system that owns the physical movement over the system that owns the ledger; treat in-transit stock as an estimate, not a count |
| Finance and working capital | What the decision costs, and when cash actually moves | Invoices per settlement cycle; rate cards per contract; FX daily | Prices lag reality — the rate card is the last artefact updated after a market move, so cost views stay stale for weeks | Accessorials, demurrage and detention, which are invisible until they are invoiced long after the decision | For anything spot, weight recently realised invoices over rate cards; for contract lanes, the reverse |
| Risk | What could invalidate the plan, and roughly where | Weather hourly; port congestion daily; labour, geopolitical and carrier-health signals episodic | High recall and low precision — acted on literally it generates constant churn, so teams learn to ignore it entirely | The specific: it can tell you the region, the port or the corridor, but almost never your box | Never a value. Risk widens the interval and lowers the gate threshold; it does not move the point estimate |
The last row is the one most often got wrong, and it is worth stating as a rule: risk signals are interval-wideners, not estimate-movers. A storm forecast does not tell you that this container will be four days late; it tells you the distribution of possible arrivals has a longer tail this week. Feed it in as a shift in the point estimate and you generate churn — every plan moves, most of the moves are wrong, and within two cycles the operation stops believing risk signals at all. Feed it in as a widening of the interval and a lowering of the gate threshold, and the effect is exactly what you want: commitments get more conservative and more disagreements get escalated, precisely during the period when the network is least predictable. Public forecasting institutions have handled this correctly for decades — ECMWF's operational forecasts are ensemble-based (opens in a new tab), describing a range of scenarios and their likelihood rather than a single answer, and NOAA (opens in a new tab) publishes probabilistic products for the same reason.
The frame: identifiers, events, time and units
One key that resolves an order, a logistics unit, a location and an invoice to the same physical thing; one written definition per event; one time base; one unit per measure. GS1's identification keys (opens in a new tab), the EPCIS event standard (opens in a new tab) and master-data syndication through GDSN (opens in a new tab) exist precisely so this does not have to be invented per trading relationship. Build it before the estimator, not after.
The error store: what each source claimed, and what happened
A table of source × horizon × lane holding the measured error distribution and the sample size behind it. It is built by writing down every claim at the time it is made and joining it to the realised outcome later — cheap to start, impossible to backfill, which is why it should start the week you decide fusion is the direction.
The correlation register: which feeds are secretly one feed
A documented list of which sources derive from a shared upstream, with a cap on the combined weight of each cluster. Most estates contain at least one cluster nobody has noticed — typically a visibility provider reselling carrier messages the operator already receives directly, counted twice as agreement.
Those three artefacts are the whole prerequisite list. Notice what is not on it: a new platform, a data lake migration, or a model. The estimator is genuinely the easy part — an inverse-variance weighting over four sources is a short function, and the sophisticated alternatives buy less than practitioners expect once the frame and the error store exist. What buys the accuracy is knowing which instrument to believe, at which horizon, on which lane. That is bookkeeping, and it is the bookkeeping this page is about.
Resolving disagreement: the order that makes fusion defensible
Frame, then correlation, then weight, then gate. Getting the order wrong is what produces a confident wrong number.
Disagreement is resolved by rule, in a fixed order: reconcile the frame, cluster the correlated sources, apply measured weights, and gate what remains. The order is not stylistic. Weighting before the frame is reconciled produces confident nonsense, because the arithmetic is being applied to claims about different events. Weighting before correlation is registered inflates confidence, because echoes are counted as witnesses. Gating last is what catches the residual — the disagreement that survives all three steps and therefore means something.
The disagreement-resolution order
1 · Prove the sources are describing the same event
Before any comparison, confirm that both claims refer to the same instant, the same object and the same unit. Berthing is not first-lift; first-lift is not gate-out; gate-out is not receipt. A large share of what teams experience as source disagreement dissolves here, and the resolution is permanent — a definition fixed once stops generating disagreements forever, whereas a weight tuned around a definitional mismatch has to be re-tuned every time the mismatch moves.
2 · Cluster sources that share an upstream
Trace each feed to its origin and group the ones that are re-serving the same underlying data. Cap the weight of the cluster, not the members. Two providers rendering one terminal's events are one witness with two microphones, and the most common way an estate ends up with a fused interval far narrower than its actual error.
3 · Weight by measured error at the decision horizon
Use the error store, conditioned on horizon and lane. The horizon condition matters more than practitioners expect: sources rank differently at fourteen days than at thirty-six hours, and a fixed ranking is wrong at one end of that range by construction. Where a cell is thin, widen the interval rather than pretending to a precision the sample does not support.
4 · Gate what still disagrees, and never silently average it
If a source's claim lies outside what the fused estimate expected given both uncertainties, exclude it, record it, and route it. A gate firing is an event worth a person's attention; a cluster of gate firings on one lane or one counterparty is a network change worth acting on before it becomes an incident.
5 · Escalate with both numbers, not with an answer
When a disagreement reaches a human, it should arrive as two claims, their histories and their measured reliabilities — not as a pre-averaged value with an amber icon. The person is being asked to exercise judgement about which instrument to believe in an unusual situation, and stripping the evidence to make the alert tidy removes the only thing that makes their judgement better than a coin toss.
6 · Write down who owns the tie-break before you need one
Every fusion needs a named owner per decision class, with the authority to override and the obligation to record why. The record is the point: an override without a reason is indistinguishable from noise six months later, and the accumulated reasons are the highest-quality training data any fusion layer will ever get.
What to do when sources disagree
Plot how much the sources agree against how expensive it is to be wrong. Only one quadrant genuinely needs a human, and the top-right is the one that quietly catches people out — agreement is only reassuring when the agreeing sources are independent.
Commit and move on
- Sources agree; the commitment is cheap to reverse
- Automate; sample-audit rather than review
- Do not spend governance budget here
Commit, but check independence
- Agreement is only evidence if the sources are independent
- Check the correlation register before trusting the narrow interval
- Store the contribution vector — this is the quadrant that gets disputed
Pick a rule and log it
- Sources disagree, but being wrong is cheap
- Deterministic tie-break, no human in the path
- Log the disagreement — it is free error-store data
Escalate with both numbers
- Disagreement plus expensive commitment: the only quadrant needing a person
- Send both claims, their histories and their reliabilities
- Record the decision and the reason; it sets tomorrow's rule
Different data owners are involved per port call: it is always a combination of both the port and the terminal, yet also nautical service providers and services related to vessels and cargo.
That sentence is the whole cross-party problem in one line. No single participant in a port call holds the truth about it, and each holds a fragment that is authoritative for their own part and speculative about everyone else's. The same structure repeats at every handover in a supply chain — factory to forwarder, forwarder to carrier, carrier to terminal, terminal to haulier, haulier to warehouse. Fusion across companies is not a data-engineering exercise with a contract attached; it is a contract with a data-engineering exercise attached, and the sequencing follows from that.
Which frame you adopt depends on the mode, and in every mode somebody has already done the work. Container shipping has DCSA's data standards (opens in a new tab); air cargo has IATA's ONE Record (opens in a new tab), which defines a single record view of a shipment shared through a standardised API rather than a chain of per-party message copies; the message layer between trading partners has GS1's EDI standards (opens in a new tab); and the assurance wrapper around all of it — how a management system is documented, reviewed and audited — sits in the ISO catalogue (opens in a new tab). Adopting an existing frame is almost always cheaper than negotiating a bilateral one, and it carries a second-order benefit that compounds: a counterparty already emitting the standard can be added to your fusion without a project, which is what turns cross-party fusion from a series of integrations into a capability.
The trust machinery that has to sit on top of the frame is short and specific. A partner's number can carry weight only when four things are written down: what the field means, how often it updates, how far it may be revised after the fact, and what happens commercially when it is persistently wrong. Add a fifth for fusion specifically — a scorecard the counterparty can see. Making the score visible is what converts a source-quality measurement from surveillance into a contract term, and it is the difference between a supplier engineering team fixing a regression you detected and a supplier account manager disputing it.
