Course / Lesson 1 of 6
15 min · sanitized real case
The Truth Contract
Objective: Classify every derived statement into exactly one of eight truth classes — and refuse promotions driven by repetition, popularity, or author confidence.
First, in plain language
In a large corpus, most sentences are someone claiming something — not something proven. The Truth Contract forces every derived statement to carry an honest label: what I observed with my own bytes, what the source merely declares, what two independent origins confirm, and what collides.
The golden rule: no repetition, popularity, or author confidence ever promotes a claim. A sentence said a thousand times by the same origin is still that origin's declaration.
Open the technical layer
Every derived claim uses exactly one of these eight classes and keeps a short receipt, textual locator or media id, confidence, and corroboration/conflict relations:
| Class | Operational meaning |
|---|---|
OBSERVED_LOCAL | Observed directly in local bytes/objects, no interpretation. |
SOURCE_CLAIM | The source asserts it; we did not verify. Repetition does not promote. |
CORROBORATED | Two or more independent origins converge. |
INFERENCE | Derived by a declared rule; the label is never hidden. |
PRODUCT_HYPOTHESIS | A product idea; stays a hypothesis until real proof. |
VALIDATED | Proven on the real path by an independent verifier. |
CONFLICT | Evidences collide; the claim points to all of them. |
UNKNOWN | No evidence in either direction — declared, not disguised. |
In the real mission, of the 1,087 final claims: 990 SOURCE_CLAIM · 66 CORROBORATED · 16 PRODUCT_HYPOTHESIS · 8 INFERENCE · 5 OBSERVED_LOCAL · 2 CONFLICT. The validation queue with ~998 declarations stays open — and that is fine.
The core case: the headline vs. its own chart
A public seed article carried the headline "more than 1/3 of PRs are already agentically approved". The headline is a legitimate SOURCE_CLAIM — until you look at the article's own chart: the visual series shows a brief peak of ~34–35% and an endpoint of ~28%.
The headline is not casually "true" or "false": it collides with the same source's visual evidence. The claim was recorded as CONFLICT (claim_3f369af58cbd) pointing at both evidences — A (headline text) and B (chart series). Resolving the collision would require primary data; declaring the conflict is the available truth.
Promote or not?
For each case, decide: promote the claim's class, keep the class, or mark CONFLICT.
Case A · "supports 40 integrations"
A repository's README claims support for 40 integrations. The sentence appears on three pages of the same README.
Case B · the same count, twice
Two independent builders derive the same item count from the same bytes, through different code paths.
Case C · "more than 1/3 of PRs"
The headline says ">1/3 of PRs agentically approved"; the article's own chart shows a brief ~34–35% peak and a ~28% endpoint.
Case D · "this would make a great product"
A recurring pattern in the corpus inspires a product idea, written down with conviction by the analyst.
Case E · the pilot that ran twice
A reconciliation pilot ran twice over ~19 GB of real data, byte-identical, with an independent verifier recomputing every receipt.
Case F · semantic recall
Question: did the processing cover all the corpus's meaning? There is no measurement in either direction.