The mechanism narrative — the complement to the GO layer

The GO-overlap analysis evaluates Affinage's mechanism_profile GO
terms and finds them systematically over-general (1/42 specific primary-function
matches). Note the distinction from the raw script metric: compare_affinage.py's
core_mf_captured flag is True for 4/42 genes (AATF, ABL1, KRAS, GAPDH — the
per-batch summary.md headers show these un-refined counts), because it fires on any
authored core-MF term. On manual inspection only KRAS (GTPase activity) is the
gene's specific primary activity; AATF/ABL1/GAPDH match on general/secondary core terms
(RNA binding, DNA binding). So "1/42" is an editorial refinement of the computed "4/42",
not a separate hard-coded number.
But that is Affinage's weakest output. Its actual product is the
citation-anchored narrative.mechanistic_narrative plus the structured
timeline.discoveries. This note evaluates that layer, because judging Affinage by
its GO grounding alone is unfair to it.

Sampled genes: GPX4, CASP3, MAPK1, ADRB2 (full records re-fetched; the committed
cache is trimmed and does not include these narrative fields — re-fetch with the API
to reproduce). Quotes below are verbatim from the mechanistic_narrative field.

The narrative recovers exactly what the GO layer drops

For every gene where the GO layer collapsed to a top-level parent, the prose names
the specific mechanism — matching the curated core function the GO grounding
missed — and cites it densely (distinct inline PMIDs in parentheses):

Conclusion: the GO mechanism_profile is a lossy down-cast of a narrative that
clearly encodes the specific function. The GO layer does not measure what Affinage
knows; it under-reports it.

The discoveries are structured and confidence-graded

Each entry in timeline.discoveries is not free text but a structured object:

year, finding, method, journal, confidence, confidence_rationale,
pmids: [...], is_preprint

plus a parallel teleology track recording what each advance explained (e.g. for
GPX4: "established the catalytic basis for GPX4's distinctive substrate range…").
This is a richer, more curator-usable artifact than the GO layer — closer to a
graded literature review than an annotation set.

Two real failure modes of the narrative

1. Recency / novelty bias on canonical genes. The ADRB2 narrative surveys recent
   specialized findings (HCC sorafenib resistance, amyloid-β, CAR-T checkpoint,
   osteoclastogenesis) but omits the textbook core mechanism: no cAMP, no
   adenylyl cyclase, no Gs coupling, no β-arrestin desensitization. For a
   heavily-studied pleiotropic receptor it reads like a literature-mining digest, not
   a mechanism primer — high on specific recent detail, low on canonical completeness.
2. Symbol collisions break even the prose. ADA's narrative is a chimera of three
   "ADA/Ada" entities (E. coli Ada, the eukaryotic ADA2/ADA3 SAGA/ATAC subunits,
   and human adenosine deaminase) — see the project page
   §3.

A useful built-in triage signal

Affinage's own evaluation.pairwise (win / tie / loss vs the curated UniProt
reference) tracks these tiers cleanly in the sample: GPX4, CASP3, MAPK1 = win
(specific, accurate); ADRB2 = tie (recency-biased, canonical gaps); ADA = loss
(entity collision). That field is a ready-made flag for which narratives to trust
before ingesting them.

Implication for AIGR — largely redundant with deep research

The narrative reads well, but as an input source it does not clearly beat what
AIGR's existing deep-research step already produces, and most of its apparent
advantages are trivial to replicate:

• Biological content overlaps heavily with our deep research. For well-studied
  human genes our falcon/perplexity reports already recover the specific mechanism
  (the falcon GPX4 report independently gives Sec46, the catalytic tetrad, the
  system xc⁻–GSH–GPX4 axis, ferroptosis) — Affinage adds no new biology.
• PMID-anchored citations are not a differentiator. DOI/[n]→PMID resolution is
  a one-liner (NCBI/EuropePMC ID converter), so the fact that Affinage emits inline
  PMIDs while falcon/perplexity emit DOIs or footnote numbers is cosmetic.
• "More citations" is not value. For GPX4, Affinage cites 37 PMIDs vs the 10 in
  our completed review (1 overlap). But curation selects the few core papers; the
  extra 36 skew toward the pleiotropic/recent material a curator prunes (the same
  recency bias as ADRB2). Count ≠ quality, and the novel PMIDs were not shown to be
  primary evidence for core function that we lack.
• No independent perspective. Affinage is Claude-generated (Sonnet read + Opus
  synthesis); AIGR's review agent is also Claude, so ingesting it adds no model
  diversity and risks amplifying shared biases — unlike our multi-provider ensemble.
• Human-only. AIGR spans 190+ species; Affinage can never touch the non-human
  majority of reviews.

Residual (minor) value. It is a free, precomputed artifact for all 19,293 human
genes, so for the human review backlog it could save the compute of a first-pass
deep-research run — a convenience, not new capability, and subject to the
correlated-error and human-only caveats above. Its flag of ~10% of the human proteome
as mechanistically dark is a weak prioritization signal (project-management, not a
per-review source).

Recommendation: do not wire Affinage in as a distinct review source on this
evidence. The durable output of this exercise is the evaluation result (the GO layer
grounds to the specific function ~1/42 while the prose gets it right), not an
integration.