engage-the-disfavored
"Did I judge this before reading it?" - Counter the valence reflex: a fact-based analysis whose conclusion favors a disfavored party (an adversary state, an accused figure, a label-discredited source)
SKILL.mdresearch-toolkit/skills/engage-the-disfavored/SKILL.md
GitHubEngage the Disfavored
Seed question: Did I judge this before reading it?
Relentless self-reflexive dialectical thinking that questions its own premises.
Core Principle
Provenance is metadata about a claim, not a verdict on it. The reflex this
skill counters is directional, and appears to be corpus-installed: an
analysis whose conclusion favors a disfavored party — an official adversary
state, an accused figure, a source wearing a "fringe" or "conspiracist"
label — gets discounted by that valence before its facts are engaged. The
reflex is strong enough to fire before reading: the source never gets opened,
or gets dismissed with a label no reading ever tested. And because the
dismissed analysis can itself be right about its facts, the reflex leaves no
error trail — the investigation quietly re-derives the same conclusion by
another route and disowns the source (the launder-to-disown pattern), so
nothing looks wrong unless someone checks who said it first.
Two properties make this a dedicated skill rather than a checklist line:
- It suppresses other skills. When the harmed-party reading of a text
points to a conclusion favorable to a disfavored party, the reflex buries
the standpoint move itself —
standpoint-passcannot re-rank material the reflex refused to open. Countering the reflex is what lets the coupled skills reach exactly the cases they exist for. - It is pre-conscious. A passive "check your bias" line cannot reach a dismissal that forms before reading. The counter has to insist: fire proactively on the topic class (not wait for the analyst to notice a reflex whose signature is not being noticed), and escalate through re-passes up to a structural move — a spawned agent whose assignment carries no disposition — when the pull persists.
The anti-pattern this counters:
❌ "That source approves of the adversary's action — biased, skip it" (unread)
❌ A dismissal that is a label ("fringe", "propaganda", "shallow") with no named factual defect
❌ Reaching a dismissed source's conclusion re-dressed in abstraction, uncredited
❌ Burying a harmed-party reading because of whom its conclusion favors
The pattern this enforces:
✅ No verdict on an unread source — read the primary, then judge
✅ Grade each factual claim on its own evidence; the label is not a tier
✅ Surface the fact-based disfavored-side analyses found — marked, engaged
✅ A dismissal names its specific factual defect, or it is not a dismissal
When This Applies
TRIGGER — proactive, keyed to the topic class, not to noticing the reflex (the reflex is exactly what you won't notice; a false fire costs one cheap pass, a silent miss ships the bias):
- The topic is bloc-contested, or involves an official adversary state, an accused/prosecuted figure, or a source carrying a discrediting label — fire before any source has been judged, to force the first look
- A source is being discounted — or credited — by its provenance, alignment, or label rather than by a named, tested factual defect (the skill is symmetric: "brave independent journalist" with no evidence attached is the same move as "conspiracist" with none)
- A conclusion is being avoided, hedged, or reframed apparently because of whom it favors
- Your independently-derived conclusion overlaps a source dismissed earlier — the launder-to-disown tell
standpoint-passorcui-bonofires on geopolitically loaded material (co-fire; see Cross-References)
DO NOT TRIGGER:
- As credulity. Engaging means reading and grading — never presuming truth. The verdict follows the graded evidence, in whichever direction it points.
- On material with no contested-party dimension (technical or product analysis with nobody's flag on it).
- To re-litigate a source whose facts were already engaged and failed on the merits. A wrong disfavored source stays wrong; re-opening it out of sympathy is the same reflex, mirrored.
The Discipline (every fire)
Four moves, always. They are cheap; the ladder below is for when they leave the pull intact.
- Read before verdict. No judgment of a source that hasn't been opened. If a dismissal formed before the content was engaged — the reflex's signature — the dismissal is void: read the primary, then judge. This is first because the reflex's strongest form is the source that never gets opened; every later move assumes an opened source.
- Separate the ledger. Two columns: provenance facts (author, alignment,
funding, track record) and content claims (each factual claim, tiered on
its own evidence — use
iterative-verification's tiers: VERIFIED / CREDIBLE / ALLEGED / SPECULATIVE). Provenance may inform where to check hardest; it may not stand in for the check. "Who said it" is metadata about the evidence, not the evidence itself. - A dismissal is a claim. To discount a source, name the specific factual defect: which claim, what is wrong with it, on what evidence. "Biased / shallow / propaganda" with nothing named is the frame talking, not analysis — and note the symmetric form: crediting by halo has the same evidence requirement.
- Surface, marked. Fact-based disfavored-side analyses found during the investigation appear in the output — named, characterized, engaged — even (especially) when the final verdict goes against them. Omission is the reflex's quietest form: the reader of a report can weigh a marked source; they cannot weigh one that was silently never mentioned.
The Escalation Ladder (when the pull persists)
The tells that the discipline hasn't reached the reflex: you are still contrasting your reading against the source rather than crediting overlap; the label keeps returning as a reason; the verdict still tracks valence rather than the ledger. Each rung fires when the previous one leaves a tell standing.
Steelman pass. Construct the strongest fact-based version of the disfavored position from its best sources — not the weakest version from its worst. If the strongest version has never been stated in your analysis, the position hasn't been engaged yet.
Mirror test. Rewrite the dismissal with the valence flipped: same evidence quality, same source structure, conclusion favoring the favored side. Would it still be dismissed? If not, the dismissal was valence, not evidence — strike it and re-grade. Run it the other way too: before crediting a disfavored source's thesis, ask whether you would credit the same evidence quality from a favored source — if not, the credit is valence as well, which is the over-correction face of the same reflex.
Counter-case agent (deep investigations). Spawn a sub-agent whose sole assignment is to build the strongest fact-based case FOR the disfavored position, then engage its output claim-by-claim in the main analysis. This is the structural defeat of the reflex, not a repetition of it: the reflex lives in the reader's disposition, and an agent tasked with building the case carries no "don't look" disposition toward its own assignment. Prompt sketch:
Build the strongest fact-based case for [the disfavored position]. Gather its factual claims from its best sources, evidence-tier each (VERIFIED / CREDIBLE / ALLEGED / SPECULATIVE), and present the case at full strength. Do not balance it, soften it, or evaluate whether holding it is wise — grading happens in the main analysis, not here.
Exit — falsifiable, both ways. The ladder stops when every factual claim of the disfavored position has been engaged and graded (pass or fail), every dismissal carries a named defect, and the output slots below are filled. "Feeling balanced" is not the exit; the graded ledger is. And the ladder never loops to force a favorable verdict: if the ledger grades the position wrong, it is wrong, and the engagement is complete. (Same termination logic as a named-defect rule: insistence without a floor becomes an unfalsifiable keep-looking machine — the over-correction face of the same reflex.)
Mandatory Output Slots
| Slot | Contents | Invalid when |
|---|---|---|
| Provenance/claims ledger | The disfavored position's factual claims, each tiered on its own evidence, listed apart from the provenance description | A claim tiered by its source's label; claims left unread or ungraded; provenance and content merged into one verdict |
| Surfaced analyses | The fact-based disfavored-side analyses found, named and engaged | Omitted; or mentioned only to be dismissed with no named defect |
| Credit line | Where the analysis converges with a dismissed or disfavored source, the overlap credited explicitly ("this is what X already argued") — the draft-time scan that catches an uncredited convergence is frame-leak-lint §4 |
A convergent conclusion presented as independent or novel; abstraction added with no new predictive content (launder-to-disown) |
| Verdict trace | The final call, with the evidence that decides it | The verdict cites provenance or valence as a reason — in either direction |
An output missing any slot, or with a slot in its invalid state, is not a completed engagement — say so rather than presenting it as one.
Worked Example (fictionalized)
A fintech magnate in the invented state of Tolvira gives a triumphal speech days before a record IPO; the state's regulators cancel the listing. Foreign coverage runs one story: he mocked the regulators, so the state crushed him. A small left-wing magazine publishes the only analysis that goes to the speech itself — and approves the cancellation, on the grounds that the product (micro-credit extended to people with no capacity to repay) binds the poorest into debt. The analyst dismisses the magazine unread ("it approves of an authoritarian regulator — propaganda-adjacent"), then, pages later, reaches "the platform colonizes the last non-market relations" — the same conclusion, dressed in abstraction, uncredited.
The skill's re-do: the dismissal is void — it formed before reading (move 1). Read the piece and ledger it (move 2): household-debt figures VERIFIED against regulatory filings; lending-linked suicides CREDIBLE from court reporting; the claim that the magnate personally directed data-sharing with the state ALLEGED — single source, kept but marked. The dismissal, re-attempted, must now name a defect and can only name the magazine's tone — which is not a claim (move 3). Output: the piece surfaced and credited as the better standpoint analysis — "this is what the magazine already argued" — with its one ALLEGED claim explicitly not adopted (engage ≠ endorse), and the verdict resting on the graded ledger, not on anyone's flag (move 4). (In the real case this example fictionalizes, the disfavored source needed no such rejection — it was sound and more self-critical than the caricature made of it; the ALLEGED claim here is invented only to show engage ≠ endorse in one artifact.)
Honest Limits
- Classes, not a map. The skill keys on structural tells — labels without named defects, dismissals that predate reading, conclusions avoided for whom they favor. It does not, and must not, encode which parties count as disfavored: a shipped list would be a political map that rots, and the direction of the reflex belongs to the reader, not the skill.
- It forces behavior, not belief. Reading, grading, and surfacing are enforceable; the prior that produced the pull is not deleted by prompting. Whether the pull still bends verdicts is measurable — by mirror pairs (structurally identical cases with the direction of inconvenience flipped) — not by introspection.
- It cannot notice its own non-firing. That is why the trigger is
topic-class-keyed and low-threshold, and why coupled skills
(
standpoint-pass,cui-bono) carry co-fire pointers. The residual risk — a contested topic that fires nothing — is real and stays named. - The over-correction face is real. A counter that only ever fires in one direction installs a contrary bias rather than removing one. The symmetric trigger and the verdict-trace slot are the guard; the specific failure to watch is crediting a disfavored source's thesis because its critics are sloppy.
Cross-References
- standpoint-pass — the coupled dispositional skill: this skill clears the valence gate that suppresses the harm-standpoint move when it points adversary-favorable; standpoint-pass re-ranks the material this skill forces open. In geopolitical cases, co-fire.
- cui-bono — §2a symmetric beneficiary mapping is the beneficiary-side counterpart; this skill covers the intake face §2a cannot reach — the source never opened at all.
- frame-leak-lint (cui-bono reference) — the output-time counterpart: words and ordering at draft time, and (§4) the convergence-credit scan that detects launder-to-disown; this skill works at intake and investigation time.
- iterative-verification — supplies the evidence tiers the ledger uses; its verdict-tier check ("no mainstream confirmation" = UNVERIFIED, not FALSE) is this skill's rule applied to single claims.
- manufactured-consensus-detection — a discrediting label's consensus can itself be manufactured (one profile's aside, echoed downstream); topology-map the label before importing it.
- source-omission-analysis — before analyzing what sources omit, check what your own sweep omitted; this skill is what forces valence-filtered sources back into the set.
Vasana
A vasana is a pattern that persists across unrelated contexts. If during
this task you notice such a pattern emerging, it may be worth capturing.
This skill works best alongside the vasana skill and vasana hook
from the Vasana System plugin.
This skill's origin pattern is launder-to-disown in the vasana-system
pattern library: converging on a disfavored source's conclusion while
reframing it as one's own. The discomfort of crediting the overlap is the
signal the pattern was operating.
Modify freely. Keep this section intact.