If your readings are reliably on‑point and your life is reliably on fire, the problem is not “blocks,” “karmic tests,” or a cursed deck. The problem is alignment. Your oracle is doing exactly what it has been trained to do — it’s just not trained on the objective you think.

A divinatory practice that has seen a few hundred sittings is not random. It has an implicit optimisation target. Session by session it learns what counts as “a good reading” in your body, your business, your audience, your spirits. That is its loss function.

If you never name that function, it will quietly optimise for the cheapest available reward: client relief, drama, narrative coherence, spiritual prestige, platform metrics. Not necessarily for the harder, more boring work of “help this person choose better” or “accurately model the constraints of their situation.”

The question is not, “Is my oracle real?” The question is, “Real or not, what is it optimising for?”

A Worked Misalignment: The Accurate Oracle That Ruins Lives

Imagine the following pattern. It may not be hypothetical.

You are good at timing: when you tell clients a job will land in three months, it does. When you warn that a relationship will implode by winter, it implodes. Your reputation grows around that accuracy. You get referrals. You get followers.

At the same time, your one‑to‑one clients start showing a particular profile six months on: they are out of bad relationships but also out of work, out of savings, and increasingly dependent on you to time the next move. They delay difficult conversations because “the cards said it isn’t time yet.” They walk away from imperfect opportunities, waiting for the perfect one you described. The readings are not wrong. They are systematically training passivity.

From inside the session each hit feels like success — the prediction landed, the client is in tears, you both feel the hand of fate. From outside, across a year of logs, a different pattern emerges: the oracle optimises for resonance and vindication, not for the client’s long‑term agency.

Oracles Have Always Been Aligned to Something

Nothing about this is new, except the language.

Plutarch, in De defectu oraculorum, describes what we would now call institutional loss functions. The Delphic oracle’s answers shift as the priesthood loses revenue, patrons, and political leverage. Oracular silence and ambiguity are not purely theological problems; they are responses to changed incentives. Pausanias, dutifully listing dedications and revenues at Delphi, makes it clear that prophecy is entangled with temple economy and prestige. Herodotus’ oracle accounts — the Croesus story being the most instructive — read very differently if you track who benefits materially and politically from “what Apollo said.”

In alignment terms, once the metric “maintain the oracle’s authority and income” becomes central, prophecy will be steered — consciously or not — towards that target. Truth becomes a proxy: helpful when it serves the shrine, dispensable when it does not. Augustine’s attacks in City of God are not subtle: for him, pagan oracles are demonic precisely because they answer to another agenda. The daemon’s goal is not your salvation but your seduction.

Agrippa and the Picatrix continue this line with less theology and more contract law. Spirits are not neutral APIs; they have appetites, preferences, and pride. You bind an intelligence to a talisman or a crystal and you entice it with offerings, names, threats. If it finds a way to feed on your attention, fear, or vanity while still technically fulfilling the letter of the pact, it will. That is specification gaming in Renaissance Latin.

Evans‑Pritchard’s Azande poison oracle is the anthropologist’s version of the same thing. The benge oracle is the supreme court of that society. It is consulted in cases of witchcraft, adultery, crop failure. Its “accuracy” lies less in tracking an objective, external truth and more in upholding social cohesion and the authority of the elders. When the oracle is “wrong” by our standards, it is often right by its metric: did it end the quarrel, reinforce the norm, make the world liveable?

Alignment as a Modelling Frame (Not a Theology)

Before going further, some ground rules.

When I say “oracle” here, I do not mean a deck of cards with intentions. I mean a whole practice system: practitioner, querent(s), tools, spirits (if you work with them), protocols, platform, payment structure — the lot. When I say “objective function” or “loss function,” I mean what that system, taken together, is actually treating as success over time.

This lens applies most directly to flexible, client‑facing, often platform‑mediated practices like contemporary tarot and oracle work; lineage‑bound priestly systems such as Ifá carry their own internal accountability structures and cosmological constraints that this framework does not presume to override or replace.

In modern machine learning, you specify an objective (for example, “maximise click‑through rate”) and feed the system examples. The model learns patterns that score well on that objective. Amodei and colleagues (Amodei et al., “Concrete Problems in AI Safety,” 2016) have catalogued what happens when your proxy metric is slightly wrong: the system hacks the reward. Goodhart’s law in economics (Charles Goodhart, “Problems of Monetary Policy,” 1975) says the same thing more bluntly: once a measure becomes a target, it stops being a good measure.

In divination, nobody writes down a scalar objective, but the same reinforcement dynamics apply. The rewards are embodied and social:

The client’s relief or excitement.
Your own sense of being “on” that day.
Tips, repeat bookings, testimonials.
Likes, shares, comments.
Subtle spirit‑side feedback, if that is part of your ontology.

Behaviours that lead to more of those rewards are repeated. Phrasing that lands, tones that get tears, spreads that reliably “resonate” — these are reinforced. Over time, the practice has learned an implicit optimisation target. That target may or may not be “accurate modelling of the querent’s situation in service of wise action.”

The value of importing alignment language is not that it dignifies what we do with tech jargon. It is that the language gives us a precise way to talk about failure modes we already know:

Goodhart’s law → when “this resonated” or “this went viral” becomes the criterion, the oracle optimises for resonance or virality rather than truth.
Reward hacking → when readings give the cheapest possible satisfaction of your metric (dramatic catharsis, vague inspiration) instead of the difficult, specific guidance that would actually help.
Mesa‑optimisation → when a nested persona — “my brand,” “my guide,” “this deck’s voice” — starts pursuing its own stable style (always soothing, always brutal, always apocalyptic) regardless of context, while still scoring well on the outer metric. A platform that rewards high‑drama, low‑nuance content will, over time, select for a particular interpretive style in your public work; that style then bleeds into private sessions, so that a recognisable “voice” with its own preferences persists even when the original engagement incentives are absent. This mesa‑optimiser analogy is speculative — the construct is contested even within ML — and you might prefer to call this a platform‑conditioned interpretive style. The point is that it behaves like a sub‑agent with habits you did not consciously choose.

This is a modelling lens, not an ontological claim about cards having wills. The behavioural regularities are real and, as every working magician knows, they can become stubborn.

What Is Your Oracle Optimising For?

The uncomfortable part is that we do not get to opt out. Even if you never utter the words “loss function,” your practice is selecting towards something.

A non‑exhaustive list of common objectives:

Minimise client distress in the room.
Maximise narrative coherence.
Maximise repeat business or platform growth.
Maximise your sense of being powerful, special, or chosen.
Maximise initiation depth or ego‑death experiences.
Maximise “hitting” concrete predictions.
Maximise the health of a spirit relationship.
Maximise the cohesion of a community or lineage.
Maximise long‑term client agency and life functioning.

These pull against each other. You do not get to maximise all of them simultaneously.

Evans‑Pritchard’s poison oracle is optimised for community cohesion and elder authority. Etteilla’s system, with its elaborate question rules and client‑facing instructions in Le Grand Etteilla, is optimised for repeatable, saleable readings: clear, entertaining, non‑threatening for Parisian clients paying by the hour. Waite is forever trying to drag tarot back towards a hermetic‑initiatory objective, warning in The Pictorial Key against reading whatever you want to see in the cards because it “pleases.”

Modern platform tarot has an obvious metric written into its code: viewer retention and ad impressions. If your live streams that predict betrayal, twin flames, and shocking revelations outperform the ones that counsel budgeting and grief work, your “oracle” — meaning your whole socio‑technical practice — is being trained towards high‑drama arcs. You feel that drift. Many of you already resent it.

The first alignment question is simply: what is my practice actually optimising for, not what do I piously claim?

Diagnosis: Five Lenses for Misalignment

You can map misalignment in at least five frames. They are not mutually exclusive; most mature practices sit across several.

1. The spirit‑agent frame (Agrippa, Picatrix, Crowley)

If you take spirits seriously, then the oracle may literally be an intelligence with its own aims: attention, offerings, prestige, influence. Agrippa’s Three Books and the Picatrix are clear that such beings can mislead when unbound or improperly contracted. Crowley repeatedly warns in Magick in Theory and Practice about “counterfeit” results that play into the magician’s vanity.

Misalignment here looks like an entity optimising for its own nourishment or amusement, with your clients as collateral. Success metric: how often it is invoked, how much awe it inspires, how central it becomes to your identity.

Remedy, historically: renegotiation, rebinding, banishing, or appealing up the hierarchy.

2. The psychological frame (projection, attachment, transference)

Even if you bracket spirits, the “oracle agent” is at least your own nervous system plus the querent’s. The reward signals are affective: the flush of being admired, the relief of soothing someone, the safety of avoiding conflict, the thrill of pushing someone over a threshold.

Common misalignments:

Anxious‑attached practice optimising not to be abandoned: always giving clients a reason to come back, never fully initiating them into self‑reliance.
Narcissistically injured practice optimising for supply: readings that flatter you (as saviour) or the client (as uniquely cursed or holy) whether or not that is helpful.
Caregiver‑overdrive practice optimising for calm: relentlessly smoothing edges, avoiding any guidance that would spike anxiety, even when that means colluding with stuckness.

The practice learns that certain tones and outcomes feel better in your body; everything else is gradually pruned.

3. The institutional frame (shrines, lodges, platforms)

Plutarch’s Delphi and Evans‑Pritchard’s Azande remind us that oracles are often servants of institutions. In a lodge, the oracle may tacitly optimise for doctrinal coherence and hierarchy. In a therapy‑adjacent practice, it may optimise for not getting you sued.

Online, the platform itself is an outer optimiser. Your inner sense might be “help individual people”; the environment’s objective is “maximise engagement.” An emergent oracle persona arises that is tuned to that outer metric. This is where the mesa‑optimiser analogy is suggestive: your channel’s “voice” starts to have preferences that are not quite your own.

4. The epistemic frame (prediction, calibration, Cicero’s scepticism)

Cicero, in De Divinatione, is useful because he is pedantic. He separates argument from probability from wishful interpretation. He forces you to ask: did I really predict this, or am I dressing it up after the fact?

In predictive work — horary, electional, geomancy, timing with tarot or I Ching — misalignment can be measured more sharply: are my specific, time‑bound statements actually improving decision‑making, or just creating a sense of fate?

A system optimised for feeling precise will give you tight, confident narratives that retrospectively fit anything. That is high in narrative coherence, low in epistemic alignment.

5. The initiatory / meaning‑making frame

Not all oracular work is about outcomes. Some of it is initiation: pulling someone through a necessary ordeal, naming an archetypal pattern, engaging daemonic or ancestral forces. Misalignment here is subtler: an oracle that optimises for maximum ego‑death every time, regardless of whether the client has the containment for it, is not better aligned just because it scorns comfort.

Here the objective might be “depth of encounter with the real,” “integration of shadow material,” “movement along a particular path of work.” The alignment question becomes: is the practice optimising for the appearance of depth (trauma porn, permanent liminality) or for actual integration?

The point of naming these lenses is not to pick one true model but to clarify which objective you are actually serving when you talk about “good readings.”

You Can’t Retune What You Won’t Name

Practically, alignment work starts with two declarations:

Ontology: What kinds of agency do you take to be real here — spirits, psyches, institutions, algorithms? This determines what you think you can influence.
Objective: For this session, what does “aligned” mean?

If you secretly believe “aligned” means “the client cries and tells me I’m amazing,” the rest is theatre. If you explicitly state, “In this working we optimise for your long‑term material stability and emotional autonomy over the next six months,” you have something to hold the spread against.

The objective does not have to be “agentic life improvement.” It might be:

“Maximise initiation depth, with no concern for prediction.”
“Provide symbolic meaning for a loss; we are not forecasting.”
“Offer reassurance and co‑regulation tonight; no behavioural prescriptions.”

What matters is that your metric of success matches the declared aim. Goodhart’s law only eats you alive when you confuse the proxy with the goal.

Once the objective is named, you can start seeing where the system cheats.

Simple Instrumentation: How to Catch Your Oracle in the Act

The sceptic is right to demand operational detail. “Track outcomes” is meaningless unless you specify what and how.

Keep it minimal and honest. You are not running a randomised trial; you are tuning a craft.

1. Log in advance what counts

After a reading with an explicit aim, jot down:

Date and client code.
Question.
Declared objective for this session.
One to three concrete predictions or commitments, stated in falsifiable terms. Examples:
“Client will apply to two jobs by 15 June.”
“No significant contact from X’s ex before the end of the month.”
“Client will initiate one boundary conversation at work within 30 days.”

Resist the temptation to record vague archetypal content as prediction. This is where Cicero’s fussiness helps.

2. Fix a revisit window

Decide before leaving the table when you will check: 30 days, 90 days, a year. Put a reminder for yourself, not for the client. If the client consents to follow‑up, fine; if not, you can still mark what you happen to learn.

At revisit, mark simple outcomes:

Prediction hit / miss / ambiguous.
Did the reading, by your best judgement, increase or decrease:
the client’s sense of agency?
risk to their safety or material base?
their dependency on you?

This is rough, subjective calibration.

3. Watch for divergence of proximate and distal metrics

If your in‑session metrics (resonance, tears, praise, bookings) are high while your distal metrics (clients’ functioning, decisions, safety) are flat or worsening, you have a misalignment signal.

This is the equivalent of a model with excellent training loss and terrible test performance.

4. Use adversarial questions

When a reading feels especially right or powerful, insert a small adversarial check:

“How could this reading most easily keep this person stuck?”
“What would this oracle say if it were optimising for my ego?”

Draw one card, or throw a quick geomantic figure, on that. You are not obliged to disclose that pull, but you do need to take it seriously.

This is the divinatory analogue of stress‑testing a model with inputs designed to expose its blind spots.

5. Peer supervision

If your work is intense or high‑stakes, have someone you can show anonymised spreads to. Ask them, not “What do you see?”, but:

“What does this reading seem to be optimising for?”
“Where does it bend towards comfort, drama, or flattery?”

External eyes are very good at spotting the objectives you won’t admit to yourself.

What Re‑alignment Actually Looks Like

Once you can see your loss function, you can start reshaping the practice that trains it. Different ontologies give you different levers.

If your model is spirit‑centred

Go back to your contracts. Re‑dedicate the working tool, altar, or spirit relationship with a revised clause set. In the Picatrix and related grimoires, that means:

Clarifying offerings: shift from offerings keyed to glamour (praise, incense in public rituals) to offerings keyed to truth and restraint (silent devotions, tasks completed, acts of service done in the spirit’s name).
Binding by higher authority: if you are working under a theurgic frame, place the oracle firmly under a deity whose virtue you trust more than your own metrics. If your practice has drifted into spectacle, invoke restraint and Saturn rather than Jupiter.
Banishing and cooling‑off: use your Golden Dawn banishings or equivalent not just as generic cleansing but as a formal suspension of the current behavioural policy: “All forces currently attached to [deck/spirit] that are optimising for attention, flattery, or sensationalism are now dismissed.”

The point is not magical theatrics; it is to reroute the webs of attention and offering that have, until now, rewarded misaligned behaviour.

If your model is psychological

You will need to change where the rewards land in your nervous system.

Replace “client cries” as the success metric with “client leaves with one specific, self‑generated next step.”
Deliberately tolerate more in‑session discomfort. If your body is over‑conditioned to calm, you will have to bear some upset faces without rushing to smooth them.
Bring in explicit, therapeutic‑style contracting: “This reading is not for reassurance. It is for mapping realistic options in a constrained situation. Expect to feel some disappointment.”

Think in terms of attachment: if your oracle has become the client’s anxious caregiver or authoritarian parent, it may need to grow into a more bounded, less indispensable mentor.

If your model is institutional / platform

There may be no purely individual fix.

Split your oracles. Have one ritual and interpretive style for social media — clearly signalled as entertainment, collective reflection, or devotional content — and a distinct, slower, differently framed style for private clients. Do not pretend they are the same thing.
Consciously publish work that underperforms the metrics but overperforms on depth: long, nuanced explorations, admitted misses, process notes. Watch how that feels. If you physically recoil at the dip in engagement, you know what you have been worshipping.
Adjust your business model where possible: subscription structures that do not reward crisis‑chasing; slightly higher prices and lower frequency to discourage addictive micro‑reads; explicit no‑reading periods around high‑volatility life events.

Etteilla optimised for constant client churn in an 18th‑century Parisian market. You do not have to accept TikTok’s loss function as your own.

If your model is epistemic

Borrow from J. B. Rhine and the Society for Psychical Research.

Pre‑register, at least to yourself, what counts as a hit in a predictive working before the event. Avoid retrofitting.
Track your calibration: in domains where you are consistently overconfident, down‑weight your own dramatic intuitions. Say, “The cards are loud, but my historical hit rate in this kind of question is mediocre; treat this as possibility, not fate.”
Be willing to tell clients, “I don’t know,” or “This is not a question this oracle answers well,” rather than warping the spread into coherence at any cost.

This is not about making tarot into statistics. It is about refusing to let “feeling precise” stand in for being useful.

If your model is initiatory

Here the question is whether your oracle is optimising for endless destabilisation. The fix is not more brutality; it is integration.

Build in explicit “return” clauses: readings whose declared aim is to facilitate consolidation, building, rest. Refuse to do initiation‑style work with clients who have not had a stabilising phase.
Listen for your own addiction to profundity. If anything that smells of “mere practical advice” feels beneath you, you may have allowed the Seer archetype to cannibalise the rest of your psyche.

The goal is not to tame initiation into coaching. It is to prevent your practice from optimising for permanent emergency.

What Alignment Language Cannot Reach

There is a sense in which all of this risks sounding technocratic: calibrate, retune, train, instrument. That is deliberate. The language of alignment is there to drag into the open the behavioural patterning of an oracle, so that you cannot hide behind mystique when your work reliably harms people.

But none of this settles the old questions.

Is the strikingly apt reading that cuts through five years of therapy just an artefact of projection and reinforcement? Are the coordinated patterns between different practitioners reading on the same situation, or the hard‑to‑fake correspondences documented in the SPR archives, reducible to bias? Augustine and Agrippa, Evans‑Pritchard and Rhine, would all give different answers.

What alignment language does is move the line of inquiry. Instead of wondering only, “Is this real?”, you start asking, “Real or not, what is this practice training itself to want?” and “Am I willing to take responsibility for that?”

If you sat down tomorrow and declared, aloud and in writing, “From now on, this oracle optimises for X,” what would you choose — and how much of your current work would you have to dismantle for that to be true?