Skip to content

Feedforward

Prime #
862
Origin domain
Engineering & Design
Subdomain
control and information systems → Engineering & Design

Core Idea

Feedforward is the structural arrangement in which a system supplies predictive information about the consequences of an action before the action is committed, so that the actor can pre-correct rather than wait for a deviation to feed back. Where feedback closes a loop after output is realized — sense the error, then adjust — feedforward opens a window of cheap pre-commitment correction based on a model of what the action will produce. The correction is shifted from the reactive arc (deviation-and-recovery) to the anticipatory arc (model-consequence-and-adjust). The essential commitment is that a predictive model is interposed between intention and irrevocable commitment: the actor acts on the modeled consequence of the action rather than on the realized consequence.

Every feedforward arrangement specifies four elements. There is an actor with an action available but not yet committed; a measurable or modelable disturbance or consequence whose effect can be anticipated; a predictive model mapping intended action and disturbance to expected outcome; and a pre-action correction channel that modifies the action based on that prediction. A residual feedback loop typically remains, catching whatever the model missed. The pattern is licensed by an asymmetry of cost: a model of the consequence-of-action, computed before the action, is often vastly cheaper than the realized error it averts. The substrate of the prediction is incidental — it may live in a controller, an interface, a forecast document, or a brain — and what matters structurally is only that a model is placed upstream of commitment, where it can still change the action.

How would you explain it like I'm…

Catch It Early

When you reach to catch a ball, you don't wait until it hits your hand — you guess where it's going and move there ahead of time. Feedforward is using a good guess about what will happen next, so you can get ready before it actually happens instead of fixing it afterward.

Brace Before the Bump

Imagine you're carrying a tray and you see someone about to bump you. You can wait until they hit you and then stumble to recover (that's reacting after the fact), or you can brace yourself a second before the bump using your guess of what's coming. Feedforward is that second thing: you act on a prediction of what an action will cause, before you fully commit, so you can correct ahead of time. There's usually still a backup that catches whatever your guess got wrong. It works because a good prediction is often much cheaper than cleaning up a mistake.

Predict-Then-Act

Feedforward supplies predictive information about the consequences of an action before the action is committed, so the actor pre-corrects instead of waiting for an error. Contrast it with feedback, which closes a loop after the output is realized — sense the deviation, then adjust. Feedforward instead uses a model of what the action will produce, shifting correction from the reactive arc to the anticipatory arc. Every feedforward setup has four parts: an actor with an action ready but not yet committed, a disturbance or consequence that can be anticipated, a predictive model mapping intended action and disturbance to expected outcome, and a channel to adjust the action before it's committed. A residual feedback loop usually remains to catch what the model missed.

 

Feedforward is the structural arrangement in which a system supplies predictive information about the consequences of an action before the action is committed, so the actor can pre-correct rather than wait for a deviation to feed back. Where feedback closes a loop after output is realized — sense the error, then adjust — feedforward opens a window of cheap pre-commitment correction based on a model of what the action will produce, shifting correction from the reactive arc to the anticipatory arc. The essential commitment is that a predictive model is interposed between intention and irrevocable commitment: the actor acts on the modeled consequence rather than the realized one. Every feedforward arrangement specifies four elements: an actor with an action available but not yet committed, a measurable or modelable disturbance or consequence, a predictive model mapping intended action and disturbance to expected outcome, and a pre-action correction channel. A residual feedback loop typically remains, catching whatever the model missed. The pattern is licensed by a cost asymmetry: modeling the consequence before acting is often vastly cheaper than the realized error it averts, and the substrate of the model is incidental.

Structural Signature

the actor with an uncommitted action — the measurable or modelable disturbance — the predictive model mapping action to expected outcome — the pre-action correction channel upstream of commitment — the residual feedback loop catching the unmodeled — the cost-asymmetry that licenses the model, gated by model accuracy

The pattern is present when the following components are jointly in play:

  • The actor with an available-but-uncommitted action. A party holding an action it has not yet irrevocably taken. The correction must be possible while the action can still change.
  • The disturbance or consequence (the anticipable quantity). A measurable or modelable influence whose effect on the outcome can be predicted before it is realized.
  • The predictive model (the interposed element). A mapping from intended action and disturbance to expected outcome, placed upstream of commitment. Its substrate — controller, interface, forecast, brain — is incidental; its placement is the pattern.
  • The pre-action correction channel. A path that modifies the action on the basis of the prediction, shifting correction from the reactive arc (deviation-and-recovery) to the anticipatory arc (model-and-pre-commit).
  • The residual feedback loop. A retained after-the-fact loop catching whatever the model missed; feedforward and feedback compose as complements partitioned by which disturbances are measurable in advance.
  • The accuracy-gated cost-asymmetry invariant. Pre-correction pays only when a model of the consequence is cheaper than the realized error and accurate enough that pre-correcting beats not correcting; an inaccurate model injects bias worse than none.

Composed, these interpose a predictive signal between intention and irrevocable commitment, routing the measurable share of disturbance through a model placed upstream and leaving the residual to feedback.

What It Is Not

  • Not feedback. feedback corrects after a deviation is realized by routing measured output back to the input; feedforward corrects before commitment using a model of the action's consequence. The arrows differ — feedback closes a loop on realized error, feedforward opens a window of anticipatory pre-correction.
  • Not prediction itself. foreseeing_prediction is the model that forecasts a consequence; feedforward is the structural placement of that model upstream of irrevocable commitment, where it can still change the action. A forecast that arrives after the action commits is prediction without feedforward.
  • Not foresight. foresight is the broad capacity to anticipate; feedforward is a specific loop topology that interposes an anticipatory correction channel between intention and commitment. Foresight is a disposition; feedforward is a control structure.
  • Not predictive coding. predictive_coding is a particular neuroscientific theory in which the brain propagates prediction errors up a hierarchy; feedforward is the substrate-neutral pattern of placing a consequence-model upstream of action, of which predictive coding's forward model is one biological instance.
  • Not learning. learning updates an internal model from experience; feedforward uses a model to pre-correct an action. They interact — error data validates the forward model — but using a model upstream is not the same as revising it.
  • Not a learning curve. learning_curve_effects describe performance improving with cumulative practice; feedforward is a real-time control arrangement, not an effect of accumulated repetition.
  • Common misclassification. Treating a faster or tighter feedback loop as feedforward. If correction still waits for the output to deviate before acting, no model has been placed upstream and the system is doing feedback, however responsive. Catch it by asking whether the disturbance is knowable — and the action still changeable — before commitment.

Broad Use

  • Control engineering. A feedforward controller measures a disturbance (incoming load on a heater, raw-material temperature) and adjusts the actuator before the disturbance moves the controlled variable, avoiding the lag inherent in pure feedback.
  • Human-computer interaction. Interfaces surface the consequence of a hover, drag, or click before commitment: preview thumbnails, "this will delete 37 files" dialogs, dry-run modes. The user predicts the outcome from the interface rather than from experiencing the action.
  • Organizational planning and governance. Pre-mortems, environmental-impact statements, and budget projections give decision-makers a model of downstream consequence before resources are committed.
  • Training and instruction. A worked example, a model performance, or a simulator run shown before the learner attempts the task moves the cost of errors from recovery to pre-commitment.
  • Neuroscience and motor control. Efference-copy and forward-model circuits predict the sensory consequences of intended motor commands, letting the system distinguish self-caused from externally caused signals and pre-tune the actuator.
  • Software operations. A deployment system runs a dry-run plan (feedforward) before applying a change; a monitoring system (feedback) catches whatever the dry-run missed.

Clarity

Feedforward names a class of corrections that are otherwise mistaken for "just better feedback." It separates two question-types that look similar but call for opposite interventions: what happened? (feedback) and what will happen if I do this? (feedforward). Once a designer holds the distinction, redundant feedback-loop remediation — faster sensors, tighter gains, quicker reporting cadence — gets reframed as a possible feedforward gap, because the latency cost of learning by error can sometimes be eliminated outright by predicting before commitment rather than reacting faster after it.

The framing also disciplines a recurrent diagnosis. When users repeatedly trigger irreversible actions they regret, the clarifying move is to recognize that the remedy is not a faster undo (better feedback) but a preview (feedforward). When a planning process discovers risks only after commitment, the remedy is a pre-mortem or impact statement, not a faster reporting cycle. The viewpoint shift — from deviation-and-recovery to model-and-pre-commitment — is what makes these reframings available.

Manages Complexity

Pure feedback systems must absorb every disturbance through the controlled variable, which requires fast actuators, tight loop gains, or tolerance of overshoot. Feedforward decomposes the disturbance problem along a clean seam: measurable disturbances are pre-corrected with an open-loop model, while unmeasurable residuals are left to feedback. The whole problem becomes the sum of two simpler problems, each solved with the cheapest mechanism appropriate to it. The same decomposition recurs across substrates — pre-mortems (measurable risks pre-mitigated, residual risks watched live), worked-example pedagogy (predictable pitfalls pre-warned, residual misconceptions caught by formative feedback), and UI design (preview catches the predictable, undo catches the residual).

The compression also relocates where preparation lives. Instead of provisioning the reactive loop to handle the full disturbance budget, the system invests in a model that handles the predictable share up front. This is leverage precisely when prediction is cheaper than recovery and the model is accurate enough to help — and the frame makes that condition explicit rather than assumed.

Abstract Reasoning

Feedforward exposes an asymmetry of information cost: the price of a consequence-of-action model computed before the action is often far below the cost of the realized error, which is what makes interposing the model worthwhile. This unlocks reasoning about where the model lives — in the controller, the interface, the forecast, the cerebellum — while holding the structural role of the model constant across all of them. The substrate of the prediction is incidental; the pattern is the placement of a predictive signal upstream of irrevocable commitment, where it can still change the action.

The abstraction also exposes its own failure mode with equal clarity. An inaccurate feedforward model injects systematic bias that the residual feedback loop must then correct, which can be worse than no feedforward at all. So the porting question always includes a sharp test: is the predictive model accurate enough that pre-correction improves on no correction? Reasoning at this level distinguishes feedforward from prediction itself (prediction is the model; feedforward is the structural placement of that model upstream of action) and from feedback (which corrects after the fact), and treats the two loops as complements whose right balance depends on which disturbances are measurable in advance.

Knowledge Transfer

The structure transfers as a portable set of role mappings and interventions. The roles hold steady across domains: the predictive model maps to the feedforward controller, the preview interface, the forecast document, the brain's forward model; the disturbance maps to the upstream load, the user's intended action, the anticipated risk, the intended motor command; the pre-action correction channel maps to the actuator adjustment, the dialog that alters the click, the plan revision, the pre-tuned actuator; and the residual feedback loop maps to the monitor, the undo, the live risk watch, the sensory check that catches what the model missed.

The interventions travel as a recognizable playbook. If a feedback loop is too slow, ask what disturbances are measurable upstream that could be pre-corrected. If users repeatedly trigger irreversible actions they regret, add a preview rather than a faster undo. If a training program over-relies on learning-from-mistakes, insert worked examples or simulators to shift cost from error-recovery to pre-commitment correction. If a planning process discovers risks only after commitment, install a pre-mortem or impact statement rather than tightening the reporting cadence. Each transfer carries the same accuracy caveat — a bad feedforward model is worse than none — so the porting move always pairs the intervention with a check on model quality. The unifying residue, which survives the strip-the-jargon test, is a predictive signal interposed between intention and commitment: in HCI it surfaces as affordance preview, in governance as impact assessment or pre-mortem, in neuroscience as efference copy, but the loop topology and its asymmetric-cost rationale are identical. Because the pattern pairs naturally with feedback — most well-designed systems run both, feedforward handling the measurable disturbances and feedback the residual — the transferable diagnostic is always to ask which share of the disturbance is predictable in advance and to route that share through a model placed upstream of commitment.

Examples

Formal/abstract

A heat exchanger with a feedforward controller is the canonical control-engineering instance, and it exhibits every role end-to-end. The actor with an uncommitted action is the steam-valve actuator that has not yet set its position; the measurable disturbance is the inlet temperature of the cold process fluid, sensed upstream of the exchanger before it perturbs the controlled outlet temperature. The predictive model is an energy-balance equation: given the inlet temperature and flow rate, it computes the valve opening required to hit the outlet setpoint. The pre-action correction channel drives the valve to that opening before the cold slug reaches the heat-transfer zone — so the disturbance is cancelled in anticipation rather than after the outlet has already drifted. The residual feedback loop is a conventional outlet-temperature controller that trims whatever the energy-balance model got slightly wrong (fouling, unmodelled losses). The cost-asymmetry that licenses the model is the lag: pure feedback must wait for the outlet to deviate, and the exchanger's thermal mass makes that lag long and the recovery expensive, whereas the inlet-temperature measurement is available immediately and the model is cheap to evaluate. The accuracy gate is sharp and worth stating: if the energy-balance model is biased (wrong heat-transfer coefficient), the feedforward term injects a systematic error the feedback loop must then fight — worse than no feedforward at all. The diagnosis this enables: when an outlet temperature oscillates because feedback cannot keep up with inlet swings, the fix is not a faster sensor but a feedforward term on the measurable inlet disturbance.

Mapped back: The valve is the uncommitted action, inlet temperature is the measurable disturbance, the energy balance is the predictive model, the pre-set valve position is the upstream correction, and the outlet trim controller is the residual feedback — with model bias the accuracy-gate failure.

Applied/industry

A destructive-action confirmation dialog and an organizational pre-mortem instantiate the same upstream-model structure in software and governance. When a user clicks "delete," a well-designed interface interposes a predictive model of the consequence — "this will permanently delete 37 files, including 4 shared with others" — before the action commits. The actor with an uncommitted action is the user mid-click; the disturbance is the user's own intended action whose full consequence they have not computed; the pre-action correction channel is the dialog's Cancel button, which lets the action change while it still can; the residual feedback loop is undo, catching whatever the preview did not convey. The cost asymmetry is stark: rendering a preview is cheap, whereas recovering from an irreversible deletion is expensive or impossible — which is exactly why the prime's diagnostic says the remedy for regretted irreversible actions is a preview (feedforward), not a faster undo (better feedback). The governance parallel: a pre-mortem asks a team to imagine the project has already failed and to model the consequences before resources are committed, surfacing predictable risks (the measurable disturbances) for pre-correction, while live risk monitoring (the residual feedback) catches the rest. The accuracy gate appears here too: a pre-mortem built on a bad causal model of the project injects false confidence or false alarm, so the predictive model's quality gates the value of the exercise — the same caveat that governs the heat-exchanger feedforward term.

Mapped back: The delete-dialog preview and the pre-mortem are predictive models placed upstream of commitment; the click and the resource decision are the uncommitted actions; Cancel and the risk-mitigation step are the pre-action corrections; undo and live monitoring are the residual feedback loops.

Structural Tensions

T1 — Anticipatory versus Reactive Correction (temporal, boundary with feedback). Feedforward corrects before commitment; feedback corrects after deviation; the two are complements partitioned by which disturbances are measurable in advance, and confusing which arc a problem belongs on yields the wrong remedy. The failure mode is treating a feedforward gap as a feedback-tuning problem — chasing faster sensors and tighter gains when the latency could be eliminated outright by predicting before acting (or, inversely, modeling an unmeasurable disturbance that only feedback can catch). Diagnostic: ask whether the disturbance is knowable before the action commits; if yes, pre-correct; if it is only knowable after the fact, no model placement helps and feedback is the only arc available.

T2 — Model Accuracy versus No Model (sign-flip). Pre-correction helps only when the predictive model is accurate enough that acting on it beats not correcting; below that threshold the feedforward term injects systematic bias that feedback must then fight, making it worse than none. The value of the move flips sign with model quality. The failure mode is adding feedforward reflexively — a pre-mortem on a bad causal model, a controller with a wrong coefficient — and degrading a system that pure feedback handled adequately. Diagnostic: before installing feedforward, test whether the model's predictions beat the realized error it aims to avert; an unvalidated forward model is a bias generator, not a correction.

T3 — Cost of Modeling versus Cost of Recovery (measurement asymmetry). The pattern is licensed by an asymmetry — modeling the consequence is cheaper than recovering from the error — but that asymmetry is an empirical claim that does not always hold. When the model is expensive to build and maintain and the realized error is cheap to recover from, feedforward is net-negative effort. The failure is mounting elaborate preview or forecast machinery for actions whose mistakes are trivially reversible, paying more for prediction than the errors would ever cost. Diagnostic: compare the standing cost of the model against the expected cost of the deviations it averts; feedforward earns its place only when recovery is expensive or irreversible relative to prediction.

T4 — Predictable Share versus Residual Share (scopal partition). Feedforward handles the measurable disturbances; an irreducible residual — fouling, unmodeled losses, novel conditions — must be left to feedback, and the partition between the two is itself a design choice that can be drawn wrong. The failure mode is over-trusting the model to cover the whole disturbance budget and provisioning no residual loop, so whatever the model missed propagates uncaught. Diagnostic: ask what share of disturbance the model genuinely captures and confirm a feedback path exists for the rest; a feedforward system without a residual loop assumes its model is complete, which no model is.

T5 — Prediction versus Placement (kind-confusion). Feedforward is not the predictive model itself but the structural placement of that model upstream of irrevocable commitment, where it can still change the action; a forecast that arrives after the action is committed is prediction without feedforward. The two are routinely conflated. The failure is producing accurate consequence-models that land too late to alter the action — an impact assessment delivered after resources are spent, a preview shown after the click resolves. Diagnostic: ask not "is the prediction good?" but "can the action still change when the prediction arrives?"; if commitment has already occurred, there is no feedforward however accurate the model.

T6 — Pre-Commitment Confidence versus Earned Experience (coupling/side-effect). Shifting cost from error-recovery to pre-commitment averts mistakes, but mistakes are also how reactive systems and learners acquire the model in the first place; pre-correcting too thoroughly can starve the feedback loop of the error signal that would improve the model. The intervention that prevents errors can also prevent the learning that errors provide. The failure mode is a training program or automated controller so good at pre-empting deviation that its model never gets corrected against reality and silently drifts. Diagnostic: ask whether the residual feedback loop still receives enough error to keep the forward model honest; perfect pre-correction that eliminates all deviation also eliminates the data that validates the model.

Structural–Framed Character

Feedforward sits at the structural pole of the structural–framed spectrum, with a structural label and an aggregate of 0.0 — the same paradigm-structural reading as its complement, feedback. It is a pure loop topology: a predictive model of an action's consequence interposed upstream of irrevocable commitment, so the actor pre-corrects rather than waits for a deviation to feed back. Every diagnostic points one way.

The pattern carries no home vocabulary that must travel with it: the identical structure is told in each field's own words — a steam-valve actuator pre-set from an inlet-temperature energy balance, a "this will delete 37 files" dialog interposed before a click, a pre-mortem run before resources are committed, the cerebellum's efference-copy forward model pre-tuning a motor command — and none of these borrows a control-theory lexicon to be understood. It carries no evaluative weight: placing a model upstream of commitment is neither good nor bad until you specify what the action does. Its origin is formal — a control-and-information-systems loop describable purely as the placement of a predictive signal between intention and commitment, with a residual feedback path catching the unmodeled remainder — and it appeals to no human institution. It is not human-practice bound: the substrate of the prediction is explicitly incidental, running indifferently in a controller, an interface, a forecast, or a brain, so the pattern operates in physical and biological media without any human role. And invoking it recognizes a loop arrangement already wired into a system rather than importing an interpretive frame — to spot feedforward is to notice that a consequence-model sits where it can still change the action. On every criterion the reading is structural, which is exactly what the 0.0 aggregate records.

Substrate Independence

Feedforward is a maximally substrate-independent prime — composite 5 / 5 on the substrate-independence scale, a pure loop topology recognized, never translated, wherever it recurs. Its domain breadth is at the ceiling (5 / 5): the predictive-pre-correction loop recurs with identical structural force across control engineering (the feedforward controller pre-setting an actuator against a measured upstream disturbance), human-computer interaction (preview thumbnails, "this will delete 37 files" dialogs, dry-run modes), organizational planning and governance (pre-mortems, environmental-impact statements, budget projections), training and instruction (worked examples and simulators shown before the attempt), neuroscience and motor control (efference-copy and forward-model circuits in the cerebellum), and software operations (dry-run plans before applying a change) — spanning physical, computational, biological, and institutional media indifferently. Its structural abstraction is likewise maximal (5 / 5): the signature names only an actor with an uncommitted action, a modelable disturbance, a predictive model placed upstream of commitment, a pre-action correction channel, and a residual feedback loop, with the substrate of the prediction explicitly incidental — controller, interface, forecast, or brain. Transfer evidence is concrete and strong (4 / 5): the identical loop is documented as feedforward control, affordance preview in HCI, pre-mortem and impact assessment in governance, and efference copy in neuroscience, with the same asymmetric-cost rationale and the same accuracy gate carried across — the score sits at 4 rather than 5 only because much of the cross-domain instantiation is recognized convergence rather than a single formal model demonstrably exported. The substrate-neutrality is genuine: there is no home vocabulary that must travel and no human role the pattern presupposes, which is exactly why the composite reaches the top of the scale.

  • Composite substrate independence — 5 / 5
  • Domain breadth — 5 / 5
  • Structural abstraction — 5 / 5
  • Transfer evidence — 4 / 5

Neighborhood in Abstraction Space

Feedforward sits in a moderately populated region (43rd percentile for distinctiveness): it has near-neighbors but no dense thicket of synonyms.

Family — Anticipation & Forward Models (15 primes)

Nearest neighbors

Computed from structural-signature embeddings · 2026-06-14

Not to Be Confused With

Feedforward's defining contrast is with feedback, its nearest neighbour by embedding and its structural complement. Both are corrective loops, and both route information to adjust an action — which is exactly why they are conflated and why the distinction is load-bearing. Feedback closes a loop after output is realized: it senses the deviation between actual output and a reference, then drives a correction, so its information is measured error and its timing is reactive. Feedforward opens a path before commitment: it interposes a model of the action's consequence between intention and irrevocable action, so its information is a prediction and its timing is anticipatory. The two partition the disturbance space by a clean seam — measurable-in-advance disturbances belong to feedforward, only-knowable-after-the-fact residuals belong to feedback — and most well-designed systems run both. The reason the distinction matters is that the two arcs demand opposite remedies. When a loop is too slow, the feedback instinct is to add faster sensors and tighter gains, chasing the deviation more quickly; the feedforward insight is that the latency can sometimes be eliminated outright by predicting the disturbance and pre-correcting before it lands. A practitioner who sees only feedback will keep tightening a reactive loop where a feedforward term on a measurable upstream disturbance would dissolve the problem — and, inversely, will try to model an unmeasurable disturbance that only feedback can catch. The diagnostic that separates them is whether the disturbance is knowable, and the action still changeable, before commitment.

Feedforward must also be held apart from foreseeing_prediction, with which it is merged because feedforward contains a predictive model and is sometimes loosely called "predicting ahead." The structural difference is that prediction is the model and feedforward is the placement of that model. A forecast, however accurate, is not feedforward unless it is positioned upstream of an irrevocable commitment where it can still alter the action. An impact assessment delivered after the resources are spent, a preview rendered after the click resolves, a weather forecast no one acts on — these are predictions without feedforward, because the action could no longer change when the prediction arrived. Conversely, feedforward's value is gated not on the prediction's mere existence but on two structural facts the bare concept of prediction does not carry: that the model sits before commitment, and that an accuracy threshold is met (an inaccurate forward model injects systematic bias the feedback loop must then fight, making it worse than no feedforward at all). The practical consequence is that "is the prediction good?" is the wrong first question; the right ones are "can the action still change when the prediction arrives?" and "does acting on this model beat not correcting?" Confusing prediction with feedforward leads to producing accurate consequence-models that land too late to matter, or trusting a forward model that is too biased to help.

These distinctions matter because each frame points at a different lever. If the problem is reactive correction, tune the feedback loop. If it is anticipatory correction, place a validated model upstream — and confirm both that the action can still change when it fires and that the model beats no correction. Collapsing feedforward into feedback chases speed where placement was needed; collapsing it into prediction produces forecasts that arrive too late to alter the action they were meant to guide.

Solution Archetypes

No catalogued solution archetypes reference this prime yet.