# Appendix E — Development History and Its Evidential Limits

## E.1 Scope and Source Basis

This appendix consolidates developmental history that helps explain the choices
examined in the main argument. Its implementation basis remains commit
`a602024416ace9652f61bb2043889c307fd4e09d`. Historical descriptions establish
what was proposed or recorded at an identified time; they create no present
obligations and do not refresh product or regulatory claims.

Original January–April repositories were deleted during migration. The owner
reports a January governance prototype, a February graph-first attempt, March
rebuild code and a largely quiet April. Surviving migrated Git history starts
with the May 18 snapshot. Earlier intent is testimony, not reconstructed commit
evidence. This limitation applies to causal and continuity claims about that period.

The regulatory mapping preserves its **2026-07-02** source editions and authorial
interpretation. Revising the architecture does not establish those sources'
present currency or legal compliance. Chapter 6 and Appendix C retain that limit.

## E.2 Development of the Production Environment

From the surviving basis, July separated roles, methods, tools and deliverable
content. August reconsidered delegation and the separation of private knowledge
from public mechanics. September reduced specialist roles and supplier wrappers.
October removed administrative duties and their enforcement together. These
changes are relevant because Root's artifacts shape the means through which
participants produce and revise later artifacts.

The October work included instruction revision, dependency migration, verification
selection and preservation of live conditions before records were archived.
[PR #1212](https://github.com/sgttomas/chirality/pull/1212),
[App v4 PR #1213](https://github.com/sgttomas/chirality/pull/1213),
[Piping PR #1216](https://github.com/sgttomas/chirality/pull/1216) and the archive
batches document that work. Recoverable Git tags preserve prior material.
Independent reviews found omissions, including technical conditions and frozen
numerical references with maintained consumers. Repairs support an account of
examination and correction; scripted production alone would not establish completeness.

## E.3 Representations of Work and Warrant

### E.3.1 Deliverable Lifecycle and Location

The six-state model used `OPEN`, `INITIALIZED`, `SEMANTIC_READY`, `IN_PROGRESS`,
`CHECKING` and `ISSUED`, with `_STATUS.md` as the intended canonical indicator.
The `1_Working`, `2_Checking` and `3_Issued` folders coupled artifact location to
production stage. These mechanisms made some distinctions explicit, but maintained
another description of work that could diverge from artifacts and imply more
than a label supported.

The dependency migrations replaced phase folders and status files with
ScopeOfWork, useful Design content and consumer-owned needs in `deliverable.yaml`.
Conditions remain distinct even when two deliverables have several relationships.
Reverse links and graph views are derived. Source-row accounting and accepted
basis comparisons in PRs #1213 and #1216 supply bounded preservation evidence;
they cannot establish that the original declarations were complete or satisfied.

The claim that every professional work product corresponds to an `ISSUED`
deliverable confused a practice concept with a workflow label. Content, purpose,
intended reliance, examination and applicable acts are the relevant distinctions.
A draft can contain consequential technical information before authentication.

### E.3.2 The Warrant Sequence

Earlier versions represented warrant development as:

```text
UNWARRANTED → CITED → REVIEWED → AUTHENTICATED
```

The notation emphasized that fluent output does not authorize reliance. Its
limitation was combining unlike relations and suggesting monotonic improvement.
A claim can have grounds without a citation; a cited claim can be poorly supported.
Review can weaken a claim or disclose missing premises. Authentication identifies
an accountable act and accepted use rather than a higher truth value.

The main argument separates grounds, examination and reliance, each with its
scope and revision. A later change leaves an act attached to its original subject
and raises an applicability question. It does not erase every prior decision or
convert uncertainty into fact.

### E.3.3 Documents, Labels and Invariant Catalogs

Datasheet, Specification, Guidance and Procedure were assigned to the four pillars
in earlier document kits. These were implementations of the questions rather than
necessary forms. A current ScopeOfWork or Design document can address several
questions and use derived views for others.

Dependencies.csv required `EvidenceFile`, `SourceRef` and related provenance
fields. FACT, ASSUMPTION, PROPOSAL and TBD conventions helped identify assertion
types. Source-field validation detected field presence rather than faithfulness,
relevance or sufficiency. Neither an instruction against invention nor a label
could guarantee that every unsupported claim was found.

The R-, I- and K-series catalogs combined commitments, representations,
procedures, tool properties and accommodations for then-current hosts. Their
historical term *invariant* did not give each rule mathematical semantics or an
enforcement mechanism. [Appendix A](appendix_a_invariant_catalog.md) preserves
the detailed crosswalk. Adopted product contracts retain their own standing.

The former claim that epistemology was the load-bearing pillar, with all other
pillars serving it, overstated the proposed thought experiment. Changing a schema
or removing a gate changes a mechanism rather than eliminating ontology or
praxeology. All four questions remain mutually corrective.

## E.4 Roles, Context and Continuity

The earlier registry made HELP_HUMAN read-only and some accounts treated a mandatory
agent hierarchy as the source of human authority. [PR #1223](https://github.com/sgttomas/chirality/pull/1223)
revised the role interpretation and resolver together. Authorized write targets
remain bounded by the assignment and host, with tests for absent authorization
and escaping targets, including symlink escapes. These checks concern the
resolver; they cannot establish universal host or App containment.

Earlier App versions offered a 3×4 matrix of epistemic postures and functions.
This was a navigation design, not evidence that work requires twelve positions.
The responsibility framework permits direct coordination or useful manager
contributions; TASK remains a non-delegating bounded executor.

The sealed-context model sought explicit inputs before delegated execution.
That concern remains useful, but a closed brief cannot remove model priors or
prove that execution used only intended information. The basis of a result and
the actual enforcement of privacy or access boundaries require separate evidence.

Immutable run snapshots, `_LATEST` pointers, receipts, handoff chains and authored
progress graphs preserved repeated working-tree accounts. The reset replaced
the general duty with current sources, queries, Git recovery, selected tags and
records justified by consumers, acceptance or recovery. Some records remain
necessary: relied-upon inputs, consequential acts and uncertain operation outcomes
can require durable evidence. Neither universal copying nor universal deletion
follows from this revision.

The formulation “if it is not in a versioned file, it does not exist for purposes
of reliance” was too broad. Conversation can establish purpose and authorization;
applications and domain engines own relevant operational state. Git also cannot
capture unsaved work, external services or substantive engagement automatically.
The relevant basis must be identified and recoverable for the use in question.

## E.5 Examination and Product Learning

The former five-gate REVIEW protocol was one staffing and sequencing arrangement.
Its important examination subjects can survive independently: scope and contributed
work; design and operational conditions; risks; technical accuracy; requirements;
quality assurance; and intended reliance. Neither completed gates nor a count
of reviewers establishes adequate examination.

[PR #1214](https://github.com/sgttomas/chirality/pull/1214) revised check selection
around consequential behaviour and exercised selection and failure propagation.
Subsequent evidence reuse distinguishes an applicable prior pass from a build
cache and states when no new tests executed. The evidence-management mechanism
supplies no general exemption from examination after a relevant change.

Archive reviews identified active technical references and contracts among intended
removals. Relocation and consumer repair were necessary even when generated
product results were already committed. A green product suite alone would not
establish that the next generator invocation remained usable.

[PR #1224](https://github.com/sgttomas/chirality/pull/1224) removed App v3, Runtime
and PEC from the working tree after accounting for active consumers. Their
[recovery references](../../projects/FROZEN.md) preserve access without creating
current verification targets. App v2 and v3 included evolution and redesign;
v4 identifies reuse candidates. A new PRD and version do not prove retained
learning: tests, native-use findings, interfaces and owner decisions need examination
for their present relevance.

[PR #1240](https://github.com/sgttomas/chirality/pull/1240) records native-use
findings about role-header accumulation and input autocorrection, followed by
repair. This supports the value of actual application interaction while leaving
general usability unestablished. [PR #1244](https://github.com/sgttomas/chirality/pull/1244)
records the App's host-placement commitment for writes, supplier interaction and
human-act capture. A commitment is not evidence that all implementation paths
conform. [PR #1238](https://github.com/sgttomas/chirality/pull/1238) provides the
accepted manuals for the working model at the source basis.

## E.6 Corrections of Earlier Claims

Several broader statements required narrowing. Agents were said to “not make
decisions,” although they can select, infer, assess and take substantial authorized
action. The main text reserves *judgment* for situated, committed human engagement
and uses *reckoning* for those sophisticated agent activities.

The prediction that capability changes would never justify a K-* amendment mixed
professional allocations with contingent mechanisms. Greater capability can
change supervision, delegation and controls without itself transferring professional
standing. The former list of future tools also presumed every lifecycle duty
should eventually be automated. A current consumer and consequential benefit,
including attention and maintenance costs, provide a more bounded basis.

The regulatory argument that conduct-focused obligations necessarily made agents
“others” exceeded its premises. The thesis preserves the July 2 mapping as a
candidate practice argument with an interpretive limit. Tool and worker analogies
can guide inquiry but cannot settle a legal term or discharge the professional's
duties. Accountability rests on its normative and institutional allocation rather
than a metaphysical proof about all future machines.

A working-session comparison with large-project delivery also exceeded the
independently established evidence. Tree, dependency graph, relationship network
and attention remain candidate correspondences, not a demonstrated equivalence
with project practice. Independent comparison and another practitioner's use
would supply stronger evidence than elaboration of the analogy by its author.

## E.7 Causal Limits and Continuing Inquiry

The owner reports improved agent behaviour and productivity following the October
reset, and subsequent PRs show connected work. Removing duties while preserving
important distinctions supplies evidence against the necessity of those mechanisms
for every undertaking. It does not isolate an effect attributable to their removal.
Stronger models, accumulated product knowledge, work begun earlier, task mix,
generated material and the owner's continuing involvement remain confounders.
File counts, line counts and merged-PR totals measure activity or size rather than
comparable accepted work or a productivity multiplier.

The historical record supports learning through production, examination and repair.
Comparative reliability, professional suitability, sustained productivity and
transfer to another practitioner remain open. Chapter 8 evaluates the specific
paths rather than promoting this chronology into those broader findings.