# AMF v0.18 — Public Evidence Projections

This directory publishes reproducible, non-authoritative projections of
selected scientific experiments recorded by the frozen AMF v0.18 campaign.

The projections expose observable evidence, lineage, adjudication and
reproducibility without exposing the Foundry's private experiment-selection
machinery.

## Projection contract

All three artifacts use:

`amf/public-evidence-projection/v0.1`

The publication boundary is defined in:

`AMF_Public_Evidence_Projection_Contract_v0.1.md`

The canonical Scientific Event Ledger, registry, receipts, identities,
measurements and dataset authorities remain authoritative over these
projections.

## Validated projections

| Projection | Intervention locus | Recorded verdict | Research state |
|---|---|---|---|
| `v018-exp-001` | Representation canonicalization | `SUPPORTED` | research-confirmed |
| `v018-exp-002` | LoRA continuation / neural mutation | `SUPPORTED` | not research-confirmed; retained in learning frontier |
| `v018-exp-003` | Raw substrate switch | `REFUTED` | SEARCH-only; not search-eligible |

### v018-exp-001

A representation intervention created a distinct prediction-system
realization while preserving the exact neural adapter and introducing
`CANONICALIZE_TO_RECORDS_V1`.

Projection SHA-256:

`5e80ccda18956e7ec588f0357cb9d7c7eda636c669661cfe2392968aaed814d0`

### v018-exp-002

A `LORA_CONTINUE` intervention preserved the representation transform while
performing recorded gradient-backed neural mutation over 140 optimizer steps.

The experiment was adjudicated `SUPPORTED`; the resulting child was retained
in the learning frontier but was not research-confirmed.

Projection SHA-256:

`bafd254d56a69b5d1b92430b9a1505019e840b2eb33a4f0d9197d1c4e7a4e138`

### v018-exp-003

A `SUBSTRATE_SWITCH` instantiated a raw Phi-4-mini realization and compared
it with the recorded `RAW_MISTRAL` contemporaneous control.

The child did not become SEARCH-eligible. No SEARCH_CONFIRM measurement was
recorded for this child, and the projection does not synthesize one. Both the
experiment and linked hypothesis were adjudicated `REFUTED`.

No physical parent is fabricated because none is recorded in the canonical
lineage of this realization.

Projection SHA-256:

`78f5f6b0de028063e55823c8749b9adceae0139395392b4e984e5bc36c623901`

## Validation

Each artifact passed four validation categories against the frozen canonical
AMF history:

1. deterministic rebuild,
2. independent trace verification,
3. targeted tamper rejection with fail-closed behavior,
4. deletion followed by byte-identical regeneration.

These results establish reproducibility, traceability and fail-closed behavior
for these specific v0.1 projection artifacts. They do not claim that every
future AMF intervention locus is already represented by a generic public
schema.

## Disclosure boundary

Published here are bounded scientific claims, source references, observable
lineage, evidence bindings, adjudication state, derived integrity checks and
reproducibility identities.

Planner scores, candidate rankings, unselected proposals, private search
heuristics, protected rows, Fresh Final rows and seeds, and proprietary
Foundry decision machinery are not published here.

**The public evidence is inspectable. Protected evidence and proprietary
discovery machinery remain controlled.**

## Relationship to the qualified realization

These projections describe selected research experiments from the campaign
that produced AMF v0.18.

They do not alter the terminal qualification verdict and do not redefine the
qualified realization published at the repository root.

**The model is a substrate. The qualified realization is the product.**