RFC 0007: The determination boundary — what a pack cannot hold

• Status: Draft
• Type: Standards-track (candidate Core amendment or profile)
• Created: 2026-07-27

This is an open proposal, not part of the specification. See RFC 0000 for the process and evidence bar. It reports a boundary found by encoding real regulatory text as a pack, and states the questions that follow. It proposes a settled answer to none of them.

Summary

Core draws its expressive boundary at computation: a pack compares facts and does not calculate. Two encodings of real policy — deliberately opposite in character, by different authors — show that the boundary which actually appears in use is determination, and that the two are not the same line. In both, part of the judgment left the pack along with the work it could not express, into a preparation layer the format cannot describe, attribute, or audit.

The size of the effect is strongly domain-dependent, and this RFC is bounded by that: encoding arithmetic-dense regulation displaced 124 facts and 13 determinations, while encoding qualitative policy displaced 5 and 1. The format fits qualitative policy well. But the escape did not vanish there — it merely changed cause. In the first domain judgment escaped because the format cannot calculate; in the second because it cannot quantify over a collection. Neither gap predicts the other, and that is the point: the boundary is not a property of arithmetic. It opens wherever the format lacks the device a policy sentence needs, and what falls through is a judgment rather than a value.

The specification has no vocabulary for that loss, because it only ever contemplated losing calculation.

Problem

§2.2 is explicit that satisfying the decimal grammar "does not imply executable comparison support", and Core defines no arithmetic. The intended reading is a clean separation: facts are prepared elsewhere; the pack states the policy over them. That reading assumes the things prepared elsewhere are inputs — observations, measurements, values of record.

In practice the preparation layer also ends up holding conclusions. Deciding which provision an input falls under is itself an application of the rules; when that decision requires comparing against a computed quantity, the format cannot express it, so it migrates. The pack then states the rules correctly given a classification it did not make and cannot show.

Three consequences, none visible in the format today:

1. The decision is split across two artifacts and only one is portable. A reader of the pack cannot see how much of the determination happened before evaluation began.
2. The disposition cannot attribute what it relied on. It reports which rules fired; it cannot report that the outcome depended on a prepared value, computed by whom, under which reading.
3. A conformant pack may be unable to reach an outcome a competent reader reaches, because the format can say a fact is unknown but not that it is too small to matter.

Evidence

Two encodings of real policy, deliberately chosen to be opposites, by different model families under a clean-room barrier. Study 002 was run specifically to falsify the generalisation drawn from Study 001; it did not.

Study 001 — arithmetic-dense regulation

Encoding the NBA Collective Bargaining Agreement excerpt published by RuleArena (ACL 2025, MIT, commit 3b9e2256) as a pack. The pack validates, and its 61 rules cover 61 / 61 of the vocabulary the benchmark uses to annotate which provisions govern each instance. Full report: expressiveness note.

E1 — 124 facts had to be prepared outside the pack. 39 always-present booleans, 18 conditional booleans, 67 conditional decimal strings. Expected, given no arithmetic.

E2 — the escape is not confined to arithmetic, and this is the finding. 13 of the 39 always-present booleans name which Salary Cap Exception a transaction invokes. The source prose never states one. Choosing "the lowest tier whose limb covers the amount" is a legal determination that happens to require comparison against a computed limb — so the format cannot hold it, and it left. Any evaluation of such a pack is partly an evaluation of its preparation layer, and nothing in the disposition says so.

E3 — unknown is not immaterial. 28 % of instances (60/216) could not be decided, almost all tracing to one absent constant (a minimum-salary schedule). A preparer that refuses to invent the figure omits the field; applicable rules go unknown; the pack escalates. A competent reader treats those contracts as immaterial and proceeds. The three-valued model distinguishes true, false, and unknown, and cannot express "absent, and below the magnitude that could change this outcome."

E4 — §8's conflict rule constrains architecture, undocumented. Two distinct candidate outcomes resolve to conflict → unresolved. A pack holding both "this is legal" and "this is illegal" rules therefore goes unresolved whenever one of each fires, which in a domain of independent provisions is routine. The pack became 61 violation detectors all resolving to illegal, with the benign outcome reachable only as fallbackOutcome. That works, and it makes every fired rule a citation — but it is a shape the author had to discover, not one the specification describes.

Study 002 — qualitative policy, run to falsify the generalisation

Encoding the "Cancel flight" decision from the airline agent policy published by τ-bench (MIT, commit 1d244f5d) — 166 lines of real customer-service policy: deontic, procedural, essentially arithmetic-free. Authored clean-room by a different model family, with a null result licensed in advance. Full report: Study 002.

E5 — the effect is heavily domain-dependent, and that bounds this RFC. Prepared facts fell from 124 to 5, prepared determinations from 13 to 1, and seven policy devices were carried with no friction. The resulting pack is two outcomes, one rule, one escalation exception. The format fits qualitative policy far better than it fits regulatory computation, and any reading of E1–E3 that treats them as a general indictment is wrong.

E6 — but the escape reproduces, through a device gap that has nothing to do with arithmetic. The single prepared determination in Study 002 is forced by this sentence:

"If any portion of the flight has already been flown, the agent cannot help and transfer is needed."

Core has no quantifier over a runtime collection. any combines a fixed, authored list of conditions; it cannot ask whether some element of a variable-length array satisfies a predicate. So the pack cannot decide "any portion already flown" — something upstream must inspect the segments and hand it the conclusion.

This is the load-bearing result of the pair. In Study 001 judgment escaped because the format cannot calculate; in Study 002 it escaped because the format cannot quantify. The determination boundary is therefore not a property of arithmetic. It appears wherever the format lacks the device a policy sentence needs, and what is lost is a judgment rather than a value. Two domains, two unrelated missing devices, the same class of loss.

Study 002 also reproduced E4 independently: with no rule precedence available, its author encoded the policy's "if … otherwise" ordering as an escalate exception plus a single any rule plus a fallback — a different shape from Study 001's 61 violation detectors, arrived at by a different author for the same reason. Two observations of one cause.

Exposure recorded by that study, and it cuts against its own finding: the authoring brief disclosed what Study 001 had found, which primes toward finding migration. Mitigations and the reasoning are in the study's report; a future replication should withhold the prior result.

Study 003 — the census, answering "how common?"

A preregistered census over all twelve separable decisions in τ-bench's airline and retail policies, with the blinding Study 002 lacked: hypothesis-blind encoders in isolated rooms producing neutral fact ledgers, then a separate measurement pass — two independent classifiers, adjudicated, disagreement counts published (zero fact-class disagreements across 84 facts). All twelve packs validate. Full report: Study 003.

E7 — in this frame the escape is universal, and one device dominates. Every decision required at least one prepared determination (12/12; 40 of 58 prepared facts). The registered "all-scalar decisions escape cleanly" clause proved vacuous rather than false: no such decision exists in either policy, because request inputs are collections — passengers, payment methods, segments, items — and every all / any / each / how many over one of them is a conclusion the format cannot draw. Collection-quantification forced 25 of the 40 determinations (63%); arithmetic — which Study 001 made look like the phenomenon — forced 6. Within the 25, ~~roughly 19 are ∃/∀-shaped~~ — amended 2026-07-27: 3, or 5 with a uniformity operator; see the amendment under question D and RFC 0008 — and 6 are count-shaped (count(filter(…)) against a limit — meaningfully more language than a quantifier).

E8 — measurement method moves the count. Study 002's decision, blind-re-encoded and independently classified, yielded 4 determinations where the original self-report found 1. Absolute counts in Studies 001–002 should be read as floors; the census's two-classifier adjudicated protocol is the reference method. The residues (55 sentences) also gave the previously unevidenced candidate areas their first evidence — deontic gradation, cross-decision references, derivation-versus-check, terminality — while confirming that most sequencing loss is agent procedure a decision format arguably should not hold.

Specification (sketch — deliberately unsettled)

Three separable questions. E2 motivates the RFC; A and B are the mechanisms that would address what was observed.

A. Attribution of prepared facts. Candidates, cheapest first:

• Nothing. Preparation is product territory, as acquisition is. Cost: the split stays invisible.
• Declaration. A pack may declare which fact pointers it expects to be prepared, with a prose definition and a source reference. No new semantics, no computation — the dependency becomes visible in the document and checkable against an input.
• Attribution in the disposition. The result records which declared-prepared inputs it read, so a decision record is self-contained about what it leaned on.
• A computation profile. An optional profile with a small arithmetic vocabulary. The largest change, and the one most likely to re-open the general-purpose-rules-language non-goal.

B. Materiality. Candidates:

• Nothing. Immateriality is a policy judgment; the preparer decides, and supplies a value or not.
• A declared threshold. A rule or evidence requirement declares a magnitude below which a missing or differing input cannot affect it, so an absent immaterial value yields no effect rather than unknown.
• A fourth condition value. Recorded so the option is visible; rejected on sight for what it would do to §7.

D. Collection quantification (E6, E7). No longer merely the best-evidenced gap — the census shows it is the dominant cause of determination escape: 25 of 40 across twelve decisions, versus 6 for arithmetic. Candidates: nothing (a prepared boolean per conclusion, at the cost E7 quantifies); or an existential/universal condition over an array-valued fact pointer — a small, bounded addition to §7, no arithmetic implied, ~~expressing roughly nineteen of the twenty-five observed cases~~. The six count-shaped cases (count(filter(…)) compared to a limit) are named and deliberately deferred: counting-with-predicates is the first step of a query language, which the non-goals exclude. Of everything in this RFC, the bounded quantifier is ~~the most likely to be worth doing~~ the cheapest to specify, and now the best evidenced.

Amended 2026-07-27 — the "roughly nineteen" figure is withdrawn. It was estimated from shape names rather than per-fact reading. A per-fact re-analysis of the same 25 facts — two separate blinded model runs with recorded adjudication; expressibility agreement 25/25 — puts a bare bounded quantifier's reach at 3, or 5 with an additional all-equal-at-a-sub-path operator; the remaining 20 need joins, counts, ordinal selection, fact-to-fact comparison, or whole-list classification. The count-shaped figure of six stands. The re-analysis is post-hoc and unregistered, not census-grade. E7's device count (25 of 40) is unaffected; only the sub-classification within it changes. The proposal, the corrected numbers, and the question of whether 3 of 25 licenses a Core amendment at all now live in RFC 0008.

C. The architectural constraint (E4). Either document the single-outcome / violation-detector pattern as non-normative authoring guidance, or revisit whether §8's conflict rule should treat a legal-vs-illegal opposition differently from genuine ambiguity. Guidance is almost certainly the right answer; the point is that today there is neither.

The strong prior: A-declaration and B-threshold are the smallest changes that fit the evidence, and a computation profile should be resisted until a second domain shows the same need.

Alternatives

• No change. Viable, and should be taken seriously. Everything here can be pushed onto the preparer, which is where it already is. The cost is auditability: the pack is inspectable and the half of the decision sitting beside it is not.
• Profile, not Core. Both mechanisms could live in an optional profile, leaving Core minimal. Probably right if a computation vocabulary is ever adopted.
• Guidance only. Document the boundary and the E4 pattern as authoring guidance and change nothing normative. Cheapest, and possibly sufficient for E4 alone.

Compatibility

Declaration and threshold are additive; no existing pack becomes invalid. A computation profile would be a new optional profile, not a Core change. The document-conformance classes are untouched under every option.

Security and privacy

Attribution helps: a decision record naming its prepared inputs is harder to launder, and E2 is exactly the case where a reader would want to know. A materiality threshold cuts the other way — it lets an author declare an input immaterial and thereby suppress an escalation that should have happened. Any threshold mechanism needs its abuse case written before adoption, and a conformance case that attempts it.

Conformance

If either mechanism is adopted: positive cases where a declared prepared field is present, absent, or present-but-immaterial; negative cases where a pack cites an undeclared pointer as prepared; and adversarial cases where a materiality declaration suppresses a required escalation. The 124-field contract from Study 001 is the concrete artifact any proposal should be checked against: does the mechanism let that pack express more of its own decision, or does it merely relabel the boundary?

Implementation

None. The evidence is the absence of these features in the two evaluators that agree 13/13 on the current semantics (RFC 0006, whose caveat applies here: both trace to one maintainer's direction, so they are not independent) — the gap is in the format, not in an implementation of it.

Unresolved questions

• Is preparation inside the specification's remit at all? The non-goals exclude retrieval and orchestration. Prepared inputs sit comfortably outside; prepared conclusions (E2) do not obviously sit anywhere.
• Does materiality belong to the rule, the evidence requirement, or the fact?
• Would a computation profile re-open the general-purpose-rules-language non-goal? Probably — which is the argument for declaration over computation.
• What else escapes? E2 was found by encoding one domain and counting what migrated. That is a repeatable method and the honest way to answer this, rather than by speculation. Areas any decision format plausibly has to face, none of them evidenced here: obligations and prohibitions as distinct from outcomes; ordinal or rubric assessment as distinct from boolean conditions; precedence between provisions. Whether JPS needs any of them is an empirical question for the next encoding, not a claim of this RFC. Two of those guesses have since been resolved by Study 002: precedence is now evidenced twice (E4), and quantification — which was not on the list at all — turned out to be the concrete gap (E6). The method found something the speculation did not, which is the argument for continuing to run encodings rather than enumerating candidates.
• ~~Does E1–E3 reproduce outside arithmetic-dense regulation?~~ Answered by Study 002: yes, ~25× weaker, and via an unrelated missing device (E5, E6).
• ~~How common is the escape in qualitative policy?~~ Answered by the Study 003 census (E7): universal in its frame — 12/12 decisions — with quantification as the dominant device. The open form that remains is external validity: the frame is two policies by one benchmark team, and the rate outside it is unmeasured. A census over a differently-authored corpus is the natural replication.