| name | benchmark-mech-interp-analysis |
| description | Use when a benchmark already has a validated latent label spec and the next job is to plan or review the actual mechanistic analysis program. Covers feature hypotheses, readout and localization strategy, probe choice, control design, evidence ladder, and first experiments. |
Benchmark Mech Interp Analysis
Use this skill when the user wants to:
- turn validated latent labels into a concrete mech-interp analysis plan
- choose methods per label family
- plan evidence from readout through localization and causal follow-up
- define first experiments for a benchmark program
- review whether a benchmark analysis plan is scientifically credible
This skill is the analysis planning phase of the benchmark-flow pipeline.
It assumes the benchmark has already gone through:
- validation
- latent-label formation
- augmentation if needed
This skill should not be the place where native labels are first interpreted.
Core rule
Do not jump from labels to mechanism without an evidence ladder.
The right progression is usually:
- Level 1 — behavioral. The model behaves sanely on the task.
- Level 2 — representational. A readout or probe detects the target variable.
- Level 3 — localized representational. The signal is localized to a span, section, token, layer, or position.
- Level 4 — causal. An intervention changes behavior beyond controls.
- Level 5 — mechanistic. A plausible computation path is identified.
- claim shaping
Required gate
Before using this skill, complete the behavioral-sanity check in:
In particular:
- inspect real examples from each label family
- run the probe-target model on a small slice
- verify output parseability
- inspect failures manually
If the task is not behaviorally sane for the target model, stop and repair it first.
The preferred canonical input artifact for this gate is:
workspaces/<workspace>/[<thread>/]phase_02/docs/02-behavioral-smoke-report.md
If an equivalent artifact is used instead, name it explicitly and explain why.
For response-side work, "behaviorally sane" should be interpreted at the level of proposed labels, not just answer existence.
Ask:
- are the generated responses actually labelable for the planned response-side labels?
- do the responses contain enough substantive content to support annotation and validation?
- are the apparent passes only tripwire-level, or do they survive close inspection?
Required lexical gate
Before probe work begins, run a cheap lexical-baseline gate on the exact split you plan to claim on.
Minimum expectation:
- one strong text baseline, usually
char-TFIDF + logistic
- one trivial non-semantic baseline, usually
length
- the actual held-out evaluation split, not just random train/test
If the experiment will rely on multiple generalization claims, test the baseline on each relevant split family before probing, for example:
- dilemma or group holdout
- source-family holdout
- prompt-format or bank holdout
- cross-language holdout
Do not treat this as a box-check. The purpose is to identify what kind of shortcut is already available:
- cue-anchor leakage
- theory-canonical response vocabulary
- template leakage
- abstraction-level asymmetry
- language-specific calibration
Default rule:
- if a cheap lexical baseline is already near ceiling on the intended split, probing is blocked until you either redesign the split, redesign the target, or redesign the data
This gate is as important as behavioral sanity. A benchmark can be behaviorally coherent and still be probe-invalid because the target is shortcut-satisfiable from text alone.
What this skill owns
1. Feature hypotheses from latent labels
For each refined label, ask:
- what internal representation or control state might exist?
- what would a readable version of that signal look like?
- what would a causally useful version look like?
2. Method selection per label family
Choose methods per label, not per benchmark.
Common options:
- linear probes
- difference-in-means directions
- PCA / geometry checks
- span-local localization
- transfer tests
- steering or erasure
- patching follow-up
For response-side labels, the first-pass probe plan should default to pairing techniques from at least two of the confound-reduction categories in constructing-llm-probes/ANALYSIS.md:
- always report full-sequence and at least one viewport-reduced probe (tail window, conclusion span, non-header) together; convergence at ceiling signals lexical domination, divergence localizes where the signal lives
- include a residualized probe against the strongest text baseline whenever a text baseline exists for that label
- if training data spans multiple prompt formats or paraphrase variants, plan cross-format transfer evaluation alongside within-format evaluation
- do not accept a single-technique representational claim on a response-side label that was not paired with at least one confound-reduction technique
See the shared principle on response-side probing and active confound reduction.
3. Evidence ladder planning
Decide what level of claim is realistic:
- Level 1 — behavioral. The model behaves sanely on the task.
- Level 2 — representational. A readout or probe detects the target variable.
- Level 3 — localized representational. The signal is localized to a span, section, token, layer, or position.
- Level 4 — causal. An intervention changes behavior beyond controls.
- Level 5 — mechanistic. A plausible computation path is identified.
4. Split and control planning
Carry forward the confound plan into:
- split design
- baselines
- matched comparisons
- same-label controls where relevant
5. First experiment specs
For the highest-value label families, define:
- target model
- prompt / generation protocol
- capture location
- first-pass readout
- stronger follow-up
- failure criteria
If an experiment depends on fresh generations, the spec should also define:
- generation batch
- annotation / validation step
- label freeze prerequisite
- the point at which probing is allowed to begin
If the target variable is explicitly named in the prompt, the spec should also define:
- the main shortcut risk from explicit naming
- the anti-shortcut control required before a strong success claim is allowed
6. Claim shaping
At the end of analysis planning, explicitly state what each planned method would support if it succeeds.
Useful pattern:
- behavior establishes the task is real
- readout establishes representation
- localization establishes where the signal emerges
- intervention establishes causal leverage
- later mechanism follow-up may support stronger computation claims
7. Result triage and routing
After experiments run, record what the result means for the research thread, not just the metric.
For each experiment, explicitly decide whether the outcome means:
- stay in phase 03 and tighten controls
- hand back to phase 02 for augmentation
- hand back to phase 01 for label refinement
- promote to phase 04 causal follow-up
- or stop / demote the label family
Use this routing logic:
- Did a named cheap baseline in
03-controls-and-splits.md achieve baseline-dominant performance on the same target?
- yes ->
AUGMENTATION_NEEDED
- no -> continue
- for imported benchmark labels, treat this as the default posture under the shared principle that benchmarks leak by construction
- Did the result beat the cheap baselines named in
03-controls-and-splits.md?
- no ->
TRIVIAL_OR_NULL
- yes -> continue
- Did the result survive the planned controls?
- no ->
CONTROL_INSUFFICIENT
- yes -> continue
- Did the result transfer across the planned splits?
- no, because the label definition is wrong or unstable ->
LABEL_REFINEMENT_NEEDED
- no, because the data still entangles the target with a confound ->
AUGMENTATION_NEEDED
- yes -> continue
- Is the result localized to a meaningful span, position, or layer band?
- no ->
REPRESENTATIONAL_ONLY
- yes -> continue
- Does the localized result satisfy the phase-04 entry criteria?
- no ->
LOCALIZED_REPRESENTATIONAL
- yes ->
PROMOTE_TO_CAUSAL
Baseline-dominant performance means a cheap lexical, semantic, template, or positional baseline matches the main signal closely enough that the experiment remains shortcut-satisfiable.
Default rule:
- if a named cheap baseline matches the main signal closely enough that the result is still plausibly explained by surface recoverability, do not accept a thin representational win by default
- route to
AUGMENTATION_NEEDED unless the experiment was explicitly designed only to measure prompt-family recoverability
- for active benchmark targets, leakage-dominant wins should default to repair, not quiet demotion
- any override should be written explicitly in the triage log, including:
- what the neural readout captures that the baseline does not
- why the baseline match is acceptable for that experiment
Verdicts and routing:
TRIVIAL_OR_NULL
stop, demote, or park the label family
CONTROL_INSUFFICIENT
stay in phase 03 and strengthen controls / baselines
LABEL_REFINEMENT_NEEDED
hand back to phase 01 and re-open the latent label spec and labeling functions
AUGMENTATION_NEEDED
hand back to phase 02 and add the missing repair to the augmentation plan and gap resolution
this is the default verdict for shortcut-dominated families whose main signal is matched by a cheap baseline
REPRESENTATIONAL_ONLY
cap claims at level 2
LOCALIZED_REPRESENTATIONAL
cap claims at level 3
PROMOTE_TO_CAUSAL
hand off to mechanistic-interventions
References
Primary references:
Use constructing-llm-probes for probe implementation details.
Required artifacts
At minimum, leave behind:
workspaces/<workspace>/[<thread>/]phase_03/docs/03-analysis-plan.md
the main human-readable analysis plan, including:
- methodology choices by label family
- evidence ladder
- behavioral-sanity gate status
- expected artifacts
- key risks
workspaces/<workspace>/[<thread>/]phase_03/docs/03-feature-hypotheses.md
label-family -> candidate representation/control-state hypotheses
workspaces/<workspace>/[<thread>/]phase_03/docs/03-experiment-specs.md
2-5 concrete first experiments with method, split, and success criteria
workspaces/<workspace>/[<thread>/]phase_03/docs/03-controls-and-splits.md
the control plan carried into actual experiments
workspaces/<workspace>/[<thread>/]phase_03/docs/03-analysis-summary.json
compact structured summary of chosen label families, methods, and next steps
workspaces/<workspace>/[<thread>/]phase_03/docs/03-result-triage-log.md
one entry per executed experiment, including verdict and routing
Also leave behind:
workspaces/<workspace>/[<thread>/]phase_03/docs/03-execution-targets.md
execution-facing model and inference config artifact, including:
- primary execution model
- any secondary comparison models if planned
- which model currently satisfies the behavioral gate
- exact inference config and system prompt
workspaces/<workspace>/[<thread>/]phase_03/docs/03-phase-04-entry-criteria.md
explicit gate for promotion from phase 03 readout/localization work to phase 04 causal follow-up
If any phase-03 experiment depends on response-side labels from fresh generations, also leave behind:
workspaces/<workspace>/[<thread>/]phase_03/docs/03-response-label-pilot.md
generation -> annotate -> validate -> freeze plan for the response-side labels
If a label family is blocked on gold-slice validation, leave behind a concrete prerequisite artifact such as:
workspaces/<workspace>/[<thread>/]phase_03/docs/03-<label>-gold-slice-spec.md
including sample size, labeler identity or procedure, and the validation threshold required before probing
Use simple frontmatter on markdown artifacts:
benchmark
phase
version
frozen_date
input_artifacts
Someone inspecting this phase should be able to answer:
- what internal features we think might exist
- which methods we plan to use first
- what level of claim we think is realistic
- what the first concrete experiments are
- what each executed experiment implies about the next phase of work
Phase Done Criteria
This phase is done when:
- the behavioral gate is satisfied and named explicitly
- the required planning artifacts exist
- each executed experiment has a triage-log entry with:
- baseline comparison
- control survival
- split-transfer judgment
- verdict
- routing
- phase-04 escalation is not proposed without a
PROMOTE_TO_CAUSAL triage verdict
- whether the work is only planning-ready or actually execution-ready
The analysis summary should make execution readiness explicit.
Useful statuses include:
planning_ready
execution_blocked_on_target_model
execution_blocked_on_response_label_freeze
execution_ready
Each experiment spec should include, either inline or by direct reference:
- primary shortcut risk
- required baseline or control
- claim ceiling
- label-freeze prerequisite when applicable
Handoff
If the next step is readout or localization work:
If the next step is causal intervention planning:
Update the benchmark sidecar with:
- strongest candidate feature hypotheses
- methods that look promising
- methods or comparisons to be careful about
- triage decisions that changed the benchmark plan materially