AI Coding Test 2026 · Reliability Platform

Does the system know when it’s right, wrong, or unsure?

One evidence-first platform, proven on two hard tasks. Every result is pass, fail, or unknown — lack of evidence can never become success. All numbers below come from real runs, not self-report.

Task 2 · SEC 10-K item extraction

Item status across the sweep

Lower pass-rate here is higher honesty: stubs that point elsewhere are labelled, not faked.

Confidence discriminates

Before the audit, confidence was a flat ~0.958 for everything. Now:

Independent oracle · Item 8 vs SEC XBRL facts

The answer to “how do you ensure status is trustworthy?” — not an LLM judging itself, but SEC’s own structured numbers. A real Item 8 must contain the reported revenue / net income / assets; a wrapper stub cannot.

Corner case · cross-reference-index filings

Task 1 · Browser agent — selector self-repair

Same search task on two site versions. v2 drifts the UI; the agent repairs itself via the accessibility tree and the verifier confirms real results.

Task 1 · Eval hardening — calibrate the judge, perturb the world

The verifier itself is measured against labelled corruptions, impossible tasks probe silent failure, and mutation sites bend the site along three axes.

Verifier calibration · labelled triples

Ground truth by construction: known-good runs vs four corruption classes. Positive = verdict pass; unknown never counts as success.

Silent failure & stability

Impossible-task probes, pass@k repeats, trajectory hygiene and the opt-in false-success triage — all measured, none self-reported.

Degradation curve · 3 axes × 3 intensities

Teal = final success, grey = mid-task checkpoint (query reached the real field). A red stub marks success 0. Checkpoint ≠ success localises where the failure happened.

Task 2 · Eval hardening — three engines, official oracles

A third extraction engine triangulates every item, SEC’s own CYD iXBRL block-tags anchor Item 1C, char-offset F1 locks span boundaries, and landmine cases are executable regressions.

CYD iXBRL oracle · Item 1C span vs SEC’s own tags

Coverage = share of the officially tagged span our segment covers. The two disagreements are wrapper 10-Ks whose 1C is an honest reference stub.

Format-source coverage · era × filing agent

Where the pipeline is proven vs honestly unsupported — pre-2001 plain-text SGML fails cleanly (0 fabricated items).

Adversarial audit · how the silent failures were found