Source-Label Preserving Summary Template¶

Source-preserving template — instantiates Use-Time Source Attribution Calibration

A summary format that forces each condensed statement to carry its source class through compression, so shortening a document can't quietly flatten observed, reported, and generated content into equally-confident prose.

Summarizing is where source labels go to die. Compressing many pages into a few sentences drops detail — and the source tag is usually the first detail dropped, so a randomized trial, a rumor, and a model's guess all emerge as "research shows." Source-Label Preserving Summary Template is a summary format whose defining constraint is that the source class must survive the compression: every condensed statement carries a tag for where it came from, and the template's language rules bind the hedge of each sentence to that tag. Its distinguishing move is that it operates at the output/compression step rather than the judgment step — it does not decide an item's source, it refuses to let an already-known source be lost when the item is shortened and restated for a downstream reader.

Example¶

An analyst compresses a sixty-page evidence dossier into a two-page brief for a decision committee. Left to ordinary summarizing, the brief would read as one smooth, uniformly-confident narrative. The template forbids that. Each synthesized statement must carry a source-class tag and phrase itself accordingly: a well-supported finding reads "multiple independent reports indicate…" and is tagged corroborated external; a claim resting on one unverified source reads "a single source states…" tagged uncorroborated; a projection reads "we estimate…" tagged inferred, never "will." Any sentence the drafting assistant generated rather than sourced keeps a visible [generated — unverified] marker until a human confirms it and removes the tag.

The brief is still two pages. But a committee member skimming it can no longer mistake the projection for the finding, or the single rumor for the corroborated pattern, because the compression preserved exactly the distinction that ordinary summarizing erases. The template changed what the sentences say, not just what a footnote records.

How it works¶

Tag rides with the claim. Each statement in the summary carries its source class inline, so the label travels with the content through condensation rather than being stored apart where it drops off first.
Hedge bound to source. The template's language policy maps each source class to a required verbal frame — observed / reported / a single source states / we estimate / generated–unverified — so the sentence's confidence is dictated by its origin, not by the summarizer's prose style.
Generated content stays marked. Any passage produced by a model or drafting tool retains an explicit unverified marker until a human verifies and clears it — the label is opt-out on verification, never opt-in.
No unlabeled synthesis. The format has no slot for a source-free assertion; a statement that cannot name its class cannot be entered as a clean claim.

What distinguishes it is where it acts: on the restatement of already-classified items for a reader, keeping provenance attached precisely when compression is trying to shed it.

Tuning parameters¶

Label granularity — a few coarse tags versus a fine source taxonomy. Fine tags preserve more distinction but clutter a summary meant to be short; coarse tags stay readable but blur real differences.
Hedge strictness — how rigidly source class dictates wording. Strict binding guarantees the phrasing matches the warrant but can make prose stiff; loose binding reads naturally but lets confident phrasing creep back over weak sources.
Label placement — inline in the sentence, versus a parallel column or bracket. Inline is unmissable but intrusive; set-aside is cleaner but is the first thing a hurried reader ignores.
Generated-marker persistence — how much verification is required to clear a [generated] tag, and whether it can be cleared at all in some contexts. Stricter prevents laundering of unverified drafting; looser keeps the summary tidy but trusts the drafter.
Compression ratio — how aggressively the source is condensed. Tighter compression saves the reader time but strains the template's ability to keep every claim's label attached and distinct.

When it helps, and when it misleads¶

Its strength is defeating the specific damage that summarizing does: it stops compression from laundering weak-source and generated material into the same confident register as observed fact, and it does so by changing the sentence a reader actually reads, not by burying provenance in an appendix nobody opens. It is where the archetype's careful source labels are made to survive contact with a downstream audience.

Its limits begin where its inputs end: it preserves labels, it does not produce them, so it faithfully carries forward a wrong source tag just as well as a right one — garbage-in, labeled-garbage-out. Kept mechanically, its tags can become wallpaper the reader tunes out, and over-tagging can bury the message under provenance. Its classic misuse is running backward for appearance — sprinkling source labels to look rigorous while the summary body still asserts everything flatly, so the format signals discipline it does not enforce. The failure it guards against is source amnesia,^[1] the loss of origin while content is retained, which compression accelerates. The discipline that keeps it honest is that the label must change the hedge of the sentence, not merely sit beside it — provenance you cannot ignore, not provenance you can.

How it implements the components¶

Source-Label Preserving Summary Template fills the downstream-preservation side of the archetype — the components that keep source distinctions intact when items are restated for a reader:

source_content_separation — it keeps the source label bound to each claim through compression, so the "where it came from" is never silently merged into the "what it says."
downstream_language_policy — its rules map each source class to a required hedge, so a summary's phrasing preserves the uncertainty its sources warrant.
generated_output_source_label — model- or tool-generated passages retain a visible unverified marker until a human clears it, keeping generation distinguishable from retrieval and observation.

It does NOT decide source class from the source_cue_feature_set or grade it with a source_confidence_annotation (the Reality Monitoring Checklist and Source Attribution Confidence Rubric do), nor resolve a provenance_backlink_or_trace or borrowed_material_credit_path against external records (Provenance Lookup Before Publication does).

Instantiates: Use-Time Source Attribution Calibration — this template is the downstream carrier that keeps source and confidence distinctions alive when items are summarized for use.
Consumes: the source labels produced upstream by the Reality Monitoring Checklist and Source Attribution Confidence Rubric — it preserves those labels through compression rather than generating them.
Sibling mechanisms: Provenance Lookup Before Publication · Reality Monitoring Checklist · Source Attribution Confidence Rubric · Memory Source Probe · Borrowed Idea Attribution Scan · Chain-of-Custody or Lineage Check · Generated Content Disclosure Gate · Hallucination Intrusion Triage · Observation Recheck or Replication · Source Attribution Training Set · Source Confusion Matrix Review

References¶

[1] Source amnesia — retaining a piece of content while losing all record of where it came from. Summarization and compression accelerate it, because the source tag is usually the first thing dropped for length — which is exactly what a source-label-preserving template refuses to drop. ↩