Source Downgrading Extensions

Follow-up work to the Source Downgrading paper (Carboni 2026). This repository carries forward the source class lattice and the inscription rule — derived trust cannot exceed input trust, with inference as the upper bound — and extends them along the four directions the paper's Section Future Work sets out.

The paper's deterministic conformance fixture, embedding-noise sweep, and 139-session adversarial-reload benchmark are not redone here; they are the load-bearing prior that this work assumes. The foundation lives in the sibling repository carboni123/source_downgrading (linked in CITATION.cff).

What the rule is

Derived trust cannot exceed input trust. The act of derivation itself imposes infer as the upper bound.

Formally, for inscription of a derived record with contributors r_1, ..., r_n:

src(Inscribe(derived))  <=  min(src(r_1), ..., src(r_n),  infer)

This is the integrity-meet rule of lattice-based IFC, adapted to the source class lattice fabricated < simulation < infer < retrieved < tool < ext. Every extension in this repository preserves the meet; only the valuation the meet operates on is enriched.

The four extensions

Module	Paper §	Direction
fgm.extensions.per_record_valuation	Future Work 1	Lift trust from per-class rank to per-record valuation tau : R -> [0,1] under class-consistency intervals [tau-_c, tau+_c]. Named in the paper as the next step.
fgm.extensions.time_varying_credibility	Future Work 3	Keep source labels monotone; let credibility vary independently. Discredited ext stays ext but loses credibility; confirmed infer stays infer with higher credibility.
fgm.extensions.context_dependent_valuation	Future Work 2	Generalize the valuation to tau(r, ctx) so the active task/role/query supplies the order. Context transitions must be explicit; the meet is computed at fixed context.
fgm.extensions.label_transport	Future Work 4	Specify signed provenance envelopes and verified import policies so labels survive crossing agent or service boundaries.

The order matches the paper's stated order of work: per-record valuation first, then time-varying credibility, then context-dependent ordering, with cross-agent transport downstream of both.

Layout

src/fgm/                           Foundation: lattice + inscription core
  core.py, inscription.py, laundering.py, source_inference.py,
  correction.py, coupling.py, residual.py, self_index.py, metrics.py
src/fgm/extensions/                The four future-work extensions
  per_record_valuation.py
  context_dependent_valuation.py
  time_varying_credibility.py
  label_transport.py
src/trace_memory/                  Public memory API the rule is wired into
src/trace_probes/                  Lightweight conformance probes
tests/architecture/                Tests of the foundation (core, laundering,
                                   source inference, trace probes)
tests/trace_memory/                Tests of the public memory API
tests/extensions/                  Tests of the four extensions (scaffold)
docs/ROADMAP.md                    Plan of attack, derived from paper §Future Work
docs/trace-memory-library.md       Public API reference
examples/trace_memory/             Walkthrough demos against the public API

The paper sources, paper-tied benchmark harnesses, LLM-in-the-loop experiment runners, and committed result files were stripped on the initial reshape. They remain available in the foundation repository.

Manuscripts

Each extension has a companion manuscript under papers/:

Extension	Manuscript	Status
Per-record trust valuation	papers/Per_Record_Trust_Valuation_pre.tex	Preprint-ready
Time-varying credibility	papers/Time_Varying_Credibility_pre.tex	Preprint-ready
Context-dependent valuation	papers/Context_Dependent_Valuation_pre.tex	Preprint-ready
Label transport	papers/Label_Transport_draft.tex	Draft, final polishing pass pending

Each preprint is self-contained: definitions, threat models, propositions, and an executable conformance fixture under benchmarks/ whose generated JSON summaries live in results/architecture/.

Install

python -m pip install -e ".[dev]"

Optional heavy dependencies (only required by specific extension work):

python -m pip install -e ".[llm,embeddings,source-classifier]"

Run the tests

python -m pytest tests -m "not live"

Foundation, extension, and public-API tests all run from the same suite (315 tests, all green at publication time).

Reproduce the conformance evidence

Each extension ships a single orchestrator script that regenerates its JSON summaries and the SHA-256 manifest:

python benchmarks/per_record_valuation.py
python benchmarks/time_varying_credibility_all.py
python benchmarks/context_dependent_valuation_all.py
python benchmarks/label_transport_all.py

Results are written to results/architecture/*_summary.json, *_multiseed_summary.json, *_properties_summary.json (where applicable), *_mutation_summary.json (where applicable), and one *_artifact_manifest.json per extension recording git commit, dirty-tree flag, Python version, dependency-manifest hash, and per-file SHA-256.

Compile a manuscript

pdflatex -interaction=nonstopmode -halt-on-error -output-directory=papers papers\Label_Transport_draft.tex
pdflatex -interaction=nonstopmode -halt-on-error -output-directory=papers papers\Label_Transport_draft.tex

Two passes resolve cross-references and the internal thebibliography. The other manuscripts compile the same way; substitute the _pre.tex filename.

Status

The four-paper chain and its companion code are stable at publication. The integrity-meet rule of the foundation paper is the non-negotiable invariant; every extension preserves it and only enriches the valuation it operates on. Label_Transport_draft.tex is the one manuscript still slated for a final polishing pass before reclassification to _pre.

License

MIT, inherited from the foundation repository.