Tilelli-llm / results /AUDIT_TRAIL.md
TilelliLab's picture
Mirror small files (code, paper, results)
f86dc09 verified
|
Raw
History Blame Contribute Delete
3.3 kB

Audit trail — from the original kit to this one

This file documents what was removed, changed, or rewritten when the project went from tilelli-kit-2026-05-23-audited.zip (the original shipping kit, SHA 00af96d8…) to this humble public repo.

Headline

The original kit shipped 6 forward-looking claims ("emergent metacognition", "3/3 seeds beat vanilla", "matched at parameter count", etc.). This kit ships 3 verifiable positives + 3 verified negatives, and only the v4 deployed ckpt.

Removed

Artifact Why
BENCHMARKS.md (original) Headline overstated: "3/3 seeds beat vanilla" rested on heterogeneous best_val definition + asymmetric eval_every. Replaced by results/claim_01_benchmark.md (preliminary 1-seed directional).
HOW_TILELLI_WORKS.html 17-scene marketing-style explainer. Replaced by this README + the audit trail.
PAPER.md §4.2 "emergent metacognition" Empirically disproven, see results/claim_02_metacog.md.
Ckpts: tilelli_chat_v5_metacog_cpu.pt, tilelli_chat_v6_metacog_chat_step100.pt, tilelli_chat_v7_metacog_v2.pt, tilelli_chat_v4splice_v7abstain.pt All failed their respective deploy gates. Available on request for negative-result replication.
runs/ (28 of 30 sub-dirs) Historical engineering churn. Only the 3 audit-relevant ones (metacog_2026-05-23, metacog_v7_2026-05-23, metacog_v4v7splice_2026-05-23) are referenced by claims here.
swap_weights.sh Deprecated per 2026-05-20 memory; could cause prod regression.

Rewritten

File Change
README.md Karpathy-style honest framing. Every claim links to a script. Negative results first-class.
results/claim_01_benchmark.md Inherits the audited BENCHMARKS.md (21:39 2026-05-23) honest version.
results/claim_02_metacog.md New, documents the DISPROVEN router-entropy claim + the v7 sub-result + the splice failure.
pyproject.toml Trimmed to torch + numpy only.

Preserved as-is

File Notes
src/tilelli/ (the package) The actual model code. Unchanged from the kit version.
checkpoints/tilelli_chat_v4.pt The deployed ckpt. SHA-pinned in results/claim_01_benchmark.md.
prompts/probe_210.jsonl The audited eval set.
LICENSE Apache 2.0.

What replication looks like

If you want to reproduce the negative results (not just the positives), the original kit + the metacog run dirs (v5 / v6 / v7 / v8a / v8b ckpts and the full probe JSONLs) are preserved in the lineage's working repo — available on request via hello@tilelli.tech.

What "humble" means here

Three rules:

  1. The README is the contract. Every numerical claim has a script that produces it from a bundled artifact, and the script exits non-zero if it doesn't.
  2. Negative results are first-class. A finding of "this doesn't work, and here's exactly how we know" is as worth shipping as a positive result of equal effort.
  3. One ckpt, one story. Don't ship five ckpts and let the user guess which one corresponds to the headline. Ship the one that anchors the claims; document the others as audit trail.

This is the Karpathy-style framing of nanoGPT and minBPE adapted to a small routed LM with an audited claim trail.