SYS.02 · PARALLAX · TRUTH-FINDING IN CAPTURED DOMAINS · STATUS // IN BUILD · POINTED AT BELIEF

When you can't trust either side.

Some questions have no clean source. Parallax fixes a claim's true position the way a surveyor does: two sightlines that don't share an interest.

REV // 386d109 LAST OBSERVED // 2026.07.10 OUTPUT // A DATA POINT AGAINST // THE CAPTURED BOTH-SIDES

The person this serves is stuck between two confident stories, and the stakes are theirs. Picking a side is the mistake, and doing nothing is also a decision.

RETURNS //a position, fixed within limits · the unresolved parts named
WILL SAY //"we don't know yet" · the answer nobody selling a side will give you
WON'T DO //persuade · crown a winner · it sharpens the operator's judgment and hands its output onward

The fix

SCROLL · TWO SIGHTLINES

A contested claim. The mainstream has something to protect. So does the counter-crowd.

One sightline gives you a direction, not a position. The consensus is one sightline — and it decides how far along it the truth sits by how confident it feels.

So take a second sighting: a body of evidence whose incentives cut against the first. Each side reads only its own sources — never the model's recall.

Where independent sightlines cross, the position fixes. The distance from where consensus put it is the reality delta.

What lands with you is a data point — calibrated, bounded, honest about what's still unresolved.

01THE WALKTHROUGH · ONE CLAIM, THROUGH THE MACHINE

Memory isn't evidence: the recall was trained on the very fight it's judging. You hold the first two gates. Follow one claim down.

STAGE 01 // THE CLAIM ENTERS

‹ THE CLAIM › contested · stakes real · no clean source

One line, worth the machine. representative — no real run shown

STAGE 02 // THE SPLIT ◂ GATE P0 · YOURS

‹ C.1 › documented fact · load-bearing ‹ C.2 › contested reading · load-bearing ‹ C.3 › pattern inference

Hairlines fork it into testable pieces — you validate the tree before anything runs. If a load-bearing piece fails, the claim fails.

STAGE 03 // THREE CORPORA · THE WALLS GO UP ◂ GATE P1 · THE SOURCE MAP IS YOURS

FORsupporting corpus · sightline 01

reads only its own sources

▸ strongest crux · survives

▸ two backers, one funder → counted as one lineage

COMMON GROUNDthe baseline both sightings share

what both sides must accept

▸ the undisputed facts

▸ the frame both share

AGAINSTchallenging corpus · sightline 02

reads only its own sources

▸ a load-bearing piece · fails

▸ an absence that should be there

▸ · source withheld

Full documents, assembled on purpose — who's captured, who's independent, co-funding counted as one lineage. The sub-claims drop inside; the walls are real. No corpus, no run.

STAGE 04 // THE COUNCILS · SEVEN METHODS, RUN INSIDE EACH WALL

The councils — walled off Forensic AccountingIntelligence Analysis EBM Critical AppraisalInvestigative Journalism Causal InferenceSource Authenticity Statistical Forensics

:: hover a council — what it checks

Each bench works one chamber at a time, from inside its walls — cite, or stay silent — and presses hardest where the load bears; that's where the answer lives. The findings etched in the chambers are theirs.

STAGE 05 // WHAT CROSSES THE WALL

Findings pass through the wall; the reasoning never does — a confident chamber can't argue the others around. And every claim that survives is attacked again, with sources chosen to break it.

STAGE 06 // THE ISOLATED ARBITER

THE ARBITER

a chamber apart — findings in, nothing else

fresh contextfixed prioritythe skeptic's veto

It never meets the debaters — only their findings converge here, into a context that can't backfill. Conflicts resolve by fixed priority, never by splitting the difference; a separate skeptic can still strike the whole result down.

STAGE 07 // OUT THE BOTTOM · TRANSFORMED

◆ ‹ THE CLAIM › → THE DATA POINT

Calibrated confidence, per sub-claim — plus what evidence would change it

rejectedweakenedunresolved provisionally-supportedpromoted

C.1 · C.2 · C.3 — EACH LANDS WITH A STATUS

The runs so far are tuning passes, not trophies. When it earns a case file, one will appear here.

01·BWHAT LANDS WITH YOU

One card per claim. Representative: no real run shown.

DATA POINT // ‹ THE CLAIM ›REPRESENTATIVE

C.1documented fact · load-bearing0.86PROMOTED

C.2contested reading · load-bearing0.55PROVISIONAL

C.3pattern inference0.30UNRESOLVED

WOULD CHANGE IT //a primary document from the counter-corpus · one independent lineage confirming C.2

A DATA POINT, NEVER A VERDICT · IT FEEDS YOUR JUDGMENT, THEN STOPS

02WHAT KEEPS IT HONEST

Closed-book, enforced

Training data is consensus residue. The corpus rule is enforced, not suggested — the machine reads the sources, never its recall.

One funder, one vote

Sources that share funding, mandate, or ancestry count as one confirmation, however many mouths repeat it.

Agreement across enemies

Two sources agreeing means little if they want the same thing. Corroboration is weighted by how sharply the interests cut against each other.

Three rules, doing most of the work — each cuts against the way a model wants to answer.

IT DECLARES ITS OWN LIMITS · EVERY RUN which of the five known limits bit what the honest confidence is which sub-claims stayed unresolved

WHY 'IN BUILD' //live use is gated on source-assembly depth · a corpus worthy of the method
THE LEDGER //confidence numbers earn trust as real runs get logged and scored · it is being kept

THE SAME TEST // RUN 2 OF 3

03AGAINST THE AVERAGE MACHINE

A GENERIC MODEL, ASKED A CONTESTED QUESTION

Answers from memory — the averaged residue of the contest itself.

Confident either way. Can't show its sources.

Hands down a verdict.

PARALLAX ///

Cut off from memory. Reads only assembled, opposed evidence.

Confidence is calibrated, logged, and scored over time.

Hands you a data point, and what would change it.

A data point, never a verdict. And an honest "unresolved" when that's the truth of it.

NEXT FILESYS.03 · SIGNAL — POINTED AT COMMUNICATION▸