# Release-packet contribution and claim map

This is a research-prototype packet. It supplies an auditable argument and
reproduction objects for the working publication sequence; it does not declare
all planned papers submission-ready. The claim labels use the repository's
science-writing vocabulary: **P** prior work, **D** design, **T** conditional theory,
**I** implementation/synthetic experiments, **E** empirical evidence, **H** open
hypothesis. No human-subject or field efficacy result is claimed here.

## Research questions and bounded answers

| Manuscript | Question | Answer supported by this packet |
|---|---|---|
| [Synthesis](manuscripts/synthesis.md) | Can a small shared boundary support work across different internal organizations? | The proposed boundary can be specified and exercised; human benefit and adoption remain open. |
| [Mathematical foundations](manuscripts/math-foundations.md) | Which allocation, completion, repair and audit guarantees survive explicit assumptions? | Four conditional models yield exact bounds and counterexamples; their premises must not be combined silently. |
| [Coupled agents](manuscripts/coupled-agents.md) | Do local allocation properties survive capacity, completion and reliance selection? | The retained synthetic pipeline exhibits failures across those stages under one shared person ledger. |
| [Local rules](manuscripts/local-rules.md) | Can a small recognition loop improve attention allocation without entrenchment or gaming? | It has a conditional attention ceiling, but retained tradeoffs and a profitable correction-credit deviation prevent a general incentive claim. |
| [Funded audits](manuscripts/funded-audits.md) | Does an announced inspection probability correspond to funded delivered inspection? | A fixed-population toy can deliver concealed quota marginals under stated resource/information premises; broken promises, early disclosure and false sanctions defeat broader interpretations. |

## Contribution categories

| Category | What is present | What is not claimed |
|---|---|---|
| P: established work | Provenance, research objects, versioned review, contribution roles, notifications, elementary probability and stability methods | Invention of those ingredients or an exhaustive priority search |
| D: synthesis and profile | Offer/check/rely/amend; exact target, scope, authority, currentness and repair commitments; shared-resource accounting | An adopted standard or a mandatory institutional rulebook |
| T: derivations | Group-panel pushforward, completion envelopes, conditional repair bound, audit constraints, local feedback and finite deviations | A whole-network cooperation equilibrium or an empirically established social law |
| I: executable evidence | Typed toy receipts, synthetic queues, event traces, funded inspections, independent oracles and adversarial regressions | Authenticated identities, actual legal authority, production transactions or real delivery |
| E: human/field data | None | Measured usefulness, lower labor, better science or community adoption |
| H: open questions | Semantic portability, human comprehension, contribution incentives, sustained independent repair and uptake | A positive answer inferred from passing code tests |

The dated [closest-work comparison](publication/positioning-and-adoption.md) and
[source ledger](publication/positioning-sources.md) cover selected feature-level
neighbors. Model-specific primary sources and derivations appear in the named
manuscripts. This is a targeted comparison, not a systematic review of every
related protocol or deployment.

## Consequential claims and evidence

| ID | Label | Claim | Evidence and stopping premise |
|---|---|---|---|
| C01 | D/I | Exact target, per-scope coverage, authority kind and amendment state are distinct | `toy_agents/protocol.py`, its tests and recorded scenarios; all identities/roles remain trusted fixture inputs |
| C02 | T/I | Redundant labels cannot change a fixed group-first panel distribution | Mathematical proof and finite enumeration; stop when true control, eligibility or weights change |
| C03 | T/I | Completion and final reliance may change population composition | Conditional probability derivations, full denominator traces and coupled sweep; no invitation guarantee transfers automatically |
| C04 | T | A common positive repair envelope bounds expected cumulative weighted work | Conditional-expectation proof; weights are not human hours unless calibrated to bound those costs |
| C05 | T/I | A hard audit quota differs from an expected expenditure limit | Exact rational quota sampler and independent endpoint/utility probes; delivery and enforcement require explicit premises |
| C06 | I | Coupled retry/capacity/reliance choices can defeat narrow allocation interpretations | 7,400 retained runs, 888 replicate intervals and independent reproduction; hypothetical utilities and known simulation oracle |
| C07 | T/I | Capped recognition bounds offered attention, but can leave poor coverage or repair capacity | Convex-update/softmax derivation and all 216 ecology runs; no endogenous cooperation model |
| C08 | T/I | Manufacturing correctable defects can pay under specified correction credit | One-step payoff derivation, 18 exact cases and independent reconstruction; externalized costs and limited action set stated |
| C09 | T/I | Concealed funded inspection can match the participating actor's conditional audit marginal | Funded-audit proof and event/accounting tests; fixed offered population, independent hidden selection and executing auditors |
| C10 | I | Early audit disclosure or inaccurate promises change the relevant incentive | Retained funded-audit negative controls; no real actors or sanctions |
| C11 | T/I | The four domain calculations expose distinct scientific/operational limits | Domain oracles, real toy-receipt replays, before/after adversarial results; synthetic inputs only |
| C12 | D/I | A legacy wrapper must retain missingness and semantic loss without creating authority | Adapter fixtures and canonical-type regression; reference syntax is not content authentication |
| C13 | I | Candidate packaging rejects the specified path, inventory and evidence-binding attacks | Harmless sentinel probes and independent repaired regressions; scanner and integrity scope are deliberately bounded |
| C14 | T/I | Tiny mock randomization designs have limited sharp-null power | Fifteen independently enumerated artificial potential-outcome cases; not a recommended real sample size |
| C15 | H | The human interface improves adequate reliance decisions at acceptable labor | Pre-results evaluation plan with strong template baseline and positive control; no study conducted |
| C16 | D/H | Existing scholarly systems can carry a useful MCRP use/capacity profile | Proposed semantic crosswalk and round-trip oracle; no external standards conformance or independent adapter validated |
| C17 | D | A curated public research seed is an appropriate next step | Internal publication assessment and collective critique; attribution, rights, destination and exact release remain human decisions |

## Methods, negative controls and review status

Each experiment states its fixed assumptions, comparison policies, negative cases
and interpretation of uncertainty. The root reproduction command retains failures
as evidence and records code, fixture, table and saved-result hashes. The
[reproduction guide](publication/reproduction.md) distinguishes core tests from
optional renderers. The [collective report](reviews/COLLECTIVE.md) records adversarial
findings and dispositions rather than treating another agent's agreement as proof.

G0-G3 have internal argument, scholarship, methods and artifact evidence at a
prototype level, with the limitations stated above. Separate AI drafting/review
lanes do **not** replace the repository's independent-review and human-release
requirements. G4/G5 are not claimed passed for journal submission or operational
use. The static protocol seed has a narrower release question; the user requested
its preparation and readiness assessment, not an invented record of human approval.

Stop an extended claim when its needed premise is missing. Prefer a simpler
structured record or uniform assignment if the additional mechanism has no
worthwhile advantage. Those are research and design stopping rules, not penalties
for contributors who refuse work or disclose an honest correction.

## Agent-first release extension

The [agent entry contract](agent_api/README.md) adds a bounded local JSON interface
and machine-readable discovery. This is an implementation result (I) and design
choice (D), not evidence of agent identity, real independence or human efficacy.
The full offered profile binds claim/evidence, scope, resolved dependencies and
required coverage; a metadata-only change therefore has a distinct version.
Aliases resolve through a separately loaded trusted policy to one principal's
ledger. Conditional alias invariance requires correct principal mapping and
does not discover undisclosed common control.

Author and independent adversarial suites exercise those boundaries. Four domain
traces preserve old pending reliance while establishing reliance on a changed
version. The [agent launch critique](reviews/agent-launch-red/REPORT.md) records
its own findings and limitations. The integration runner includes these suites,
input fixtures and regenerated outputs in its validation inventory.