Evidence Ledger

The ledger makes "every number has a source" checkable. Every factual claim in a lifecycle artifact carries a source tag. Gates check resolving pointers and source counts; the model still judges whether a source supports the claim.

Mode. A standalone answer cites its direct sources in the requested format and destination; the evidence files, gates and state writes below belong to a lifecycle initiative. Both per operating modes.

Tag Grammar

[class:identifier], at the end of the sentence it substantiates, before the period or after it, consistently within a document.

Every Tag Has a File

A pointer to nothing is an invented source. Every interview, doc and data tag has a file in evidence/ holding the raw material it cites: notes, the pasted query output with its parameters, the excerpt. Names, format and lookup order: the state schema, section Evidence Files. Gate condition E.1 checks it on every gate.

Write the evidence file when the source is captured: notes written from memory a week later are a doc about an interview, not the interview. What the user tells you in the conversation is [doc:user-{YYYY-MM-DD}-{topic}], with their words copied verbatim into the file.

Evidence From Outside the Session

Most evidence lives somewhere the session cannot reach: interview notes in a notes app, a call transcript, a CSV export, a support inbox, a survey tool. It counts the moment its raw material is in evidence/. Ask the user to paste it or give the file path, then save it with scripts/bos.mjs evidence add --class interview|doc|data --id {id} --where "{where it came from}" [--from {file}] (the raw material from the file or pasted on stdin; --shared for evidence behind PRODUCT.md, DOMAIN.md or TECH.md), which writes the format above and prints the tag to cite. One participant per interview file. A link alone is not evidence: save the excerpt the claim rests on, with the link as Where. Where commands cannot run, write the same file by hand. Resolve a connected source first when one answers the question (tickets.read, meetings.read, docs.search); importing is the floor, not a second-class path.

Source Hierarchy

Choose the source that best answers this claim by directness, relevant population, observation date, provenance and known limitations; the source class alone leaves a conflict open. Estimates and assumptions stay explicitly uncertain. Read every relevant conflicting source. Write the conflict into the artifact, recording both and explaining the choice or leaving it unresolved. Examples: examples.md, Source Conflicts.

Counting Rules

1. An evidence unit is one tag on one distinct claim. One tag repeated across five sentences is one unit.
2. Primary sources are data, interview, code. doc counts as primary only when the document is itself a record of primary contact (a transcript, a support ticket), not a summary of one.
3. Evidence thresholds count data, interview, code and primary doc units. estimate and assumption belong in artifacts, marked as what they are.
4. Distinct sources. Five quotes from one interview are one source. So are lines of one code file (regardless of line number or symlink alias) and multiple exports of one query or document; evidence files may declare **Source identity:** {stable upstream source} to identify this equivalence. A new filename does not create independent evidence.
5. Every number gets a tag. Percentages, counts, currency, dates of events. Adjectives of scale ("most", "many", "rapidly") count as numbers and need a tag or a rewrite.

Rewrite Protocol

An untagged claim is kept and made honest: rewrite it into the weakest honest form and mark it: narrow the claim to what the source supports, separate the observation from the causal story, and name the test that would settle it. All three, every time. Worked example: examples.md, Rewrite Protocol.

Audit Mode

A gate check or an artifact audit runs this pass:

1. Extract every sentence containing a number, a proportion word, or a causal claim.
2. Flag each one without a tag.
3. Count units per class.
4. Report: total claims, tagged, untagged, primary units, distinct sources, assumption ratio.
5. Fail the gate if the phase threshold is not met, naming the specific untagged sentences.

A high assumption ratio fails only when that phase's gate demands otherwise; phase 0 is almost entirely assumptions, correctly.

Watch for