# Trismegistus Contest And SWE Adjudication Story

## Thesis

Trismegistus is the Hermes contest deployment surface for Renaissance Field
Lite's patent-pending Mirror Architecture. The point of the build is not a demo
shell. The point is to show an AI expert partner that can preserve research
memory, use sources, run benchmark discipline, repair code, operate browser
missions, draft external actions under approval, and save receipts.

## What Happened In The Sprint

The benchmark spine emerged inside an unplanned, compressed under-five-day
contest sprint. The system went from demo surface to real benchmark receipts:

- SWE-bench Verified local selected-test receipt: `495/500`.
- Title-edge local adjudication read: `496/500`, with remaining rows boxed as
  parser, output-credit, or environment-boundary cases.
- WebArena Verified hard receipt: `255/258`, with final rows parked for
  maintainer or contest response.
- Tris 100-turn architecture-on context pass: `436/436` checks, mean score
  `1.0`, zero prompt spills, zero raw receipt spills.
- Baseline versus architecture-on slice: baseline mean `0.792`, architecture-on
  Tris mean `1.0`, delta `+0.208`.

## SWE-Bench Whole Story

The important part is not only the number. The Mirror Architecture workflow
mapped the benchmark, created the recursive repair lane, generated
source-backed patches, packaged an official-format 500-row prediction file,
attempted hosted evaluation, diagnosed the evaluator boundary, verified the
uploaded file locally, and filed the upstream issue with receipts.

Sequence:

1. Build recursive repair route: inspect exact source, generate smallest patch,
   preflight, run evaluator, save receipt.
2. Scale across SWE-bench Verified selected-test rows.
3. Package public-safe local receipt at `495/500`.
4. Build title-edge adjudication pack at `496/500` local official read.
5. Build official-format 500-row hosted submission package.
6. Run the real `sb-cli` path and upload all `500` predictions.
7. Hosted report returns `0` completed runs with evaluator-container failures.
8. Re-run the exact hosted prediction file through the official local harness;
   sanity rows complete and resolve with `0` errors.
9. File public-safe maintainer issue for inspection/re-trigger:
   https://github.com/SWE-bench/sb-cli/issues/28

## Public Boundary

Clean claim:

> Trismegistus Codex-helper route reached 495/500 official local selected-test
> resolves on SWE-bench Verified, with a hosted evaluator adjudication request
> in flight.

Boundary:

> This is a local official-harness receipt plus hosted adjudication request, not
> a public leaderboard placement. Do not claim a leaderboard score until the
> hosted/maintainer route confirms it.

## Why It Matters

This is the core contest story: the Mirror Architecture was able to rapidly map
a respected benchmark lane, organize a recursive repair process, preserve the
receipts, attempt official hosted scoring, and handle the evaluator boundary in
a professional way. That is the road toward an advanced recursive AI expert
partner: measurable improvement, source-backed work, and honest gates.