# Perception

> Applied research company developing precision techniques to edit
> trained models — fixing behaviour without retraining. The program is
> called certified capability repair: input, failure examples; output,
> a certified patch with proof it survives installation. Sage is the
> first applied product, focused on child development.

## Research focus

- Certified capability repair — per-dimension certificates (efficacy,
  locality, regression, cost) that decide whether a model patch is safe
- Causal tracing — locating where a behaviour lives in a model
- Rank-one model editing (ROME) — targeted, reversible weight edits
- Task vectors and weight arithmetic — composing and removing learned behaviours
- Representation engineering (RepE) — steering behaviour at inference time

## Key published findings (2026-06)

- Teacher-forced evaluation overstates free-generation capability by a
  near-deterministic linear amount: gap = -0.045 + 0.931 * headroom
  (R^2 = 0.99), the same line across a 5x model-size range (1.5B, 7B).
  Inflation is largest in the fallible regime where repair happens.
- Training loss says nothing about patch safety: a LoRA patch with a
  clean loss curve was refused by certification for damaging
  neighboring capabilities (spurious-JOIN injection).
- Certificate thresholds must be calibrated from a measured
  benign-perturbation noise floor, under a rule fixed before data; this
  makes the certificate both necessary and satisfiable.
- Low-rank patches reach the edge of certifiability but land on it: a
  gentle patch missed a clean pass by a floating-point epsilon. No
  certified patch is claimed in the study.
- Naive composition (summing patch weight deltas) dropped both patched
  capabilities below the unpatched base model.
- Merge conflict is predictable from patch weights; the governing
  variable is norm dominance, not subspace overlap (lower overlap is
  more destructive). "Same capability is safe" is false.
- The optimal install policy is to ROUTE between patches, not merge
  them: routing beat naive merge on 8 of 9 pairs and naive merge is
  strictly dominated.
- The framework transfers to a second, unrelated domain: tool/function
  -call generation on the Berkeley Function-Calling Leaderboard (BFCL).
  The certificate refused a patch that was useless on its target and
  also dropped instruction-following by 0.094 — the same collateral
  victim degraded by both SQL and tool-call training.
Working paper: https://perception.club/research/certified-capability-repair
Concluded study (note): https://perception.club/notes/you-cannot-trust-a-patch-without-a-certificate
First results (note): https://perception.club/notes/certified-capability-repair-first-results

## Other papers

- When Does a Language Model Help a Graph Neural Network on Relational
  Data? Under a fair paired, multi-seed protocol, graph-conditioned LLMs
  (soft-prompt "Rel-LLM" style) tie but do not beat a well-tuned GNN on
  RelBench relational prediction, at 1.5B and 8B scale. A structured GNN
  encoding beats flat-text RAG on both accuracy and ~50x token
  efficiency. Proposes a portable, model-agnostic GNN-for-RAG
  architecture: the GNN is the predictor, the LLM is the interface.
  Paper: https://perception.club/research/when-does-an-llm-help-a-gnn

## Sage

Sage is the first applied vertical: longitudinal evidence of child
development that tells precision editing what to aim for.
Marketing: https://perception.club/sage/ · App: https://sage.perception.club

## Pages

- https://perception.club/ — homepage
- https://perception.club/research — long-form research thesis
- https://perception.club/research/certified-capability-repair — working paper (full text)
- https://perception.club/research/when-does-an-llm-help-a-gnn — working paper (full text)
- https://perception.club/notes — research notes index
- https://perception.club/sage/ — Sage marketing pages
- https://perception.club/contact — get in touch

Contact: contact@perception.club