proxylimen

Proxy + limen: a proxy at the threshold.

📖 Read the essay (md · PDF) · ▶ Try the demo · Appendix A · See the evidence

The internet is a proxy of a proxy: text left behind by other people's models of the world, not the world itself. proxylimen asks where a learner can still recover structure from explicit, minimal contact, and where a validated diagnostic stops telling that structure apart from a control that has no geometry at all.

The world cannot be derived from nothing, but it may not have to be learned from the shadow of the internet either. There is an intermediate regime of calibrated derivation: contact is minimal but explicit; rules are fixed before outcomes; handles are named rather than smuggled; the failure boundary is measured rather than hidden.

What this is

proxylimen is a publishable extraction from the larger Ascesis research forge: canonical claims, reproducible experiments, a self-checking gate harness, and an offline visual demo. The repository is organized to separate citable results from superseded results that the instrument itself marks non-citable.

The central artifact is not just a set of numbers. It is the split between:

• experiments/harness_valid/ — decisions produced by gate_harness.runner.run_gate and verified by gate_harness.verify_decision.
• experiments/superseded_invalid/ — earlier decisions preserved as evidence that the method found and marked the author's own errors. Their decision.json files lack _harness_provenance, so they are INVALID by the independent verifier. See README_SUPERSEDED.md.

That separation is part of the thesis: a useful research instrument should not only report successes. It should make certain author mistakes physically hard to cite.

The question, made testable

The experiments do not train an LLM, use internet data, or claim semantic grounding. They ask bounded synthetic questions:

• If observations are observer-colored, can a small amount of explicit auxiliary calibration recover a latent scalar in a toy world?
• Can calibrated observations recover a 2D product-order relation under known or evaluated dimensional structure?
• Can dimension be estimated blindly from only a kNN graph, with no coordinates, distances, auxiliary variables, or truth hints?
• Where does a validated graph-only dimension diagnostic stop distinguishing a uniform-hypercube geometry from a geometry-free random k-out control?

The allowed claim strengths are recorded in canonical/RESULTS_CANONICAL.md. The README is subordinate to that file.

Findings, at the allowed strength

• B1 auxiliary calibration is a bounded toy identifiability signal, not a general disentanglement theorem. The harnessed rerun recovers the scalar with auxiliary calibration, while no-auxiliary and broken-calibration controls fail; however information_ratio = 0.04697540404970473 is below the preregistered threshold, so construction_may_be_tautological = true.

• B2 relation recovery shows high-F1 recovery of a synthetic 2D product-order relation under auxiliary calibration and known/evaluated structure. It is not a blind discovery of order dimension: the harness records classification_success_depends_on_harness_hint = true because a 3D-control classification depended on the evaluation hint truth_axes=3.

• B2.2.1 blind dimension estimation recovers the Kleindessner-von Luxburg E_CAP table behavior from the kNN graph alone for the supported literature-table rows up to d <= 7. The d=12 rows are boundary checks, not a claim of robust high-dimensional recovery.

• B2.3 discrimination crossover maps where the already-validated E_CAP + k-spread diagnostic stops separating uniform-hypercube geometry from the preregistered random k-out control. This is a discrimination boundary, not a dimension-accuracy or recovery boundary. In the recorded sweep, the discrete crossover is d*=130 at n=1000 and d*=24 at n=5000. The inward shift is established for the k-spread diagnostic only: under fixed k=15, n=1000 shows no crossover through d=200 — see Appendix A.

• Random-control mimicry is analytically explained for the preregistered directed random k-out graph: the mechanism gives L_CAP ~= 1/(k+1), which drives the recorded E_CAP k-sweep 9.40, 11.50, 13.10 for k=10,15,20.

What is enforced vs asserted

Enforced by the harness:

• A citable decision must contain _harness_provenance written by gate_harness.runner.run_gate.
• The current gate_harness/*.py working-tree hash must match the recorded harness_version.
• Preregistration locks, leakage scans, tautology checks, and evaluation-oracle checks must be recorded as true in provenance.
• Missing provenance fails closed: the verifier reports INVALID regardless of how good the numbers look.

Asserted only with caveats:

• The synthetic auxiliary-variable result is not evidence of general real-world disentanglement.
• The B2 relation result is not blind order-dimension discovery.
• The B2.3 crossover is not a proof that more data generally hurts, and it is not a theorem confirmation. It is a measured discrimination boundary for the stated uniform-hypercube family and random k-out control.
• No result here is a substrate claim, semantic-grounding claim, LLM-safety claim, or transfer-to-language claim.

Reproduction

Verify the citable decisions:

for f in experiments/harness_valid/*/decision.json; do
  PYTHONPATH=. python3 -m gate_harness.verify_decision "$f"
done

Expected: all five print VALID.

Verify that superseded decisions remain non-citable:

find experiments/superseded_invalid -name '*decision*.json' -print0 \
  | xargs -0 env PYTHONPATH=. python3 -m gate_harness.verify_decision

Expected: each prints INVALID with the missing _harness_provenance reason.

Run the harness self-checks and reproduce the citability split (Python 3.11+; the only runtime dependency is numpy):

python3 -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"        # numpy + pytest
pytest -q                      # 18 harness self-checks
python scripts/verify_all.py   # 5 VALID, superseded INVALID; non-zero on mismatch
python scripts/check_learner_view_hygiene.py   # estimator path never reads seed_id

CI runs exactly this on every push.

Demo

Try it in the browser, or open demo/blind_dimension.html locally. It is a single offline HTML file: no build step, no CDN, no external data. The demo lets a reader generate a hidden high-dimensional cube, give the machine only the kNN graph, estimate dimension blindly, then swap the world for a random k-out control and watch k-sweep catch the artifact.

Repository map

• canonical/RESULTS_CANONICAL.md — source of truth for essay claims.
• canonical/MEMO_B_BRANCH_HARNESS.md — audit and harness construction memo.
• MANIFEST.md — extraction map from Ascesis into this repository.
• experiments/ — harness-valid and superseded-invalid experiment artifacts.
• gate_harness/README.md — mechanical gate discipline and verifier semantics.
• demo/blind_dimension.html — interactive visual explanation.

The triptych

This repository is one panel of three, sharing a single thesis — do not try to certify intentions; build contact, consequences, and constraints that can be checked — and a single honest caveat: all three come from one forge, one author, the same AI partners. Independent replication is invited and would outweigh any further internal check.

• justitia — what keeps a world of powerful, evolving agents livable when no one can read anyone's soul.
• proxylimen (this repo) — where a mind's world comes from, and where blind derivation measurably breaks.
• fallacy-cutter — the fail-closed instrument both were cut with.

License

MIT — see LICENSE.