Epistemic safeguards

Pattern-finding, suspicion, rigour and researcher safety.

The literature on safeguards falls into two halves. One half, from Sedgwick, Felski and Latour, describes the failure Holagram risks: a strong reading that anticipates what it will find, absorbs counter-evidence, and hands its tools to conspiracist uses. The other half is practical and mostly comes from outside the humanities: intelligence analysis (Heuer, the UK Red Teaming Handbook), philosophy of science (Chamberlin, Longino), social psychology (Mercier and Sperber, Nemeth), qualitative methodology (Lincoln and Guba, Haven and Van Grootel) and adversarial collaboration (Melloni et al.). The rule that no reading is tested by its own author has good precedent: Chamberlin on affection for a favourite theory, Longino on objectivity as a property of a critical community, and Mercier and Sperber on people judging other people's arguments better than their own. The rule is necessary but not sufficient. Nemeth's work suggests that an assigned devil's advocate can entrench the view it attacks, and the sycophancy research (Sharma et al.) makes the same worry concrete for a synthetic Adversary. Koehler's work suggests the Fiction register may raise confidence in a reading. Evidence that named techniques work is thin; Whitesmith's tests of ACH are discouraging. On researcher safety the sources cover emotion work in sensitive research, not the pattern-finding mode itself. Nothing found tests a method like Holagram directly. Where an entry relies on memory for a source's argument, it says so.

Sources

Ideas to take

Stamp a Prediction before the Fiction, and agree the outcome map in advance.

Changes: Split the Test into a Prediction, written and dated straight after the Reading and before the Fiction, and a Result written after the check. The Prediction says what would count as 'broke' and what would make the check inconclusive (Melloni et al.'s preregistered map had a 'challenged' outcome and marked weak negatives as less conclusive). The register order becomes Trace, Reading, Prediction, Fiction, Result.

Risk: Preregistration sits awkwardly with interpretive work that is meant to change in the field; Haven and Van Grootel treat qualitative research as largely postdiction. Forbidding revision makes the method brittle, and allowing it freely empties the stamp. Log dated revisions instead of banning them.

Require two rival Readings per Trace, one of them mundane.

Changes: The Reading register takes a required rival in which nothing is hidden (the clerical, budgetary or accidental explanation). The Test is then written to discriminate between the two, as in ACH, where evidence consistent with every hypothesis is set aside.

Risk: Whitesmith's experimental tests are negative about ACH's theory, and a matrix can give false comfort. Both rivals are still produced by one mind, so the true alternative may be missing. Treat the matrix as a discipline of listing and let the outside reader add rivals.

Write the Adversary protocol as authentic dissent, not role-play, and check it for sycophancy.

Changes: The Adversary receives the Trace and the Reading but not the Fiction, must argue from sources a human can open, and has its agreement logged. Run a sycophancy check by presenting some Readings as another person's. Add a 'contra memo' step in which the author writes the strongest case that the Reading is wrong before any check. The outside reader, not the Adversary, closes a Test as held.

Risk: Nemeth's work suggests assigned opposition can entrench, Sharma et al. suggest the agent may flatter, and Costello et al. suggest an AI arguing with evidence can work. These do not settle which regime Holagram is in. A fluent Adversary can invent sources, which is why Traces and Test sources must be openable by a human. Costs reader time.

Give the outside reader a job description drawn from Longino's four norms.

Changes: Beyond the power to stop, the outside reader reads the log on a fixed day, is the only one who can mark an Entry 'held', can require a written reply from the author in the log (uptake), and works from standards written before the study. Add negative case analysis and a Trace archive for audit as standing duties, after Lincoln and Guba.

Risk: One reader is a thin community. If reader and author share a frame (discipline, politics, friends), the rule fails in the way Longino and Mercier and Sperber predict. Choose a reader likely to disagree, and for Traces a second reader from outside planning. For a solo PhD such readers are scarce.

Choose some Traces by rule, tag them, and hunt for negative cases.

Changes: Tag each Trace as 'drawn' (selected by a rule set in advance: every nth page of the document, every fixed distance on the walk) or 'found' (noticed). Report the ratio, and mark Readings that rest only on found Traces. For each Reading, name one Trace that should exist if it were true and look for it.

Risk: Rule-based sampling dilutes the serendipity the method depends on, and Ginzburg's clue-reading rests on the odd detail. Someone still sets the rule. Berkowitz's anecdote of decoding accidental scraps shows that the rule limits selection, not interpretation.

Keep the Fiction quarantined and use it twice.

Changes: Fiction is written by other hands (Alter) after the Prediction is stamped, and a second short line is added, set in the world where the Reading is false and explaining why it looked true. Eco's question (does the rest of the text bear it out?) is applied to the Fiction as well: where it needs material the Reading does not supply, mark it.

Risk: The second line adds length to a 150-word form and may harden into boilerplate.

Put wellbeing criteria into the stop rule and cover the agent sessions.

Changes: The outside reader's stop rule gets written criteria agreed before the study (sleep, patterns seen outside reading blocks, refusal to let a Reading go, distress). It covers sessions with Alter and Adversary as well as fieldwork. The consent line says plainly that the method asks for suspicious reading. Reading-mode on and off each get a short routine.

Risk: The literature concerns sensitive topics and emotion work, not pattern-finding as such, so transfer is by analogy. Written criteria can medicalise a creative method or be ignored by an author who is, by design, enjoying the reading. Østergaard's argument is a hypothesis.

Present the rules as the method's rigour claims, with a breach log.

Changes: The method statement lists the rules as commitments an examiner can audit (hands, no self-testing, failure log, outside reader) and adds a breach log recording each time a rule was broken and why. The conventional write-up states how many Entries were drawn or found, how many held, broke or stayed unchecked, and who closed each.

Risk: Biggs and Büchler were not read in this run, and examiners may want an established criteria set (Lincoln and Guba, Tracy) in place of a new one. A breach log is honest but hands a hostile reviewer material.

Counter-arguments

Gaps