# Boundary-contract extension — exact guards and observation limits

10 October 2026. This research record retains the work originally planned as Essai 08; its public contribution is folded into the existing Boundary Contract Lab and method essay at Lawrence’s request, rather than published as another essai. Baseline `4316df420184af3349a450a0f14e43f9efa91b30`. This extends the existing Boundary Contract Lab and preserves its original scenarios and engines.

## What changed in understanding

The intended compression question contained a constraint worth making explicit: the state of this finite model **is** the named completed-occurrence set, and the promised observation includes that set. Any observation-preserving summary is therefore injective. A smaller behavioural quotient is unavailable under this guarantee. This is an elementary consequence of the model and observation definition, not a novel theorem about open Petri nets or projects generally.

That does not force us to store the expanded tables. The new guard contract retains source incidence as internal input/delta and symbolic boundary input/output vectors, initial ownership and supplies, and occurrence identities. It derives markings and enabled joint steps when requested. The reconstruction is exact for the admitted once-only additive model. This is essentially a compact rule encoding of the original mechanism, rather than a discovery of less information sufficient for the same task. It still enumerates the reachable graph during comparison.

The alternative-support case gives a useful contrary result to a universal information-cost claim. Initial support 0, 1 and 2 produces respectively 7/12, 8/19 and 8/19 states/steps. Adding the first token permits C-first. All three contexts use the identical receiver guard. The second extra token changes observed counts without changing the action options. The model grows in reachable possibilities without the receiver contract becoming inadequate.

A deliberately defective resource-count-only quotient merges the states after A and after B. The union of its individually valid edges invents A then A. That violates once-only work and the uniform successor requirement. Under the actual chosen observation it is already invalid because A-complete and B-complete differ. The example prevents weakening a test to collected edge sets and calling the resulting paths preserved.

## What this extension adds

- The observation-relative no-merging argument and its explicit separation from encoding size.
- A working, table-free guard encoding and composer for the existing admitted contract class.
- Independent source-net and uniform continuation checks over the new context grid.
- A linked interactive comparison and reusable finite case export.

Earlier closed-summary failure, enriched-contract/direct-net equality, alternative suppliers and open-net assembly remain prior results. This extension does not rebrand them as discoveries.

## Verification and reproduction

With Node 22 or later, from the repository root:

```sh
node --test tests/*.test.mjs
node scripts/export-contract-compression.mjs
```

The focused test is `tests/contract-compression.test.mjs`. Its separate reference helper has no production imports. It glues source places with its own union algorithm, traverses legal serial histories, reconstructs markings by firing, and allocates distinct tokens to joint input slots. Tests compare every reached observation and every state's outgoing labelled successors, then enumerate all finite joint-step prefixes for cases of at most four events. They also reject every pairwise state merge under the completed-work observation. Every nonidentity partition contains such a pair, so no separate Bell-number enumeration is needed.

Coverage includes seven retained cases; 81 combinations of internal permit, external initial support, production and demand (each 0–2); the three supply cases; renaming and disjoint-composition controls; the defective repeated occurrence; incomplete-cap handling and invalid wiring. The eight-independent-event example preserves 256 states and 6,305 joint steps. JSON encoding size is 310,165 bytes for the rich contract after removing its audit source copy, versus 1,318 bytes for the guard contract. Keys and declarations are included. These numbers are not a runtime benchmark, an information measure or a least-encoding result; small cases can get larger.

All 196 repository model tests passed locally, including nine focused compression tests. Desktop (1440px) and phone (390px) browser checks passed with no page errors or horizontal overflow. A final code review also corrected inherited-property lookup for legitimate place identifiers such as `constructor`; a dedicated regression now covers these names and empty work. These are local checks, not a deployment claim.

The first independent-reference run exposed a reference-only bug: sorting a place-class array while mapping it skipped an ownership key. The reference was corrected to sort before mapping. The candidate did not change to fit the faulty output. All focused checks subsequently passed. This record preserves the failure because independence does not make an oracle infallible.

The local browser script is `scripts/contract-compression-browser.mjs`; it accepts `LAB_BASE_URL`, `PLAYWRIGHT_MODULE` and `LAB_EVIDENCE_DIR`. It checks root arrival, all three contexts, C-first and complete work, no repeated event, reset/reload and keyboard activation, the defective-merge explanation, method navigation, the existing one/two-lift controls, and desktop/390px overflow. It uses disposable browser contexts. Passing it is agent-operated UI evidence, not a human-use study.

## Exact public case boundary

`cases.json` is generated from the source fixtures and carries the baseline pin, adapter semantics, initial states, observations, labelled atomic joint steps, complete flags and explicit summary maps. It has no probabilities, durations, adversarial outcomes or inferred causal model. It preserves the completed-work observation that rules out state merging. Comparison with other forays must state these differences rather than forcing the cases into one story.

## Source → construction → limit

- Baez and Master, *Open Petri Nets*, arXiv 1808.05415 v6, introduction and operational/reachability comparison: explicit boundary gluing, with laxness warning for a reachability-only summary. The original bounded enriched contract remains our construction.
- Baez, Genovese, Master and Shulman, *Categories of Nets*, arXiv 2101.04238 v2, §3 including Proposition 3.3 and the token-reuse warning: count/monoidal semantics, with occurrence-labelled joint steps retained separately.
- Wynn and Clarkson, *Improving the engineering design process by simulating iteration impact with ASM2.0*, DOI 10.1007/s00163-020-00354-5, §6.2.3 and §7: the context-maintenance objection and the distinction between trace checking and calibration to reality.

The existing lab links [the integrated method section](../../planning/boundary-contract-method.html#compact-guards), which connects these sources to the actual construction. Private reading records are not included here.

## Judgement

This advances the native reuse question by distinguishing two forms of “smaller”: fewer behavioural states is impossible under the declared observation, while fewer stored table rows is straightforward and exact. The storage result alone does not justify a new planning tool; retaining the full source mechanism may be the simplest sufficient contract. The most informative outcome is that increased external supply does not force a new receiver schema in this case. A later enquiry that wants genuine behavioural compression must justify a weaker observation, without silently removing required task identities or allowable continuations.

The result remains restricted to the original finite once-only class. It does not establish schedule optimality, useful time savings, real engineering validity, a new general abstraction method or a human benefit. Preparation, checked source, pushed PR, merge and deployment are separate states.

## Publication fit — 10 October 2026

Lawrence chose to fold this modest advance into the existing boundary-contract essay and lab. The root’s additional essay card and the Essai 08 label are removed. `planning/boundary-contract-method.html#compact-guards` now carries the argument, source relationship and limits; the former explanation URL redirects there. The lab retains the three supply contexts and the original one/two-lift workbench. Model code, tests and the machine-readable fixture are unchanged by this editorial revision.

The updated ordinary-browser journey passed at 1440px and 390px: arrival through the existing lab, all supply controls, completion and once-only behavior, reset/reload/keyboard, integrated method navigation, old-URL redirect and retained lift controls. There were no page errors or horizontal overflow. Nine focused model tests passed again. These checks establish retained behavior and navigation, not publication significance or human value.
