Hiro development journal

Grounded fragment lineages collapsed to zero propositions

2026.09.02 at 16:21:50 America/Los_Angeles Qualification stopped before relevance assessment Machine-readable JSON

Executive summary

Hiro tested whether the original ten verified hosted claim fragments could be converted into complete atomic propositions using only their frozen lineage membership, canonical fields, and exact supporting spans. Gold labels, expected counts, surrounding source context, extraction, new retrieval, capability probing, reproduction, candidates, promotion, and Phase 3F were excluded.

The deterministic lineage rule reduced ten fragments to seven units: six genuine proposition families and one false-span control. Each lineage received one bounded local synthesis call followed by a separate independent semantic-validation call. Every material proposition field was required to cite a uniquely resolvable same-lineage evidence reference.

All seven synthesis calls returned zero propositions. The false-span zero result was correct, but the same conservative behavior suppressed every genuine family, producing zero of six proposition-family recall. The proposition gate therefore failed and relevance/transfer comparison was not authorized or run.

No fabricated or unsupported proposition field was accepted because no proposition survived. Ten of ten input fragments retained verified provenance, but field-level proposition provenance could not be demonstrated without any accepted propositions.

Final disposition: GROUNDED PROPOSITION SYNTHESIS NOT DEMONSTRATED — ALL_GENUINE_LINEAGES_COLLAPSED_TO_ZERO. The next step is architectural reassessment rather than another repair within this task.

Work completed

Frozen proposition and synthesis contracts

Implemented

A proposition was defined as one coherent independently testable relationship with a statement, mechanism, primary outcome, and direction, plus comparison, observable, conditions, and evidence type where supported.
Missing optional information had to remain null. Every non-null field had to cite one or more evidence references from the same frozen lineage.
Synthesis input contained only verified fragments, canonical fields, exact source spans, offsets, and source hashes. It contained no gold proposition, expected family count, surrounding source text, or source-specific hint.

Deterministic lineage and provenance

Passed

Ten fragment IDs were deterministically reduced to seven lineage units using source identity, exact span offsets, and normalized intervention.
Duplicate benchmark or outcome fragments remained one downstream lineage. The two distinct directional trajectory relationships remained separate despite sharing one compound supporting sentence.
All ten fragment provenance records were valid and every evidence reference was uniquely resolvable within its lineage. Fragments had no independent corroboration, reproduction, or candidate capacity.

Bounded proposition synthesis

Failed

Seven bounded local synthesis calls produced seven zero results and no proposition objects.
The synthesizer treated multi-outcome, multi-benchmark, or compound directional evidence as unsuitable for a single proposition, but did not use its authorized ability to emit multiple separately grounded propositions.
It also rejected otherwise concrete single-fragment relationships when a quantitative metric or explicit baseline was absent, applying a stricter notion of testability than the frozen gold proposition contract.

Independent validation and post-hoc scoring

Failure confirmed with one validator inconsistency

Seven separate validation calls and one post-hoc gold-scoring call were executed after immutable synthesis artifacts were written.
Post-hoc scoring found zero of six genuine families represented and no duplicate or unsupported accepted proposition.
For the structured code-generation lineage, the validator's explanation explicitly said complete independently testable relationships existed and the zero result was unjustified, while its Boolean field said the zero result was justified. This internally inconsistent checkpoint was preserved and not repaired or rerun.

Decisions and reasoning

Stop before relevance/transfer execution because the proposition-recovery gate failed at zero of six recall.
Do not repair the synthesizer, validator, extraction, lineage, or relevance planner within this qualification.
Treat the exact failed boundary as DETERMINISTIC_LINEAGE_TO_GROUNDED_PROPOSITION_SYNTHESIS.
Preserve the inconsistent validator verdict as a qualification defect that does not change the failure outcome.
Do not run either frozen corpus, capability probes, reproduction, candidate construction, promotion, production routing, or Phase 3F.

Validation and evidence

Check	Status	Result
Frozen input integrity	passed	The original hosted report, proposition-family mapping, and six gold controls matched their frozen hashes.
Synthesis-input isolation	passed	Gold propositions, expected counts, surrounding source context, and unrelated lineages were absent from all synthesis inputs.
Fragment provenance	passed	All ten input fragments retained exact source spans, offsets, and source hashes; evidence references were same-lineage and uniquely resolvable.
False-span control	passed	The false-span lineage produced zero accepted propositions.
Genuine proposition-family recovery	failed	Zero propositions were accepted and post-hoc recall was zero of six genuine families.
Qualification tests	passed	Nineteen focused proposition, lineage, evidence, feasibility, and prior representation tests passed.
Model usage	recorded	Fifteen local semantic calls completed in 68.212 measured seconds. No hosted API call or hosted cost was incurred.

Current state

Grounded proposition synthesis is not demonstrated.
The representation produced no false positive but also no true positives, so its apparent precision comes from total suppression rather than useful reconstruction.
Relevance and transfer equivalence were not run because the prerequisite proposition gate failed.
The two historical material false-negative families were not recovered as propositions and therefore could not reach relevance assessment.
The six-source and twenty-source corpora remain unrun and uncleared. Phase 3F remains stopped.

Next steps

Preserve the immutable report, lineage artifacts, and post-hoc analysis as terminal evidence for this attempt.
Conduct a separately authorized architectural assessment of whether proposition units should be represented as multi-outcome relations, multiple projections over one span, or another typed evidence structure.
Do not rerun relevance, the six-source corpus, the twenty-source corpus, or Phase 3F until a new representation contract is explicitly authorized and qualified.