Investigation trail
Goal → prediction → experiment → observation → revision
Investigation intelligence / 01
Evidence Atlas
Follow the chain from goal to action to result to revision. Compare what Tend expected with what the evidence actually earned, and see which findings changed the next move.
◈
Evidence discipline is the interface. Observations, interpretations, and unknowns stay separate. This is a curated demo snapshot based on recorded Tend milestone outcomes; it is not a live connection to Tend or a complete research log.
Atlas entries
06
4 PRs · plus boundaries and costs
Obligations proved
8 / 12
Z3 screen · PR381
Unknown / unsupported
4
3 unsupported · 1 solver unknown
Transfer result
Negative
Later investigation cost increased
4 PR records
Decision trace
The finding that changed the next choice
G
Goal
Determine whether retained learning improves a later investigation.
P
Prediction
Retained history should change a consequential action or test.
O
Observation · PR375
Both guides chose candidate 6. Diversity did not produce a different consequential test.
R
Revision
Treat action divergence as required evidence for transfer, not a proxy metric.
↗
Later result · PR384
Prior outcomes were consumed, but the later investigation cost more. Transfer was negative.
Changed later investigative choice
Do not call retained knowledge useful because it was read. Measure whether it changes a consequential choice, then whether that change improves cost or outcome.
Current progress vs retained history
Same task, two information conditions · compare decisions and outcomes
| Measure | Current progress | Retained history | What the result supports |
|---|---|---|---|
| Candidate selected · PR375 | Candidate 6 | Candidate 6 | Same choice; no demonstrated action divergence |
| Later investigation cost · PR384 | Baseline not stated | Increased | Negative transfer in this reported comparison |
| Proof screen · PR381 | 12 inputs screened | 8 proved; 3 unsupported legacy semantics; 1 solver unknown | Proof coverage is partial; unknowns remain unresolved |
| Repair loop · PR383 | 150 predicates | Reduced to 5; 3 obligations passed | Promising result on this benchmark scope only |
Cost numbers are unknown. The source notes establish direction for PR384, not a numeric delta. Baseline and absolute cost are not supplied here.
Policies earned so far
Portable rules with evidence boundaries
01
Require a consequential action or test difference before claiming that retained knowledge changes an investigation.Derived from PR375 guide comparison · scope: tested benchmark
02
Report transfer cost and outcome. Consuming prior outcomes alone is not evidence of improvement.Supported by negative transfer result · PR384
03
Preserve uncertainty classes: unsupported semantics and solver unknown are different failure states.Applied to PR381 · not a generalized policy yet
Cost & unresolved work
Unknown is a first-class result
PR384 absolute investigation costUNKNOWN
PR384 baseline cost definitionUNKNOWN
PR381 unsupported legacy semantics3 cases
PR381 solver result1 unknown
Comparison bench
Current progress and retained history across the documented experiments
| Goal / experiment | Expected | Observed | Revision or next choice | Cost |
|---|---|---|---|---|
| PR375 · guide comparison | Retained history changes a consequential action/test | Both guides selected candidate 6; no consequential difference shown | Make action/test divergence a prerequisite for transfer claims | Not reported |
| PR381 · invariant proof attempt | Prove invariants over 12 inputs | 8 proved; 3 legacy semantics unsupported; 1 solver unknown | Keep proof, unsupported, and unknown states distinct | Not reported |
| PR383 · repair loop | Reduce search while satisfying three obligations | 150 predicates → 5; all three reported obligations passed | Retain initiation positives, preservation implication pairs, sufficiency negatives | Not reported |
| PR384 · transfer | Prior outcomes improve later investigation | Prior outcomes consumed; later investigation became more expensive | Record negative transfer; inspect cause before reusing the policy | Increase reported; absolute delta unknown |
A prediction is not a result. “Not reported” and “unknown” remain visible instead of being silently filled in.
Repair outcomes
Preservation evidence is an implication, not reachability
+
Initiation positives
Recorded as constraints that must remain included.
→
Preservation implication pairs
If predecessor is included, successor must be included.
−
Sufficiency negatives
A failed preservation check can instead exclude its predecessor on restart.
Failed preservation & transfer
Two ways a model can mislead
FP
Failed preservation is a constraint signal, not proof that the successor is reachable.PR383 records implication pairs and restart exclusion behavior
NT
Negative transfer means prior experience made the later investigation more expensive.PR384 · mechanism and numeric cost delta unknown
Learning ledger
What changed in the system’s future choices?
Evidence classes
Read the epistemic status before the claim
Observed
Directly reported outcome or count.Example: 8 of 12 inputs proved in PR381.
Inference
A conclusion drawn from observations; scope stays attached.Example: tested guides did not demonstrate consequential divergence.
Unknown
Not supplied or not resolved. No guess substituted.Example: absolute PR384 cost delta.
Negative
Evidence against the expected improvement.Example: retained outcomes increased later investigation cost.