Decision Tree with Hold State¶
Decision procedure — instantiates Observational Equivalence Resolution
Routes an unresolved case into an explicit hold branch that takes a safe, reversible action and keeps the ambiguity live, instead of forcing a premature verdict.
The Decision Tree with Hold State is the mechanism for what to do when the evidence has not decided. An ordinary decision tree branches to an action on each outcome; this one adds a first-class hold branch that fires precisely when the discriminating evidence fails to clear the bar. Instead of snapping to whichever candidate is momentarily ahead, the hold branch commits to an action that is safe across the whole live equivalence class — reversible, low-commitment, buying time — and keeps the case flagged as unresolved until better evidence arrives or a deadline forces the issue. Its defining move is treating "not yet distinguishable" as a decidable state with its own action, not a gap to be papered over. It resolves nothing itself; it governs conduct while resolution is pending, so that ambiguity is neither ignored nor allowed to paralyze the system.
Example¶
A self-driving car approaches a crosswalk and its perception stack cannot separate two readings of a figure at the curb: a pedestrian about to step into the road, or a pedestrian who will stay put. The sensor evidence is genuinely equivocal — the same posture and position fit both. A naïve controller would pick the more probable reading and act on it; if it guesses "will stay" and is wrong, the failure is catastrophic and irreversible.
The decision tree instead routes the case to its hold branch, because the threshold for committing to a normal "proceed" maneuver has not been met. It takes the action safe under both readings — ease off the accelerator, widen the following gap, continue at a speed from which it can still stop — while tagging the situation as an unresolved ambiguity the planner must keep watching. This is a minimal-risk maneuver: the conservative fallback a system takes when it cannot confidently resolve its situation.[n1] The monitor holds the case open for the next few frames; if the figure steps out, the evidence now clears the bar and the tree branches to "yield and stop"; if they turn away, normal speed resumes. Either way, the hold bought the seconds that made a premature, irreversible commitment unnecessary.
How it works¶
The tree is a branching policy with an explicit undecided outcome. Each terminal action sits behind an evidence gate: a committed maneuver fires only when the discriminating evidence for its reading clears a threshold set by the stakes and the reversibility of the action. When no branch's gate is met, control falls to the hold branch, which selects the action that is robust across the remaining candidates — reversible, staged, or delayed — and labels the case as unresolved so nothing downstream treats it as decided. A monitor keeps the held case live, watching for the new evidence that would trip a gate, or for a timeout that forces a robust default when waiting is no longer safe. The point is that the hold is bounded: it is a considered pause with an exit condition, not an indefinite stall.
Tuning parameters¶
- Commit thresholds — how much discriminating evidence each branch requires before firing. Raise them and the tree holds more often (safer, slower); lower them and it commits sooner (faster, riskier).
- Hold action — what the system does while undecided: idle, take a reversible partial step, or fall to a conservative default. The more robust the hold action across candidates, the longer it can be sustained.
- Timeout — how long a case may stay in the hold state before a default is forced. Short timeouts avoid dithering; long ones wait for evidence at the cost of delay.
- Asymmetric gating — setting a much higher bar for the irreversible or catastrophic branch than for the safe one, so the tree is biased toward caution where the cost of a wrong commit is worst.
- Re-entry rule — whether a resolved case can be pulled back into the hold state if fresh evidence reopens the ambiguity.
When it helps, and when it misleads¶
Its strength is that it lets a system act responsibly under unresolved ambiguity without either freezing or bluffing: by making "hold" a real branch with its own safe action, it converts the instinct to guess-and-commit into a governed pause, invaluable wherever a wrong irreversible commitment is far costlier than a delay.
Its failure modes are the two ways a hold can go wrong. Dithering is the hold that never resolves — the timeout is too long or absent, so the system stalls, which in a moving vehicle is its own hazard. The opposite is a hold action that is not actually robust: the fallback was assumed safe across all candidates but is quietly bad under one, so the pause causes harm while looking cautious. A tree can also launder ambiguity by labeling a case "held" while downstream consumers silently read the leading candidate anyway. The guarding discipline is to bound every hold with an exit condition, verify the hold action is genuinely safe under each live candidate rather than on average, and keep the unresolved label binding. Note the division of labor: this tree acts on unresolved cases in real time, whereas the Ambiguity Register records and routes them for later disposition — the register is the ledger, this tree is the live branch.
How it implements the components¶
The Decision Tree with Hold State realizes the ambiguity-governance face of the archetype — deciding conduct when discrimination has not resolved:
decision_rule_under_ambiguity— the branch logic that, when no candidate clears its gate, selects a robust, reversible, or delayed action instead of a committed one.uncertainty_label— the unresolved tag attached to the held case and carried into whatever action the hold branch takes, so nothing downstream treats it as decided.evidence_threshold— the per-branch gate that decides whether incoming evidence is decisive enough to leave the hold state and commit.residual_ambiguity_monitor— the watch on the held case for the new evidence, or the deadline, that will move it off the hold branch.
It gathers no discriminating evidence of its own: it neither designs a probe (discriminating_test_design, that is Ablation or Perturbation Test) nor reads a tell (discriminating_observable, the discriminating mechanisms such as Differential Diagnosis Protocol and Forensic Discriminator); it governs what to do once discrimination has failed.
Related¶
- Instantiates: Observational Equivalence Resolution — this tree is the archetype's live decision governor for cases that stay unresolved.
- Consumes: the discriminating mechanisms (e.g. Differential Diagnosis Protocol, Ablation or Perturbation Test) supply the evidence its gates test; when they fail to clear a gate, this tree takes over.
- Sibling mechanisms: Ambiguity Register · Differential Diagnosis Protocol · Causal Identification Probe · Ablation or Perturbation Test · Forensic Discriminator · Frame-of-Reference Shift · Side-Channel Measurement · Controlled Disambiguation Test
Editorial Notes¶
Form Classification¶
Form family: Decision, Gate & Allocation
Rationale: Evidence gates route each case to a committed response or to an explicit unresolved hold with a safe reversible action, so the mechanism's defining output is a bounded disposition under ambiguity.
Nearest alternative: Control, Automation & Runtime — A monitor keeps held cases live and may trip later branches, but the tree primarily selects the case's current decision state rather than regulating an operating variable.
Review outcome: Adjudicated after independent review; high confidence.
Origin Attribution¶
Primary origin: Computer Science & Software Engineering
Origin pattern: Cross-disciplinary synthesis
Present-day reach: Multi-domain
Rationale: Pattern recognition and information theory are primary because classification with a reject option established an explicit no-decision branch when posterior evidence fails a threshold. Sequential decision analysis supplied bounded delay and further information gathering; safety engineering supplied the robust reversible action and monitored exit condition used while the case remains unresolved.
Related originating lineages:
- Engineering & Design — Safety engineering supplied minimal-risk hold actions, reversibility, monitoring, and time-bounded fallback when ambiguity persists.
- Operations Research — Bayesian and sequential decision theory supplied thresholded defer-or-gather-information policies and the cost tradeoff between error, rejection, and waiting.
Review resolution: Pattern recognition and information theory are primary because classification with a reject option established an explicit no-decision branch when posterior evidence fails a threshold. Sequential decision analysis supplied bounded delay and further information gathering; safety engineering supplied the robust reversible action and monitored exit condition used while the case remains unresolved.
Attribution caveat: The reject branch is established computational decision theory; the safe action and live monitor are an engineering synthesis layered onto it.
Encyclopedia synthesis: The exact catalogued form synthesizes established practice rather than reproducing a single standard historical label.
Review outcome: Researched adjudication after independent review; high confidence.
Sources consulted:
- Chow (1970): On optimum recognition error and reject tradeoff
- Classification Under Partial Reject Options
Notes¶
The hold branch is neither a low-confidence commitment to the leading candidate nor an open-ended stall — it is a distinct, decidable state with its own safe action and its own exit condition. A tree that lacks it will, under pressure, always convert ambiguity into a false verdict.
[n1] In automated-driving safety, a minimal-risk maneuver (and the minimal-risk condition it reaches) is the deliberately conservative fallback — slowing, pulling over, stopping in lane — that a system executes when it cannot confidently continue its normal task. It is the engineered embodiment of a hold state: a pre-defined safe action taken precisely when the situation cannot yet be resolved with confidence. ↩