Parallax

Render-in-the-loop self-verification for coding agents — and the study of why it works.

A second viewpoint reveals structure invisible from the first. Parallax flips a generated text artifact (SQL, a regex, plotting code) into a rendered representation, has a critic inspect the render, and reports (or corrects) silent, "looks-right-but-wrong" bugs a quick glance misses.

• Package (PyPI): parallax-verify · import: parallax · API: crosscheck(artifact, render_fn)
• Status: 🟢 M1 usable. The crosscheck primitive works end-to-end with two adapters (SQL-result-shape, regex-highlight), a CLI, and an MCP server. Offline by default (no API key); upgrades to a multimodal vision critic when one is configured. The why-it-works study (M2) is separate and ongoing.

---

What it catches

Code that LLMs (and people) generate can fail silently — it runs, returns a plausible answer, and is wrong. Nothing throws.

• SQL: a dropped/loose JOIN predicate inflates row counts (fan-out); a LEFT/INNER swap silently drops rows.
• Regex: a greedy quantifier over-matches into a region it shouldn't; a pattern quietly misses spans it should catch.

Parallax extracts a small set of neutral shape facts (row-count ratio, null fractions, how many matches land in a forbidden region, …), renders them, and has a critic flag anything that looks off. It knows only what you tell it — there is no hidden oracle.

Install

pip install parallax-verify               # core: SQL + regex + CLI, offline heuristic critic
pip install "parallax-verify[anthropic]"  # + Anthropic vision critic
pip install "parallax-verify[openai]"     # + any OpenAI-compatible vision endpoint
pip install "parallax-verify[mcp]"        # + the MCP server

Walkthrough (CLI)

Offline by default — no API key required. Here's a real session; the whole idea fits in about thirty seconds.

A greedy regex over-matches into a region it was supposed to leave alone. --forbidden-spans 3:9 says the characters SECRET must not be matched:

$ parallax regex "<.*>" --target "<a>SECRET<b>" --forbidden-spans 3:9
parallax regex: "artifact" - [!] SUSPICIOUS (suspicion 0.65)
  critic: parallax-local-heuristic
  flagged: matches_in_forbidden_region
  note: local heuristic flagged extreme bins: matches_in_forbidden_region
$ echo $?
1

Make the quantifier lazy and the bug is gone — Parallax goes quiet:

$ parallax regex "<.*?>" --target "<a>SECRET<b>" --forbidden-spans 3:9
parallax regex: "artifact" - [ok] looks fine (suspicion 0.05)
  critic: parallax-local-heuristic
  note: local heuristic saw no extreme bins
$ echo $?
0

That flip — <.*> flagged, <.*?> clean — is the whole product in miniature.

SQL works the same way. A join predicate that's too loose multiplies rows (fan-out); you tell Parallax how many rows you expected and it catches the blow-up (3 expected, 12 returned):

$ parallax sql "SELECT o.id, c.region FROM orders o JOIN cust c ON o.cust = c.id" \
    --setup "CREATE TABLE orders(id INT, cust INT); INSERT INTO orders VALUES (1,10),(2,10),(3,10);
             CREATE TABLE cust(id INT, region VARCHAR); INSERT INTO cust VALUES (10,'A'),(10,'B'),(10,'C'),(10,'D');" \
    --expected-rows 3
parallax sql: "artifact" - [!] SUSPICIOUS (suspicion 0.65)
  critic: parallax-local-heuristic
  flagged: rowcount_ratio_to_expected
  note: local heuristic flagged extreme bins: rowcount_ratio_to_expected

Exit code is 0 when it looks fine, 1 when flagged, 2 on error — so it drops straight into CI. Add --json for machine-readable output:

$ parallax regex "<.*>" --target "<a>SECRET<b>" --forbidden-spans 3:9 --json
{"kind": "regex", "item_id": "artifact", "caught": true, "suspicion": 0.65, "flagged": ["matches_in_forbidden_region"], "rationale": "local heuristic flagged extreme bins: matches_in_forbidden_region", "corrected": false, "proposed_fix": null, "critic": "parallax-local-heuristic"}

Swap the offline critic for a vision model with --critic anthropic (needs ANTHROPIC_API_KEY) or --critic openai — see Critics below.

Use from Python

from parallax import Artifact, crosscheck, render_regex
from parallax.clients.heuristic_critic import HeuristicCritic

art = Artifact(kind="regex_highlight", text="<.*>",
               context={"target": "<a>SECRET<b>", "forbidden_spans": "3:9"})
result = crosscheck(art, render_regex, HeuristicCritic(), mode="expected")
print(result.critique.caught, result.critique.flagged)  # True ('matches_in_forbidden_region',)

Run the full demo (SQL + regex, offline): python examples/demo.py.

Use as an MCP tool (for Claude / agents)

parallax-mcp starts a server exposing one crosscheck tool (kind = sql | regex, text, context, optional mode/critic). Point your MCP client at it and an agent can render-verify its own generated SQL/regex before trusting the result.

The two honest modes

Parallax has no magic oracle. It works from what you give it:

1. Supplied expectations — you provide the expected row count, the spans you meant to match, or the regions that must not match.
2. Relative anomaly — compare against a previous / reference run.

Critics

Bring your own key — local needs none, and the vision critics accept a key from an env var or the --api-key flag:

• local (default) — a fast, offline, rule-based sanity check over the same verdict-free features. No API key.
• anthropic — an Anthropic multimodal model looks at the rendered image. PARALLAX_CRITIC=anthropic + ANTHROPIC_API_KEY (default model claude-opus-4-8).
• openai — any OpenAI-compatible vision endpoint. PARALLAX_CRITIC=openai + OPENAI_API_KEY (default model gpt-4o-mini). Point --base-url (or PARALLAX_BASE_URL) at OpenAI, Google Gemini's OpenAI-compatible endpoint, OpenRouter, or a local server, and pick the model with --model.

Which critic to trust. The offline local critic is the zero-config default and is fully tested. The Anthropic vision critic is validated live on the SQL adapter — it passes clean results and catches fan-outs. (An earlier render showed only a color per bin and the model false-positived on clean results; the render now prints each bin's label as text, which fixed it.) The OpenAI-compatible endpoint and the regex render haven't had the same live check yet, so treat their vision verdicts as provisional and lean on local for CI.

# use your own OpenAI key
parallax sql "..." --setup "..." --expected-rows 3 --critic openai --api-key "$OPENAI_API_KEY"
# or a different provider via its OpenAI-compatible endpoint
parallax regex "<.*>" --target "<a>x<b>" --forbidden-spans 3:4 \
  --critic openai --base-url https://openrouter.ai/api/v1 --model "google/gemini-2.5-flash"

All settings also read from the environment (PARALLAX_CRITIC / PARALLAX_MODEL / PARALLAX_API_KEY / PARALLAX_BASE_URL) — see .env.example.

What this is (and is not)

The render-and-critique loop is not new — it exists in several verticals (maps, motion trajectories, plotting, GUI agents). Parallax's contribution is different: a general, artifact-agnostic primitive with SQL-result-shape and regex-highlight adapters, plus a controlled study the field skips.

We do not claim "rendering beats text," that images universally help, or that the idea is ours. The honest claim is a model-conditional representation effect, scoped to the cells we actually run, with nulls reported. The factorial decomposition that would justify any why-it-works statement is a separate, ongoing research program (M2) — see docs/NOVELTY.md for the full do-not-claim list and docs/RED-TEAM.md for why the design is shaped this way.

Documentation

Doc	What
docs/PRD.md	Product requirements, users, scope/non-goals
docs/INTERFACE.md	crosscheck() + the adapter contract + worked adapter specs
docs/ARCHITECTURE.md	Stack, package layout, sandboxing, model routing
docs/TASKS.md	TDD-first task list (M1 library MVP + M2 study), risk register
docs/EVAL.md	The M2 crux — pre-registration-grade factorial decomposition design
docs/NOVELTY.md	Prior-art verdict, defensible claim, do-not-claim list
docs/RED-TEAM.md	The adversarial findings that reshaped the contribution
docs/CITATIONS.md	Citation-integrity pass — all 13 sources verified real; 7 wording/number corrections applied