Skip to content

METHOD.md

Latest commit

 

History

History
156 lines (102 loc) · 5.18 KB

METHOD.md

File metadata and controls

156 lines (102 loc) · 5.18 KB

Method Contract

This repository publishes the public contract for a constrained auditable capability-upgrade workflow. It does not publish the private frontier search or full training stack.

Object Of Study

The public method object is not a single block, not a generic PEFT primitive, and not only a board artifact.

The accepted object is:

  • a frozen authoritative parent
  • a set of residual-family local repairs
  • a route-gated trunk recurrent surface
  • a promotion contract that all accepted repairs must survive
  • a derived bounded artifact for the board proof

That coupled object is the unit of evidence in this repo.

See:

Frozen Parent

All public comparisons are anchored to one frozen authoritative parent.

Why that matters:

  • gain is always measured relative to the frozen parent
  • locality probes compare candidate behavior against the frozen parent
  • the exact no-trunk ablation is interpreted as a return to the frozen parent boundary
  • additional release-boundary audits are also interpreted against the frozen parent boundary

The parent row is published in:

Residual Family

In public terms, a residual family is a bounded error slice defined by the combination of:

  • the parent competition shape
  • the local repair target pair
  • the route surface that decides where the repair may fire

This repo publishes the accepted family-local repair evidence and representative locality probes. It does not publish the full discovery frontier, rejected candidates, or support-source graph.

Route Activation

Public route semantics are documented in:

At a high level:

  • question and option token surfaces are converted into a route signature
  • a block is eligible only if its route requirements are satisfied
  • parent-topology conditions must also match before the recurrent repair can fire

The current public inventory shows:

  • 75 trunk recurrent blocks
  • median route token count 6
  • route minimum hits of 2 or 3
  • a published-eval route-disabled proxy collapses official LogiQA to 0.244548

See:

Trunk Recurrent Blocks

The current public surface contains:

  • 72 blocks in option_latent_v2
  • 3 blocks in pair_score
  • 11,325 exemplars
  • 8192 GPU hash dim

The repo publishes block-level structure and representative sanitized fields, but not full route tokens, source ids, or latent prototypes.

Public contrastive controls now also show:

  • target-only diagnostic blocks reach only 0.308411
  • a same-parent trained linear-readout control reaches 0.300623
  • a same-parent trained BitFit option-bias control reaches 0.266355
  • a same-parent trained LoRA-style hash-delta control reaches 0.303738
  • a same-parent trained low-rank adapter baseline reaches 0.303738
  • a same-parent trainable-budget-matched LoRA-style control reaches 0.303738
  • a same-parent trainable-budget-matched low-rank adapter control reaches 0.303738
  • architecture-near low-rank adapter controls stay at the frozen-parent level

See:

Promotion Contract

A local repair counts only if it survives a coupled acceptance predicate. Publicly visible parts of that predicate include:

  • official replay against the frozen parent
  • external non-regression
  • leakage and deep-pretrain audit status
  • runtime and shadow cleanliness
  • overfit and release-boundary disclosure

The public repo also documents a stricter internal mainline:

  • the formal mainline monitors multiple benchmark routes rather than one benchmark only
  • in the current public line, that includes LogiQA, GSM8K, and a narrow instruction-following benchmark
  • external dev and external blind are treated as required gates, not optional side receipts

See:

Crystallization Boundary

The board proof is a derived artifact, not the whole method claim.

Current public board scope:

  • aggregate board-proof summary mode: evaluation_mode = logiqa_batch_compiled_probe_aggregated
  • default single-batch board readback mode: evaluation_mode = logiqa_batch_compiled_probe
  • compiled_probe_mode = host_full_exact
  • fixed published batches only

So the board proof demonstrates bounded crystallization correctness, not unrestricted open-input reasoning on MCU.

See:

Published Versus Withheld

Published here:

  • accepted mechanism contract
  • replayable public evidence
  • representative locality probes
  • exact no-trunk ablation
  • board acceptance and raw readback

Withheld here:

  • full family-mining pipeline
  • rejected frontier
  • support-source graph
  • curriculum construction rules
  • full checkpoint internals needed to reproduce the private training moat